{
 "groupbg::산림_범행지": {
  "input_fingerprint": "d2a119c0b346a68e",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "산림_범행지",
    "tags": [
     "S1sh1",
     "S1sh5",
     "S3sh3"
    ]
   },
   "context_sig": "2cd39f2e434ae5db"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n금월산 숲속 범행·수사 현장: 우거진 수풀이 빽빽하게 들어찬 산속 등산로 주변의 어둡고 습한 흙바닥. 노란색 폴리스 라인이 처져 있고 감식 장비들이 놓여 있다. (특징: 우거진 수풀과 나뭇가지들; 바닥에 파인 흔적과 핏자국; 주변을 둘러싼 폴리스 라인; 현대식 과학수사 장비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#1. 숲속 / 해 질 무렵\n- 씬1, 2 현장을 감식하는 경찰, 폴리스 라인을 치고 시신을 수거한다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n금월산 숲속 범행·수사 현장: 우거진 수풀이 빽빽하게 들어찬 산속 등산로 주변의 어둡고 습한 흙바닥. 노란색 폴리스 라인이 처져 있고 감식 장비들이 놓여 있다. (특징: 우거진 수풀과 나뭇가지들; 바닥에 파인 흔적과 핏자국; 주변을 둘러싼 폴리스 라인; 현대식 과학수사 장비)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#1. 숲속 / 해 질 무렵\n- 씬1, 2 현장을 감식하는 경찰, 폴리스 라인을 치고 시신을 수거한다.\n\nTIME OF DAY (lock): dusk.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_산림_범행지_a8cfa8.png",
  "asset_id": "ebe5401b-668a-4420-9d48-1057e290897c",
  "input_asset_ids": [
   "30ef44c4-73ef-4348-a59c-c7073b4d8ea6"
  ],
  "origin_tag": "S1sh1",
  "place_text": "On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell.",
  "origin_inputs": {
   "place_text": "On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell.",
   "time_of_day_en": "dusk",
   "conti_asset_id": "30ef44c4-73ef-4348-a59c-c7073b4d8ea6"
  }
 },
 "S1sh1::bgfirst_bg": {
  "input_fingerprint": "316784d82fa4d0bf",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 흙바닥에 짓눌린 금월산 피해 남성의 맨살 허벅지 위로, 핏물이 고인 커다란 이빨 자국이 선명하게 파인 클로즈업.\n\nLOCATION (lock): On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From ground level within close-up distance, aim slightly downward across the injured thigh as the bite wound sits near frame center and the man's hip and the soil provide scale at the edges. Hold the stage-entry composition nearly static before the beginning of the prescribed pullback, using shallow focus to isolate the pooled blood and tooth impressions.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 흙바닥 (The man's injured thigh is pressed against it); used as Visible beneath and around the thigh as the immediate ground plane and scale reference; 우거진 수풀 (Dense around the ground-level view); used as Soft peripheral framing that preserves the forest location without competing with the wound.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diminishing sunset ambient light holds a dark, desaturated forest palette while the blood supplies the restrained crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 흙바닥에 짓눌린 금월산 피해 남성의 맨살 허벅지 위로, 핏물이 고인 커다란 이빨 자국이 선명하게 파인 클로즈업.\n\nLOCATION (lock): On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From ground level within close-up distance, aim slightly downward across the injured thigh as the bite wound sits near frame center and the man's hip and the soil provide scale at the edges. Hold the stage-entry composition nearly static before the beginning of the prescribed pullback, using shallow focus to isolate the pooled blood and tooth impressions.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 흙바닥 (The man's injured thigh is pressed against it); used as Visible beneath and around the thigh as the immediate ground plane and scale reference; 우거진 수풀 (Dense around the ground-level view); used as Soft peripheral framing that preserves the forest location without competing with the wound.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diminishing sunset ambient light holds a dark, desaturated forest palette while the blood supplies the restrained crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S1sh1__bgfirst_bg.png",
  "asset_id": "beb1db85-9c58-4701-bf13-0d6efeb53eec",
  "input_asset_ids": [
   "30ef44c4-73ef-4348-a59c-c7073b4d8ea6",
   "ebe5401b-668a-4420-9d48-1057e290897c"
  ]
 },
 "S1sh1": {
  "input_fingerprint": "a13f32a26178ee9e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 짓눌린 금월산 피해 남성의 맨살 허벅지 위로, 핏물이 고인 커다란 이빨 자국이 선명하게 파인 클로즈업.\n\nLOCATION (lock): On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From ground level within close-up distance, aim slightly downward across the injured thigh as the bite wound sits near frame center and the man's hip and the soil provide scale at the edges. Hold the stage-entry composition nearly static before the beginning of the prescribed pullback, using shallow focus to isolate the pooled blood and tooth impressions.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 흙바닥 (The man's injured thigh is pressed against it); used as Visible beneath and around the thigh as the immediate ground plane and scale reference; 우거진 수풀 (Dense around the ground-level view); used as Soft peripheral framing that preserves the forest location without competing with the wound.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diminishing sunset ambient light holds a dark, desaturated forest palette while the blood supplies the restrained crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The mountain victim has one shoe missing and a large, actively bleeding bite mark on his bare thigh; both conditions must remain unchanged throughout his flight and attack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 금월산 피해 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 짓눌린 금월산 피해 남성의 맨살 허벅지 위로, 핏물이 고인 커다란 이빨 자국이 선명하게 파인 클로즈업.\n\nLOCATION (lock): On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From ground level within close-up distance, aim slightly downward across the injured thigh as the bite wound sits near frame center and the man's hip and the soil provide scale at the edges. Hold the stage-entry composition nearly static before the beginning of the prescribed pullback, using shallow focus to isolate the pooled blood and tooth impressions.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 흙바닥 (The man's injured thigh is pressed against it); used as Visible beneath and around the thigh as the immediate ground plane and scale reference; 우거진 수풀 (Dense around the ground-level view); used as Soft peripheral framing that preserves the forest location without competing with the wound.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diminishing sunset ambient light holds a dark, desaturated forest palette while the blood supplies the restrained crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The mountain victim has one shoe missing and a large, actively bleeding bite mark on his bare thigh; both conditions must remain unchanged throughout his flight and attack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 금월산 피해 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 흙바닥에 짓눌린 금월산 피해 남성의 맨살 허벅지 위로, 핏물이 고인 커다란 이빨 자국이 선명하게 파인 클로즈업.\n\nLOCATION (lock): On the dirt floor beneath dense undergrowth in a mountain forest, at the spot where the fleeing victim fell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From ground level within close-up distance, aim slightly downward across the injured thigh as the bite wound sits near frame center and the man's hip and the soil provide scale at the edges. Hold the stage-entry composition nearly static before the beginning of the prescribed pullback, using shallow focus to isolate the pooled blood and tooth impressions.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 흙바닥 (The man's injured thigh is pressed against it); used as Visible beneath and around the thigh as the immediate ground plane and scale reference; 우거진 수풀 (Dense around the ground-level view); used as Soft peripheral framing that preserves the forest location without competing with the wound.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diminishing sunset ambient light holds a dark, desaturated forest palette while the blood supplies the restrained crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The mountain victim has one shoe missing and a large, actively bleeding bite mark on his bare thigh; both conditions must remain unchanged throughout his flight and attack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 금월산 피해 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S1sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S1sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 금월산 피해 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:601806>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_산림_범행지_a8cfa8.png"
    },
    {
     "label": "CHARACTER REFERENCE — 금월산 피해 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:601806>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "핵심인 이빨 자국이 프레임 중앙에 위치하나 핏물이 고인 질감이 부족하며, 지시된 클로즈업보다 프레이밍이 다소 넓게 잡혔습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트의 핵심인 상처와 핏물이 전혀 묘사되지 않았으며, 다리의 해부학적 구조와 신발의 위치가 물리적으로 불가능하게 왜곡되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "핵심인 이빨 자국이 프레임 중앙에 위치하나 핏물이 고인 질감이 부족하며, 지시된 클로즈업보다 프레이밍이 다소 넓게 잡혔습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트의 핵심인 상처와 핏물이 전혀 묘사되지 않았으며, 다리의 해부학적 구조와 신발의 위치가 물리적으로 불가능하게 왜곡되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "지정된 클로즈업 구도와 의상(회색 셔츠, 벨트), 배경은 정확히 구현했으나, 가장 핵심적인 이빨 자국과 핏물이 완전히 누락되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 클로즈업 구도를 위반하여 상체와 다리 전체를 노출했고, 상의 탈의 오류가 있으며 상처가 평면적인 붉은 선처럼 묘사되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "지정된 클로즈업 구도와 의상(회색 셔츠, 벨트), 배경은 정확히 구현했으나, 가장 핵심적인 이빨 자국과 핏물이 완전히 누락되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "요구된 클로즈업 구도를 위반하여 상체와 다리 전체를 노출했고, 상의 탈의 오류가 있으며 상처가 평면적인 붉은 선처럼 묘사되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 8
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "totals": {
   "A": 9,
   "B": 8
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "핵심인 이빨 자국이 프레임 중앙에 위치하나 핏물이 고인 질감이 부족하며, 지시된 클로즈업보다 프레이밍이 다소 넓게 잡혔습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "프롬프트의 핵심인 상처와 핏물이 전혀 묘사되지 않았으며, 다리의 해부학적 구조와 신발의 위치가 물리적으로 불가능하게 왜곡되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_산림_범행지_a8cfa8.png"
   },
   {
    "label": "CHARACTER REFERENCE — 금월산 피해 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:601806>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "레이아웃 스케치의 허벅지 클로즈업 구도를 무시하고 피사체가 너무 작게 배치되어 지시되지 않은 맨 가슴과 다리 전체가 보입니다.",
     "fix_en": "Scale up and reposition the subject to exactly match the extreme close-up framing of the thigh from the layout sketch."
    },
    {
     "issue_ko": "프롬프트와 스케치에 명시된, 피가 고인 깊고 커다란 이빨 자국이 허벅지에 없습니다(희미한 붉은 자국만 존재).",
     "fix_en": "Render the large, deep, actively bleeding tooth marks with pooled blood into the center of the thigh exactly as sketched."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Scale up and reposition the subject to exactly match the extreme close-up framing of the thigh from the layout sketch.\n- Render the large, deep, actively bleeding tooth marks with pooled blood into the center of the thigh exactly as sketched.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지시된 클로즈업 화각과 엉덩이, 흙바닥을 이용한 프레이밍은 정확히 구현했으나, 선명하게 파인 이빨 자국과 핏물의 디테일이 부족합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "클로즈업 지시를 위반하고 신체와 배경을 너무 넓게 보여주었으며, 상처의 디테일 또한 프롬프트에 미치지 못합니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지시된 클로즈업 화각과 엉덩이, 흙바닥을 이용한 프레이밍은 정확히 구현했으나, 선명하게 파인 이빨 자국과 핏물의 디테일이 부족합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "클로즈업 지시를 위반하고 신체와 배경을 너무 넓게 보여주었으며, 상처의 디테일 또한 프롬프트에 미치지 못합니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "지정된 클로즈업 샷 크기와 스케치의 프레이밍을 잘 따랐으나, 지시문이 요구한 깊게 파이고 피가 고인 선명한 이빨 자국을 제대로 구현하지 못했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 지시를 무시하고 씬을 너무 넓게 보여주는 치명적인 프레이밍 오류를 범했으며, 상처의 묘사 역시 지시문에 미치지 못합니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "지정된 클로즈업 샷 크기와 스케치의 프레이밍을 잘 따랐으나, 지시문이 요구한 깊게 파이고 피가 고인 선명한 이빨 자국을 제대로 구현하지 못했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "클로즈업 지시를 무시하고 씬을 너무 넓게 보여주는 치명적인 프레이밍 오류를 범했으며, 상처의 묘사 역시 지시문에 미치지 못합니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 11
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S1sh1__bgfirst_bg.png",
   "bg_asset_id": "beb1db85-9c58-4701-bf13-0d6efeb53eec",
   "bg_record_key": "S1sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "산림_범행지",
   "groupbg_asset_id": "ebe5401b-668a-4420-9d48-1057e290897c"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S1sh5::variants": {
  "author_fp": "f8a81b4776b48ea2",
  "author": {
   "variants": [
    {
     "approach_ko": "수직 원근과 깊은 초점으로 숲 바닥의 질감과 겹친 두 인물을 냉정하고 법의학적으로 포착한다.",
     "prompt_en": "Use a rectilinear overhead perspective with restrained depth and edge-to-edge clarity, making the forest floor read as a flat evidentiary plane around the overlapping bodies. Keep the composition austere and nearly symmetrical, with fine texture in leaves, soil, clothing folds, and supported limbs; soft dusk ambience provides muted tonal separation while the attacker remains visibly human beneath the dense darkness."
    },
    {
     "approach_ko": "잔광의 비스듬한 명암과 압축된 원근으로 가슴 위의 무게와 폐쇄감을 강조한다.",
     "prompt_en": "Use a compressed overhead lens feel that visually tightens the brush around the fully legible bodies and concentrates attention on the crushing contact across the torso. Let the last directional dusk light skim unevenly across the ground and the victim while the attacker falls into heavier, dimensional shadow, preserving human anatomy and physical fabric detail; hold the surrounding edges darker and softer for a suffocating, motionless tension."
    }
   ]
  },
  "reused": false
 },
 "S1sh5": {
  "input_fingerprint": "8661abe076d3439a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 바닥에 쓰러진 금월산 피해 남성의 몸통 위를 짙은 그림자 같은 존재가 완전히 짓누른 채 정지된 하이앵글 구도.\n\nLOCATION (lock): On the ground amid dense brush in a mountain forest, where the victim is pinned beneath his attacker. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The crane settles directly above the bodies into a static, straight-down high-angle hold, with the dark figure laid across the center of the man's torso and the man's limbs extending into the surrounding frame. Keep both bodies fully legible against the forest floor, emphasizing the completed vertical rise rather than adding lateral movement.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 정체불명 검은 존재 in the middle-center of the frame, midground; 금월산 피해 남성 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 숲 바닥 (Both figures lie against it); used as The overhead ground plane separates the overlapping silhouettes of attacker and victim; 우거진 수풀 (Dense around the attack site); used as Irregular perimeter framing around the bodies in the vertical composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading sunset ambience creates low-key, desaturated separation, with the dark figure reading as a dense mass over the victim.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His body is collapsed on the ground with his torso completely pinned beneath the dark figure, leaving him unable to rise; the source does not establish the exact positions of his head or limbs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): His body is collapsed on the ground with his torso completely pinned beneath the dark figure, leaving him unable to rise; the man still has one shoe missing and the same large bleeding bite mark on his thigh.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 금월산 피해 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상); 정체불명 검은 존재 (식별되지 않는 얼굴, 불분명한 머리 윤곽, 인간형 머리 실루엣) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 바닥에 쓰러진 금월산 피해 남성의 몸통 위를 짙은 그림자 같은 존재가 완전히 짓누른 채 정지된 하이앵글 구도.\n\nLOCATION (lock): On the ground amid dense brush in a mountain forest, where the victim is pinned beneath his attacker. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The crane settles directly above the bodies into a static, straight-down high-angle hold, with the dark figure laid across the center of the man's torso and the man's limbs extending into the surrounding frame. Keep both bodies fully legible against the forest floor, emphasizing the completed vertical rise rather than adding lateral movement.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 정체불명 검은 존재 in the middle-center of the frame, midground; 금월산 피해 남성 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 숲 바닥 (Both figures lie against it); used as The overhead ground plane separates the overlapping silhouettes of attacker and victim; 우거진 수풀 (Dense around the attack site); used as Irregular perimeter framing around the bodies in the vertical composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading sunset ambience creates low-key, desaturated separation, with the dark figure reading as a dense mass over the victim.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His body is collapsed on the ground with his torso completely pinned beneath the dark figure, leaving him unable to rise; the source does not establish the exact positions of his head or limbs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): His body is collapsed on the ground with his torso completely pinned beneath the dark figure, leaving him unable to rise; the man still has one shoe missing and the same large bleeding bite mark on his thigh.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 금월산 피해 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상); 정체불명 검은 존재 (식별되지 않는 얼굴, 불분명한 머리 윤곽, 인간형 머리 실루엣) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a rectilinear overhead perspective with restrained depth and edge-to-edge clarity, making the forest floor read as a flat evidentiary plane around the overlapping bodies. Keep the composition austere and nearly symmetrical, with fine texture in leaves, soil, clothing folds, and supported limbs; soft dusk ambience provides muted tonal separation while the attacker remains visibly human beneath the dense darkness.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 바닥에 쓰러진 금월산 피해 남성의 몸통 위를 짙은 그림자 같은 존재가 완전히 짓누른 채 정지된 하이앵글 구도.\n\nLOCATION (lock): On the ground amid dense brush in a mountain forest, where the victim is pinned beneath his attacker. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The crane settles directly above the bodies into a static, straight-down high-angle hold, with the dark figure laid across the center of the man's torso and the man's limbs extending into the surrounding frame. Keep both bodies fully legible against the forest floor, emphasizing the completed vertical rise rather than adding lateral movement.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 정체불명 검은 존재 in the middle-center of the frame, midground; 금월산 피해 남성 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 숲 바닥 (Both figures lie against it); used as The overhead ground plane separates the overlapping silhouettes of attacker and victim; 우거진 수풀 (Dense around the attack site); used as Irregular perimeter framing around the bodies in the vertical composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading sunset ambience creates low-key, desaturated separation, with the dark figure reading as a dense mass over the victim.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): His body is collapsed on the ground with his torso completely pinned beneath the dark figure, leaving him unable to rise; the source does not establish the exact positions of his head or limbs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): His body is collapsed on the ground with his torso completely pinned beneath the dark figure, leaving him unable to rise; the man still has one shoe missing and the same large bleeding bite mark on his thigh.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 금월산 피해 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상); 정체불명 검은 존재 (식별되지 않는 얼굴, 불분명한 머리 윤곽, 인간형 머리 실루엣) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a compressed overhead lens feel that visually tightens the brush around the fully legible bodies and concentrates attention on the crushing contact across the torso. Let the last directional dusk light skim unevenly across the ground and the victim while the attacker falls into heavier, dimensional shadow, preserving human anatomy and physical fabric detail; hold the surrounding edges darker and softer for a suffocating, motionless tension.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 3,
   "B": 8
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "프레임 화면 내에 카메라 장비가 노출되었으며, 피해 남성의 다리가 여러 개로 증식되는 치명적인 해부학적 오류가 발생했습니다."
   },
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "수직 하이앵글 구도와 인물들의 배치, 벗겨진 신발과 허벅지의 물린 자국 등 이전 샷의 상태와 레퍼런스를 매우 정확하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S1sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 금월산 피해 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:601806>"
   },
   {
    "label": "CHARACTER REFERENCE — 정체불명 검은 존재: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:833745>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "검은 존재가 남성의 하반신에 위치해 있으며, 지문과 같이 남성의 몸통(가슴과 배)을 완전히 짓누르고 있지 않습니다.",
     "fix_en": "Move the dark figure upwards so it is laid directly across and completely pins the center of the man's torso (the green jacket)."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Move the dark figure upwards so it is laid directly across and completely pins the center of the man's torso (the green jacket).\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 완벽한 수직 하이앵글 구도와 두 인물의 배치, 허벅지의 물린 상처와 벗겨진 신발 등 디테일을 정확히 구현함."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "A와 동일한 이미지로, 프롬프트의 구도와 인물 상태, 물리적 환경 지시를 매우 훌륭하게 준수함."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 완벽한 수직 하이앵글 구도와 두 인물의 배치, 허벅지의 물린 상처와 벗겨진 신발 등 디테일을 정확히 구현함."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "A와 동일한 이미지로, 프롬프트의 구도와 인물 상태, 물리적 환경 지시를 매우 훌륭하게 준수함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 하이앵글 구도와 주변 수풀 환경을 정확히 구현했으며, 손의 해부학적 형태와 허벅지의 물린 자국이 더 사실적입니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "구도와 인물 배치는 훌륭하나, 피해자의 손가락 형태가 다소 부자연스럽고 물린 자국의 묘사가 덜 사실적입니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 하이앵글 구도와 주변 수풀 환경을 정확히 구현했으며, 손의 해부학적 형태와 허벅지의 물린 자국이 더 사실적입니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "구도와 인물 배치는 훌륭하나, 피해자의 손가락 형태가 다소 부자연스럽고 물린 자국의 묘사가 덜 사실적입니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S1sh1"
  }
 },
 "groupbg::도시_뒷골목": {
  "input_fingerprint": "aee9a7f47c804cbc",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "도시_뒷골목",
    "tags": [
     "S2sh3",
     "S2sh5"
    ]
   },
   "context_sig": "34be1e2f52f0fc1a"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: Beside a utility pole in a dim urban back alley at night, under faint streetlight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n도시 뒷골목: 희미한 가로등 불빛이 비추는 좁고 어두운 도심의 뒷골목. 낡은 전봇대 주변으로 쓰레기봉투와 잡동사니가 널려 있다. (특징: 좁고 어두운 골목길; 빛이 약한 구형 가로등; 전단지가 붙어 있는 낡은 콘크리트 전봇대; 어두운 그림자와 낮은 채도의 벽돌벽)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#2. 도시 뒷골목 / N\n- 도심 희미한 가로등 불빛.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: Beside a utility pole in a dim urban back alley at night, under faint streetlight.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n도시 뒷골목: 희미한 가로등 불빛이 비추는 좁고 어두운 도심의 뒷골목. 낡은 전봇대 주변으로 쓰레기봉투와 잡동사니가 널려 있다. (특징: 좁고 어두운 골목길; 빛이 약한 구형 가로등; 전단지가 붙어 있는 낡은 콘크리트 전봇대; 어두운 그림자와 낮은 채도의 벽돌벽)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#2. 도시 뒷골목 / N\n- 도심 희미한 가로등 불빛.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_도시_뒷골목_43d35b.png",
  "asset_id": "3c03c327-57b1-423a-ae8f-b63ae5ae45c6",
  "input_asset_ids": [
   "1afb33b4-a07e-48e8-8b9a-36f6d70cd679"
  ],
  "origin_tag": "S2sh3",
  "place_text": "Beside a utility pole in a dim urban back alley at night, under faint streetlight.",
  "origin_inputs": {
   "place_text": "Beside a utility pole in a dim urban back alley at night, under faint streetlight.",
   "time_of_day_en": "night",
   "conti_asset_id": "1afb33b4-a07e-48e8-8b9a-36f6d70cd679"
  }
 },
 "S2sh3::bgfirst_bg": {
  "input_fingerprint": "4934f6e8ea02b516",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 한국인 신원미상 남성의 입술이 효장동 피해 여성의 목덜미를 깊게 문 채, 검붉은 핏방울이 공중에 튄 정지 순간.\n\nLOCATION (lock): Beside a utility pole in a dim urban back alley at night, under faint streetlight.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At neck height beside the pair, finish the dolly-in on an extremely tight opposing-profile composition: the man's lips and teeth enter from one side, the woman's neck from the other, and suspended blood droplets remain in the narrow gap between them. The frame freezes the instant before the planned downward tilt, excluding almost all alley context while preserving the physical bite.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 전봇대 (The two people are pressed near it) — Only a shallow section of the side facing the pair remains visible behind them; used as A narrow blurred vertical trace behind the tightly framed profiles anchors them to the alley location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Faint streetlight illumination leaves the profiles in low-key desaturated contrast, with the airborne blood as the only restrained dark-crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 한국인 신원미상 남성의 입술이 효장동 피해 여성의 목덜미를 깊게 문 채, 검붉은 핏방울이 공중에 튄 정지 순간.\n\nLOCATION (lock): Beside a utility pole in a dim urban back alley at night, under faint streetlight.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At neck height beside the pair, finish the dolly-in on an extremely tight opposing-profile composition: the man's lips and teeth enter from one side, the woman's neck from the other, and suspended blood droplets remain in the narrow gap between them. The frame freezes the instant before the planned downward tilt, excluding almost all alley context while preserving the physical bite.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 전봇대 (The two people are pressed near it) — Only a shallow section of the side facing the pair remains visible behind them; used as A narrow blurred vertical trace behind the tightly framed profiles anchors them to the alley location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Faint streetlight illumination leaves the profiles in low-key desaturated contrast, with the airborne blood as the only restrained dark-crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S2sh3__bgfirst_bg.png",
  "asset_id": "1558005e-4252-463d-a063-0d6b0410c986",
  "input_asset_ids": [
   "1afb33b4-a07e-48e8-8b9a-36f6d70cd679",
   "3c03c327-57b1-423a-ae8f-b63ae5ae45c6"
  ]
 },
 "S2sh3": {
  "input_fingerprint": "fb59e52d0ee56066",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한국인 신원미상 남성의 입술이 효장동 피해 여성의 목덜미를 깊게 문 채, 검붉은 핏방울이 공중에 튄 정지 순간.\n\nLOCATION (lock): Beside a utility pole in a dim urban back alley at night, under faint streetlight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At neck height beside the pair, finish the dolly-in on an extremely tight opposing-profile composition: the man's lips and teeth enter from one side, the woman's neck from the other, and suspended blood droplets remain in the narrow gap between them. The frame freezes the instant before the planned downward tilt, excluding almost all alley context while preserving the physical bite.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 전봇대 (The two people are pressed near it) — Only a shallow section of the side facing the pair remains visible behind them; used as A narrow blurred vertical trace behind the tightly framed profiles anchors them to the alley location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Faint streetlight illumination leaves the profiles in low-key desaturated contrast, with the airborne blood as the only restrained dark-crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The unidentified man’s teeth remain embedded in the woman’s neck, leaving the neck torn and bloodied before she collapses.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 한국인 신원미상 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 자연스러운 짙은색 머리카락); 효장동 피해 여성 (한국인 성인 여성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한국인 신원미상 남성의 입술이 효장동 피해 여성의 목덜미를 깊게 문 채, 검붉은 핏방울이 공중에 튄 정지 순간.\n\nLOCATION (lock): Beside a utility pole in a dim urban back alley at night, under faint streetlight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At neck height beside the pair, finish the dolly-in on an extremely tight opposing-profile composition: the man's lips and teeth enter from one side, the woman's neck from the other, and suspended blood droplets remain in the narrow gap between them. The frame freezes the instant before the planned downward tilt, excluding almost all alley context while preserving the physical bite.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 전봇대 (The two people are pressed near it) — Only a shallow section of the side facing the pair remains visible behind them; used as A narrow blurred vertical trace behind the tightly framed profiles anchors them to the alley location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Faint streetlight illumination leaves the profiles in low-key desaturated contrast, with the airborne blood as the only restrained dark-crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The unidentified man’s teeth remain embedded in the woman’s neck, leaving the neck torn and bloodied before she collapses.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 한국인 신원미상 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 자연스러운 짙은색 머리카락); 효장동 피해 여성 (한국인 성인 여성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한국인 신원미상 남성의 입술이 효장동 피해 여성의 목덜미를 깊게 문 채, 검붉은 핏방울이 공중에 튄 정지 순간.\n\nLOCATION (lock): Beside a utility pole in a dim urban back alley at night, under faint streetlight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At neck height beside the pair, finish the dolly-in on an extremely tight opposing-profile composition: the man's lips and teeth enter from one side, the woman's neck from the other, and suspended blood droplets remain in the narrow gap between them. The frame freezes the instant before the planned downward tilt, excluding almost all alley context while preserving the physical bite.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 전봇대 (The two people are pressed near it) — Only a shallow section of the side facing the pair remains visible behind them; used as A narrow blurred vertical trace behind the tightly framed profiles anchors them to the alley location.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Faint streetlight illumination leaves the profiles in low-key desaturated contrast, with the airborne blood as the only restrained dark-crimson accent.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The unidentified man’s teeth remain embedded in the woman’s neck, leaving the neck torn and bloodied before she collapses.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 한국인 신원미상 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 자연스러운 짙은색 머리카락); 효장동 피해 여성 (한국인 성인 여성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S2sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S2sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 한국인 신원미상 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:824125>"
    },
    {
     "label": "CHARACTER REFERENCE — 효장동 피해 여성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:493451>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_도시_뒷골목_43d35b.png"
    },
    {
     "label": "CHARACTER REFERENCE — 한국인 신원미상 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:824125>"
    },
    {
     "label": "CHARACTER REFERENCE — 효장동 피해 여성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:493451>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "두 이미지 모두 남성이 여성의 목을 무는 핵심 행동과 핏방울을 누락했으나, B가 여성의 회색 스카프 등 의상 디테일을 더 가깝게 재현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "목을 무는 프롬프트의 주요 상황과 핏방울이 전혀 연출되지 않았으며, 여성의 스카프 색상도 레퍼런스와 다릅니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "두 이미지 모두 남성이 여성의 목을 무는 핵심 행동과 핏방울을 누락했으나, B가 여성의 회색 스카프 등 의상 디테일을 더 가깝게 재현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "목을 무는 프롬프트의 주요 상황과 핏방울이 전혀 연출되지 않았으며, 여성의 스카프 색상도 레퍼런스와 다릅니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "핵심 액션(목을 무는 모습, 핏방울)이 완전히 누락되었으나, 의상(회색 스카프)과 인물 디테일이 레퍼런스에 더 가깝습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "핵심 액션과 핏방울이 누락되었으며, 여성의 스카프 색상(빨간색)이 레퍼런스와 일치하지 않습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "핵심 액션(목을 무는 모습, 핏방울)이 완전히 누락되었으나, 의상(회색 스카프)과 인물 디테일이 레퍼런스에 더 가깝습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "핵심 액션과 핏방울이 누락되었으며, 여성의 스카프 색상(빨간색)이 레퍼런스와 일치하지 않습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 5,
     "B": 7
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 5,
   "B": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "두 이미지 모두 남성이 여성의 목을 무는 핵심 행동과 핏방울을 누락했으나, B가 여성의 회색 스카프 등 의상 디테일을 더 가깝게 재현했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "목을 무는 프롬프트의 주요 상황과 핏방울이 전혀 연출되지 않았으며, 여성의 스카프 색상도 레퍼런스와 다릅니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_도시_뒷골목_43d35b.png"
   },
   {
    "label": "CHARACTER REFERENCE — 한국인 신원미상 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:824125>"
   },
   {
    "label": "CHARACTER REFERENCE — 효장동 피해 여성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:493451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "남성의 입술이 여성의 목덜미를 물고 있지 않고 두 사람이 떨어져서 서로 바라보고만 있습니다.",
     "fix_en": "Close the gap completely so the man's lips and teeth are deeply embedded into the woman's neck."
    },
    {
     "issue_ko": "공중에 튀어 정지해 있는 검붉은 핏방울이 화면에 전혀 없습니다.",
     "fix_en": "Add suspended dark crimson blood droplets splattering in the air around the bite area."
    },
    {
     "issue_ko": "카메라 프레이밍이 남성의 입과 여성의 목덜미만 보이도록 극단적으로 밀착되지 않고 두 사람의 얼굴 전체가 화면에 담겨 있습니다.",
     "fix_en": "Zoom in significantly so that only the man's mouth enters from the left edge and the woman's neck enters from the right edge."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Lower the camera from eye level to neck height beside the pair while preserving all other visual attributes.",
     "issue_ko": "카메라가 목 높이가 아니라 두 인물의 얼굴과 눈높이에 놓여 있다."
    },
    {
     "fix_en": "Reposition the man so his profile enters from one side with his mouth physically pressed into the woman's exposed neck, rather than facing her face.",
     "issue_ko": "남성이 여성의 목을 무는 측면 배치가 아니라 두 사람이 얼굴을 마주 보는 배치로 연출되어 있다."
    },
    {
     "fix_en": "Dolly in to an extreme close-up that excludes most shoulders and faces and concentrates on the man's mouth, the woman's neck, and the narrow gap between them.",
     "issue_ko": "프레임이 너무 넓어 얼굴과 상반신이 많이 보이며, 요구된 입과 목 중심의 극도로 타이트한 클로즈업이 아니다."
    },
    {
     "fix_en": "Shift and defocus the utility pole so only a narrow blurred vertical trace remains behind the tightly framed profiles.",
     "issue_ko": "전봇대가 화면 중앙을 넓고 선명하게 차지해, 뒤편의 좁고 흐린 수직 흔적으로만 보여야 한다는 배경 배치와 다르다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Close the gap completely so the man's lips and teeth are deeply embedded into the woman's neck.\n- Add suspended dark crimson blood droplets splattering in the air around the bite area.\n- Zoom in significantly so that only the man's mouth enters from the left edge and the woman's neck enters from the right edge.\n- Lower the camera from eye level to neck height beside the pair while preserving all other visual attributes.\n- Reposition the man so his profile enters from one side with his mouth physically pressed into the woman's exposed neck, rather than facing her face.\n- Dolly in to an extreme close-up that excludes most shoulders and faces and concentrates on the man's mouth, the woman's neck, and the narrow gap between them.\n- Shift and defocus the utility pole so only a narrow blurred vertical trace remains behind the tightly framed profiles.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 '목 높이의 극단적인 입술-목덜미 클로즈업' 구도를 무시하고 인물의 얼굴 전체를 담았으며, 물어뜯는 동작과 핏방울이 전혀 연출되지 않았습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "입술과 목이 양옆에서 들어오는 극단적인 클로즈업 프레임은 매우 정확하게 구현했으나, 공중에 튄 핏방울이 붉은색이 아니며 살을 무는 묘사가 부족합니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 '목 높이의 극단적인 입술-목덜미 클로즈업' 구도를 무시하고 인물의 얼굴 전체를 담았으며, 물어뜯는 동작과 핏방울이 전혀 연출되지 않았습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "입술과 목이 양옆에서 들어오는 극단적인 클로즈업 프레임은 매우 정확하게 구현했으나, 공중에 튄 핏방울이 붉은색이 아니며 살을 무는 묘사가 부족합니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "목 높이의 극단적인 클로즈업 구도를 명확히 따랐으나, 지시된 검붉은 핏방울과 실제 무는 동작이 제대로 표현되지 않았습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "카메라 앵글이 목이 아닌 얼굴 높이에 맞춰져 구도 지시를 어겼으며, 공중의 핏방울이나 무는 동작 또한 전혀 나타나지 않았습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "목 높이의 극단적인 클로즈업 구도를 명확히 따랐으나, 지시된 검붉은 핏방울과 실제 무는 동작이 제대로 표현되지 않았습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "카메라 앵글이 목이 아닌 얼굴 높이에 맞춰져 구도 지시를 어겼으며, 공중의 핏방울이나 무는 동작 또한 전혀 나타나지 않았습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 12
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S2sh3__bgfirst_bg.png",
   "bg_asset_id": "1558005e-4252-463d-a063-0d6b0410c986",
   "bg_record_key": "S2sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "도시_뒷골목",
   "groupbg_asset_id": "3c03c327-57b1-423a-ae8f-b63ae5ae45c6"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S2sh5::variants": {
  "author_fp": "a2ebcffb8730e9ae",
  "author": {
   "variants": [
    {
     "approach_ko": "골목의 축을 정면으로 잡고 깊은 초점과 엄격한 원근으로 공허함을 강조한 저시점 와이드 숏.",
     "prompt_en": "Photograph from an austere, nearly axial low viewpoint, using a rectilinear wide-lens feel and strong one-point perspective. Hold deep focus from the textured foreground pavement to the distant utility pole, letting the converging alley lines and overwhelming negative space reduce the woman to a small but clearly human presence. Keep the sparse streetlight exposure restrained, with desaturated shadow detail and no visual embellishment, creating a severe, observational finality."
    },
    {
     "approach_ko": "전경을 부드럽게 흘리고 전봇대와 쓰러진 여성에 시선을 모으는 비대칭적이고 층위감 있는 와이드 숏.",
     "prompt_en": "Use an off-axis low composition with the alley receding diagonally, placing the distant pole and woman away from the visual center while the empty foreground dominates. Let the nearest pavement fall gently out of focus, drawing attention through layered darkness toward a narrow plane of clarity around the far spatial anchor and collapsed figure. Shape the faint streetlight into restrained islands of visibility amid broad, soft black areas, making the withdrawal feel hushed and emotionally remote."
    }
   ]
  },
  "reused": false
 },
 "S2sh5": {
  "input_fingerprint": "9bb332ef676fa8f9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 어두운 골목길 바닥에 효장동 피해 여성이 미동 없이 쓰러져 있는 적막하고 넓은 앵글.\n\nLOCATION (lock): On the deserted pavement of a dark urban back alley, beneath sparse street lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From a low position over the alley ground, complete the dolly-out into a broad, static-feeling view aimed back toward the utility pole. Place the motionless woman as a small figure near the pole while the empty alley occupies most of the frame, allowing the withdrawal itself to create the final silence.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 어두운 골목길 (Empty except for the motionless woman); used as Primary negative space extending from the low foreground toward the distant body; 전봇대 (Fixed at the far end of the composition) — Its side faces the withdrawing camera, with the woman positioned near its base; used as Distant spatial anchor locating where the woman fell.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The faint streetlights provide sparse low-key visibility across a dark, desaturated night alley.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her entire body is collapsed slack and motionless on the alley floor, with no part of her body supporting a standing or seated posture; the source does not specify whether she faces upward or downward or the exact placement of her limbs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The woman's entire body is collapsed slack and motionless on the alley floor, with no part of her body supporting a standing or seated posture, retaining the torn, bloodied wound on her neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 효장동 피해 여성 (한국인 성인 여성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 어두운 골목길 바닥에 효장동 피해 여성이 미동 없이 쓰러져 있는 적막하고 넓은 앵글.\n\nLOCATION (lock): On the deserted pavement of a dark urban back alley, beneath sparse street lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From a low position over the alley ground, complete the dolly-out into a broad, static-feeling view aimed back toward the utility pole. Place the motionless woman as a small figure near the pole while the empty alley occupies most of the frame, allowing the withdrawal itself to create the final silence.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 어두운 골목길 (Empty except for the motionless woman); used as Primary negative space extending from the low foreground toward the distant body; 전봇대 (Fixed at the far end of the composition) — Its side faces the withdrawing camera, with the woman positioned near its base; used as Distant spatial anchor locating where the woman fell.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The faint streetlights provide sparse low-key visibility across a dark, desaturated night alley.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her entire body is collapsed slack and motionless on the alley floor, with no part of her body supporting a standing or seated posture; the source does not specify whether she faces upward or downward or the exact placement of her limbs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The woman's entire body is collapsed slack and motionless on the alley floor, with no part of her body supporting a standing or seated posture, retaining the torn, bloodied wound on her neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 효장동 피해 여성 (한국인 성인 여성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph from an austere, nearly axial low viewpoint, using a rectilinear wide-lens feel and strong one-point perspective. Hold deep focus from the textured foreground pavement to the distant utility pole, letting the converging alley lines and overwhelming negative space reduce the woman to a small but clearly human presence. Keep the sparse streetlight exposure restrained, with desaturated shadow detail and no visual embellishment, creating a severe, observational finality.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 어두운 골목길 바닥에 효장동 피해 여성이 미동 없이 쓰러져 있는 적막하고 넓은 앵글.\n\nLOCATION (lock): On the deserted pavement of a dark urban back alley, beneath sparse street lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From a low position over the alley ground, complete the dolly-out into a broad, static-feeling view aimed back toward the utility pole. Place the motionless woman as a small figure near the pole while the empty alley occupies most of the frame, allowing the withdrawal itself to create the final silence.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 어두운 골목길 (Empty except for the motionless woman); used as Primary negative space extending from the low foreground toward the distant body; 전봇대 (Fixed at the far end of the composition) — Its side faces the withdrawing camera, with the woman positioned near its base; used as Distant spatial anchor locating where the woman fell.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The faint streetlights provide sparse low-key visibility across a dark, desaturated night alley.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her entire body is collapsed slack and motionless on the alley floor, with no part of her body supporting a standing or seated posture; the source does not specify whether she faces upward or downward or the exact placement of her limbs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The woman's entire body is collapsed slack and motionless on the alley floor, with no part of her body supporting a standing or seated posture, retaining the torn, bloodied wound on her neck.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 효장동 피해 여성 (한국인 성인 여성 얼굴, 자연스러운 피부색, 구체적 머리 형태 미상) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse an off-axis low composition with the alley receding diagonally, placing the distant pole and woman away from the visual center while the empty foreground dominates. Let the nearest pavement fall gently out of focus, drawing attention through layered darkness toward a narrow plane of clarity around the far spatial anchor and collapsed figure. Shape the faint streetlight into restrained islands of visibility amid broad, soft black areas, making the withdrawal feel hushed and emotionally remote.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 넓은 앵글과 피사체 배치를 잘 구현했으며, 참조 이미지의 전봇대(콘크리트 질감, 벽보 흔적)와 붉은 벽돌 배경을 충실히 재현했습니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "프레이밍은 지시문에 부합하나, 전봇대 하단의 노란 줄무늬와 주변 건물의 형태가 참조 이미지의 고정된 배경 요소와 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S2sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 효장동 피해 여성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1147652>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "여성의 상체가 전봇대에 기대어 앉은 자세를 취하고 있어, 바닥에 완전히 쓰러져 있어야 한다는 지시를 위반했습니다.",
     "fix_en": "Drop the woman's upper body so she lies completely flat on the alley ground, removing any seated or propped-up posture."
    },
    {
     "issue_ko": "화면 중앙 전봇대 표면에 텍스트가 포함되어 있습니다.",
     "fix_en": "Remove the text from the surface of the utility pole."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reposition the woman into a fully slack, gravity-supported collapse with her entire body, including torso, head, and limbs, resting on the alley floor near the pole; preserve the camera and all other framing.",
     "issue_ko": "여성이 전봇대에 기대어 상체와 머리를 세운 좌식 자세로 보여, 전신이 골목 바닥에 힘없이 완전히 쓰러져 있어야 한다는 고정 포즈와 다릅니다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Drop the woman's upper body so she lies completely flat on the alley ground, removing any seated or propped-up posture.\n- Remove the text from the surface of the utility pole.\n- Reposition the woman into a fully slack, gravity-supported collapse with her entire body, including torso, head, and limbs, resting on the alley floor near the pole; preserve the camera and all other framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "여성이 바닥에 완전히 쓰러져 있지 않고 전봇대에 기대어 앉아 있어 '서거나 앉은 자세가 아닐 것'이라는 자세 지침을 위반했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "낮고 넓은 앵글과 텅 빈 골목길의 구도를 잘 구현했으며, 지침대로 여성이 바닥에 완전히 미동 없이 쓰러진 자세를 정확히 표현했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "여성이 바닥에 완전히 쓰러져 있지 않고 전봇대에 기대어 앉아 있어 '서거나 앉은 자세가 아닐 것'이라는 자세 지침을 위반했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "낮고 넓은 앵글과 텅 빈 골목길의 구도를 잘 구현했으며, 지침대로 여성이 바닥에 완전히 미동 없이 쓰러진 자세를 정확히 표현했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트의 지시대로 여성이 골목길 바닥에 완전히 쓰러져 있는 자세와 넓은 앵글의 구도를 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "여성이 전봇대에 기대어 앉은 자세를 취하고 있어, 신체의 어떤 부분도 앉거나 서 있는 자세를 지탱해서는 안 된다는 핵심 지침을 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "프롬프트의 지시대로 여성이 골목길 바닥에 완전히 쓰러져 있는 자세와 넓은 앵글의 구도를 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "여성이 전봇대에 기대어 앉은 자세를 취하고 있어, 신체의 어떤 부분도 앉거나 서 있는 자세를 지탱해서는 안 된다는 핵심 지침을 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S2sh3"
  }
 },
 "S3sh3::bgfirst_bg": {
  "input_fingerprint": "f2b5e0132350eb4a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 무거운 안색을 띤 혜수가 폴리스 라인 앞 경찰 무리 한가운데서 무전기를 쥔 채 꼿꼿이 서 있는 상반신 정면.\n\nLOCATION (lock): In a cordoned forest clearing at the mountain crime scene, directly behind the police line among investigating officers.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Presented as regional-news footage, the camera finishes its chest-height frontal dolly-in at medium distance with 혜수 centered from the waist up, her radio held clearly within the lower portion of frame. Police remain compressed along both lateral edges and the police line crosses the rear field, turning the otherwise direct command portrait into a crowded operational image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 폴리스 라인 (Set around the investigation area) — The printed warning side faces the news camera and runs laterally behind 혜수; used as A visible boundary behind 혜수 that separates command personnel from the recovery activity; 시신 수거 현장 (Police are examining the scene and recovering bodies); used as Layered background activity establishing the active crime-scene investigation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination is rendered with subdued, desaturated contrast and the slightly flattened tonal response of news footage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 무거운 안색을 띤 혜수가 폴리스 라인 앞 경찰 무리 한가운데서 무전기를 쥔 채 꼿꼿이 서 있는 상반신 정면.\n\nLOCATION (lock): In a cordoned forest clearing at the mountain crime scene, directly behind the police line among investigating officers.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Presented as regional-news footage, the camera finishes its chest-height frontal dolly-in at medium distance with 혜수 centered from the waist up, her radio held clearly within the lower portion of frame. Police remain compressed along both lateral edges and the police line crosses the rear field, turning the otherwise direct command portrait into a crowded operational image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 폴리스 라인 (Set around the investigation area) — The printed warning side faces the news camera and runs laterally behind 혜수; used as A visible boundary behind 혜수 that separates command personnel from the recovery activity; 시신 수거 현장 (Police are examining the scene and recovering bodies); used as Layered background activity establishing the active crime-scene investigation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination is rendered with subdued, desaturated contrast and the slightly flattened tonal response of news footage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S3sh3__bgfirst_bg.png",
  "asset_id": "450c0c33-1b9b-4052-a153-4710bb23d458",
  "input_asset_ids": [
   "8aefbb80-cdc2-4b76-962d-85efdc0e29cd",
   "ebe5401b-668a-4420-9d48-1057e290897c"
  ]
 },
 "S3sh3": {
  "input_fingerprint": "8755a2923c188370",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무거운 안색을 띤 혜수가 폴리스 라인 앞 경찰 무리 한가운데서 무전기를 쥔 채 꼿꼿이 서 있는 상반신 정면.\n\nLOCATION (lock): In a cordoned forest clearing at the mountain crime scene, directly behind the police line among investigating officers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Presented as regional-news footage, the camera finishes its chest-height frontal dolly-in at medium distance with 혜수 centered from the waist up, her radio held clearly within the lower portion of frame. Police remain compressed along both lateral edges and the police line crosses the rear field, turning the otherwise direct command portrait into a crowded operational image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 폴리스 라인 (Set around the investigation area) — The printed warning side faces the news camera and runs laterally behind 혜수; used as A visible boundary behind 혜수 that separates command personnel from the recovery activity; 시신 수거 현장 (Police are examining the scene and recovering bodies); used as Layered background activity establishing the active crime-scene investigation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination is rendered with subdued, desaturated contrast and the slightly flattened tonal response of news footage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무거운 안색을 띤 혜수가 폴리스 라인 앞 경찰 무리 한가운데서 무전기를 쥔 채 꼿꼿이 서 있는 상반신 정면.\n\nLOCATION (lock): In a cordoned forest clearing at the mountain crime scene, directly behind the police line among investigating officers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Presented as regional-news footage, the camera finishes its chest-height frontal dolly-in at medium distance with 혜수 centered from the waist up, her radio held clearly within the lower portion of frame. Police remain compressed along both lateral edges and the police line crosses the rear field, turning the otherwise direct command portrait into a crowded operational image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 폴리스 라인 (Set around the investigation area) — The printed warning side faces the news camera and runs laterally behind 혜수; used as A visible boundary behind 혜수 that separates command personnel from the recovery activity; 시신 수거 현장 (Police are examining the scene and recovering bodies); used as Layered background activity establishing the active crime-scene investigation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination is rendered with subdued, desaturated contrast and the slightly flattened tonal response of news footage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무거운 안색을 띤 혜수가 폴리스 라인 앞 경찰 무리 한가운데서 무전기를 쥔 채 꼿꼿이 서 있는 상반신 정면.\n\nLOCATION (lock): In a cordoned forest clearing at the mountain crime scene, directly behind the police line among investigating officers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Presented as regional-news footage, the camera finishes its chest-height frontal dolly-in at medium distance with 혜수 centered from the waist up, her radio held clearly within the lower portion of frame. Police remain compressed along both lateral edges and the police line crosses the rear field, turning the otherwise direct command portrait into a crowded operational image.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 폴리스 라인 (Set around the investigation area) — The printed warning side faces the news camera and runs laterally behind 혜수; used as A visible boundary behind 혜수 that separates command personnel from the recovery activity; 시신 수거 현장 (Police are examining the scene and recovering bodies); used as Layered background activity establishing the active crime-scene investigation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination is rendered with subdued, desaturated contrast and the slightly flattened tonal response of news footage.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S3sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S3sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_산림_범행지_a8cfa8.png"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "지정된 샷 사이즈와 배경 배치는 잘 따랐으나, 레퍼런스의 가죽 재킷 대신 임의의 정장 재킷을 착용하여 의상 고증에서 벗어났습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "레퍼런스의 가죽 재킷 의상을 정확히 반영했으며, 명시된 상반신 미디엄 샷 구도와 폴리스 라인, 수사 현장 배경을 충실히 구현했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "지정된 샷 사이즈와 배경 배치는 잘 따랐으나, 레퍼런스의 가죽 재킷 대신 임의의 정장 재킷을 착용하여 의상 고증에서 벗어났습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "레퍼런스의 가죽 재킷 의상을 정확히 반영했으며, 명시된 상반신 미디엄 샷 구도와 폴리스 라인, 수사 현장 배경을 충실히 구현했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "레퍼런스의 복장(가죽 재킷)을 완벽하게 유지했으며, 명시된 미디엄 샷 구도와 폴리스 라인 뒤의 수사 현장 묘사가 지시사항과 잘 일치합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "구도와 분위기는 비슷하나 인물의 복장(파란색 블레이저)이 레퍼런스와 전혀 다르며, 배경 인물들의 형태와 폴리스 라인의 묘사가 다소 어색합니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "레퍼런스의 복장(가죽 재킷)을 완벽하게 유지했으며, 명시된 미디엄 샷 구도와 폴리스 라인 뒤의 수사 현장 묘사가 지시사항과 잘 일치합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "구도와 분위기는 비슷하나 인물의 복장(파란색 블레이저)이 레퍼런스와 전혀 다르며, 배경 인물들의 형태와 폴리스 라인의 묘사가 다소 어색합니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 9,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "지정된 샷 사이즈와 배경 배치는 잘 따랐으나, 레퍼런스의 가죽 재킷 대신 임의의 정장 재킷을 착용하여 의상 고증에서 벗어났습니다."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "레퍼런스의 가죽 재킷 의상을 정확히 반영했으며, 명시된 상반신 미디엄 샷 구도와 폴리스 라인, 수사 현장 배경을 충실히 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_산림_범행지_a8cfa8.png"
   },
   {
    "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:900092>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "가죽 재킷의 양팔이 모두 아래로 늘어져 있는데, 가슴 중앙에서 정체불명의 세 번째 손이 나와 무전기를 쥐고 있습니다.",
     "fix_en": "Remove the extra hand from the center of the chest and pose her actual right arm to hold the walkie-talkie."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the extra hand from the center of the chest and pose her actual right arm to hold the walkie-talkie.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "뉴스 보도 화면 형태의 미디엄 샷과 인물 및 배경 배치를 훌륭하게 구현했으나, 무전기를 쥔 오른손이 해부학적으로 불가능한 형태(오른팔에 왼손 결합)를 띠고 있어 치명적인 감점 요인이 되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "A와 시각적으로 동일한 이미지로, 지시된 구도와 톤은 잘 반영되었으나 동일하게 무전기를 쥔 손의 심각한 해부학적 오류가 존재하여 높은 점수를 받을 수 없습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "뉴스 보도 화면 형태의 미디엄 샷과 인물 및 배경 배치를 훌륭하게 구현했으나, 무전기를 쥔 오른손이 해부학적으로 불가능한 형태(오른팔에 왼손 결합)를 띠고 있어 치명적인 감점 요인이 되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "A와 시각적으로 동일한 이미지로, 지시된 구도와 톤은 잘 반영되었으나 동일하게 무전기를 쥔 손의 심각한 해부학적 오류가 존재하여 높은 점수를 받을 수 없습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지정된 미디엄 샷 프레이밍, 무전기를 쥔 혜수의 상반신 정면 구도, 양옆의 경찰 무리, 그리고 폴리스 라인 뒤의 시신 수거 현장까지 모든 지시사항을 완벽하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "후보 A와 완전히 동일한 이미지로, 프롬프트의 카메라 앵글, 인물 배치 및 배경 디테일을 정확도 높게 반영했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "지정된 미디엄 샷 프레이밍, 무전기를 쥔 혜수의 상반신 정면 구도, 양옆의 경찰 무리, 그리고 폴리스 라인 뒤의 시신 수거 현장까지 모든 지시사항을 완벽하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "후보 A와 완전히 동일한 이미지로, 프롬프트의 카메라 앵글, 인물 배치 및 배경 디테일을 정확도 높게 반영했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S3sh3__bgfirst_bg.png",
   "bg_asset_id": "450c0c33-1b9b-4052-a153-4710bb23d458",
   "bg_record_key": "S3sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "산림_범행지",
   "groupbg_asset_id": "ebe5401b-668a-4420-9d48-1057e290897c"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S4sh2::variants": {
  "author_fp": "463760bd5943c20f",
  "author": {
   "variants": [
    {
     "approach_ko": "깊은 초점과 절제된 원근으로 옥탑방의 고립과 삭막한 건축성을 강조한 관찰적 정지 화면.",
     "prompt_en": "Use restrained wide-angle perspective and deep focus for an austere architectural tableau. Hold the rooftop room as a solid central mass, with the oblique rooftop edge forming a firm foreground boundary and the rusted door remaining a clear but subordinate point of attention. Preserve subdued contrast, desaturated color, and tactile detail in every weathered surface, creating a stark, observational stillness."
    },
    {
     "approach_ko": "압축된 원근과 층층이 겹치는 옥상 면으로 거리감과 고립감을 강화한 묵직한 정지 화면.",
     "prompt_en": "Use a mildly compressed lens feel to stack the foreground rooftop boundary, the surrounding roof plane, and the rooftop room into distinct visual layers while retaining the required wide reveal. Let broad areas of unoccupied rooftop create severe negative space around the central structure, with the angled door-bearing face drawing the eye quietly toward the rusted metal door. Render the daytime ambient light with low-key tonal density, restrained highlights, and a heavy, desolate calm."
    }
   ]
  },
  "reused": false
 },
 "S4sh2": {
  "input_fingerprint": "a3633db5cd10a911",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오래된 다세대 빌라 옥상 위, 녹슨 철문이 달린 낡은 옥탑방 외부가 삭막하게 담긴 넓은 정지 구도.\n\nLOCATION (lock): On the exposed rooftop of an aging multi-unit villa, facing the weathered rooftop room and its rusted metal door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From above and outside the rooftop edge, hold a wide, slightly elevated static view that reveals the old rooftop room as the central mass and keeps enough surrounding rooftop visible to establish its isolation. Angle the structure so the rusted iron door is legible on one face without allowing it to overwhelm the broader rooftop layout.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 낡은 옥탑방 외부 (Old and worn) — The door-bearing exterior face and one adjoining side are visible from the elevated oblique viewpoint; used as Central architectural subject within the wide rooftop reveal; 녹슨 철문 (Rusted; its open or closed position is not specified) — Its broad exterior face is turned toward the camera at a slight angle; used as Secondary focal point on the camera-facing wall of the rooftop room; 옥상 가장자리 (Part of the old multi-unit villa rooftop) — Seen obliquely beneath the camera, leading toward the rooftop room; used as Foreground boundary that frames the elevated approach and locates the room on the villa roof.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is treated in a restrained, desaturated register with low-key tonal weight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오래된 다세대 빌라 옥상 위, 녹슨 철문이 달린 낡은 옥탑방 외부가 삭막하게 담긴 넓은 정지 구도.\n\nLOCATION (lock): On the exposed rooftop of an aging multi-unit villa, facing the weathered rooftop room and its rusted metal door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From above and outside the rooftop edge, hold a wide, slightly elevated static view that reveals the old rooftop room as the central mass and keeps enough surrounding rooftop visible to establish its isolation. Angle the structure so the rusted iron door is legible on one face without allowing it to overwhelm the broader rooftop layout.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 낡은 옥탑방 외부 (Old and worn) — The door-bearing exterior face and one adjoining side are visible from the elevated oblique viewpoint; used as Central architectural subject within the wide rooftop reveal; 녹슨 철문 (Rusted; its open or closed position is not specified) — Its broad exterior face is turned toward the camera at a slight angle; used as Secondary focal point on the camera-facing wall of the rooftop room; 옥상 가장자리 (Part of the old multi-unit villa rooftop) — Seen obliquely beneath the camera, leading toward the rooftop room; used as Foreground boundary that frames the elevated approach and locates the room on the villa roof.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is treated in a restrained, desaturated register with low-key tonal weight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse restrained wide-angle perspective and deep focus for an austere architectural tableau. Hold the rooftop room as a solid central mass, with the oblique rooftop edge forming a firm foreground boundary and the rusted door remaining a clear but subordinate point of attention. Preserve subdued contrast, desaturated color, and tactile detail in every weathered surface, creating a stark, observational stillness.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 오래된 다세대 빌라 옥상 위, 녹슨 철문이 달린 낡은 옥탑방 외부가 삭막하게 담긴 넓은 정지 구도.\n\nLOCATION (lock): On the exposed rooftop of an aging multi-unit villa, facing the weathered rooftop room and its rusted metal door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From above and outside the rooftop edge, hold a wide, slightly elevated static view that reveals the old rooftop room as the central mass and keeps enough surrounding rooftop visible to establish its isolation. Angle the structure so the rusted iron door is legible on one face without allowing it to overwhelm the broader rooftop layout.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 낡은 옥탑방 외부 (Old and worn) — The door-bearing exterior face and one adjoining side are visible from the elevated oblique viewpoint; used as Central architectural subject within the wide rooftop reveal; 녹슨 철문 (Rusted; its open or closed position is not specified) — Its broad exterior face is turned toward the camera at a slight angle; used as Secondary focal point on the camera-facing wall of the rooftop room; 옥상 가장자리 (Part of the old multi-unit villa rooftop) — Seen obliquely beneath the camera, leading toward the rooftop room; used as Foreground boundary that frames the elevated approach and locates the room on the villa roof.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is treated in a restrained, desaturated register with low-key tonal weight.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a mildly compressed lens feel to stack the foreground rooftop boundary, the surrounding roof plane, and the rooftop room into distinct visual layers while retaining the required wide reveal. Let broad areas of unoccupied rooftop create severe negative space around the central structure, with the angled door-bearing face drawing the eye quietly toward the rusted metal door. Render the daytime ambient light with low-key tonal density, restrained highlights, and a heavy, desolate calm.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 6,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "텍스트의 '녹슨 철문'과 구도를 사실적으로 구현했으나, 지정된 구조물(Structure Look)의 형태는 반영하지 못했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "지정된 구조물 형태를 누락했으며, 전선과 빨랫줄이 뭉개지는 물리적 왜곡(Hard Violation)이 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L04B02.png"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/background_chain/seed_bg_rooftop_villa_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "옥탑방 건물의 형태, 창문, 문 위치 및 재질이 'STRUCTURE LOOK' 참고 이미지를 무시하고 'LOCATION' 참고 이미지의 구조를 그대로 복사했습니다.",
     "fix_en": "Replace the rooftop room's shape, window placement, door position, and materials to exactly match the structure shown in the STRUCTURE LOOK reference."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reframe so the rooftop room forms the central architectural mass while preserving the wide rooftop reveal and all existing scene content.",
     "issue_ko": "옥탑방이 화면 오른쪽 상단에 치우쳐 있어, 넓은 구도 속 중앙 건축 매스로 배치하라는 계약과 다릅니다."
    },
    {
     "fix_en": "Shift the camera to a more oblique elevated axis so the door-bearing face and one adjoining side of the rooftop room are both clearly visible, preserving the wide scale.",
     "issue_ko": "시점이 옥탑방의 문 달린 면을 거의 정면으로만 보여 주어, 인접한 한쪽 측면이 명확히 드러나지 않습니다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the rooftop room's shape, window placement, door position, and materials to exactly match the structure shown in the STRUCTURE LOOK reference.\n- Reframe so the rooftop room forms the central architectural mass while preserving the wide rooftop reveal and all existing scene content.\n- Shift the camera to a more oblique elevated axis so the door-bearing face and one adjoining side of the rooftop room are both clearly visible, preserving the wide scale.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 옥탑방의 재질(흰색 패널)과 프롬프트에 명시된 '녹슨 철문'을 레퍼런스에 맞춰 매우 충실하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트의 핵심인 '녹슨 철문'이 반영되지 않았으며, 구조물의 외벽이 레퍼런스의 패널이 아닌 콘크리트로 임의 변경되어 룩 지침을 크게 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 옥탑방의 재질(흰색 패널)과 프롬프트에 명시된 '녹슨 철문'을 레퍼런스에 맞춰 매우 충실하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트의 핵심인 '녹슨 철문'이 반영되지 않았으며, 구조물의 외벽이 레퍼런스의 패널이 아닌 콘크리트로 임의 변경되어 룩 지침을 크게 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "텍스트가 명시한 '녹슨 철문'을 정확히 묘사했으며, 구조물 레퍼런스의 흰색 패널 재질을 충실히 유지했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "구조물이 레퍼런스(흰색 패널)와 전혀 다른 콘크리트/벽돌 재질로 변형되었고, 프롬프트의 '녹슨 철문' 요구사항도 누락되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "텍스트가 명시한 '녹슨 철문'을 정확히 묘사했으며, 구조물 레퍼런스의 흰색 패널 재질을 충실히 유지했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "구조물이 레퍼런스(흰색 패널)와 전혀 다른 콘크리트/벽돌 재질로 변형되었고, 프롬프트의 '녹슨 철문' 요구사항도 누락되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "플레이트+seed만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only",
  "plate_select": {
   "candidates": {
    "A": "L04B01",
    "B": "L04B02",
    "C": "L04B03"
   },
   "assigned": "L04B02",
   "choice": "Candidate B",
   "confident": true,
   "reason_ko": "원문에서 '넓은 정지 구도'로 옥탑방 외부 전체와 옥상의 삭막한 풍경을 담기를 요구하고 있으므로, 더 넓은 화각으로 공간 전체를 보여주는 Candidate B가 가장 적합합니다.",
   "kept": "L04B02"
  }
 },
 "S5sh6::bgfirst_bg": {
  "input_fingerprint": "431f7a73c9975e9b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 식탁 위에 놓인 강민숙의 맨 손목 안쪽에 새겨진 작고 기하학적인 원형 표식이 선명하게 드러난 클로즈업.\n\nLOCATION (lock): At the dining table inside the rooftop apartment’s compact living area, lit by daylight filtering through a small window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Beside the table at seated chest height, lower slightly at the end of the dolly-in and look down onto 강민숙's inner wrist resting across the tabletop. The wrist occupies most of the central frame without exceeding a natural human scale, while the small geometric circle remains sharp and the surrounding hand and table edge provide context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 식탁 상판 (The bare wrist rests on it); used as Simple surrounding plane that isolates the wrist and supplies a clear scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The sunlight admitted through the small living-room window gives the wrist natural but restrained illumination within the otherwise desaturated interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 식탁 위에 놓인 강민숙의 맨 손목 안쪽에 새겨진 작고 기하학적인 원형 표식이 선명하게 드러난 클로즈업.\n\nLOCATION (lock): At the dining table inside the rooftop apartment’s compact living area, lit by daylight filtering through a small window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Beside the table at seated chest height, lower slightly at the end of the dolly-in and look down onto 강민숙's inner wrist resting across the tabletop. The wrist occupies most of the central frame without exceeding a natural human scale, while the small geometric circle remains sharp and the surrounding hand and table edge provide context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 식탁 상판 (The bare wrist rests on it); used as Simple surrounding plane that isolates the wrist and supplies a clear scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The sunlight admitted through the small living-room window gives the wrist natural but restrained illumination within the otherwise desaturated interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S5sh6__bgfirst_bg.png",
  "asset_id": "cdebd20d-acf3-47b7-a623-d8a99164370f",
  "input_asset_ids": [
   "2c335c06-1591-4c1b-8269-ae935dc092e0",
   "6b72d257-47fa-4920-835d-132491d08d88"
  ]
 },
 "S5sh6": {
  "input_fingerprint": "37d5390b136df455",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 식탁 위에 놓인 강민숙의 맨 손목 안쪽에 새겨진 작고 기하학적인 원형 표식이 선명하게 드러난 클로즈업.\n\nLOCATION (lock): At the dining table inside the rooftop apartment’s compact living area, lit by daylight filtering through a small window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Beside the table at seated chest height, lower slightly at the end of the dolly-in and look down onto 강민숙's inner wrist resting across the tabletop. The wrist occupies most of the central frame without exceeding a natural human scale, while the small geometric circle remains sharp and the surrounding hand and table edge provide context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 식탁 상판 (The bare wrist rests on it); used as Simple surrounding plane that isolates the wrist and supplies a clear scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The sunlight admitted through the small living-room window gives the wrist natural but restrained illumination within the otherwise desaturated interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A small circular mark is permanently visible on the inside of Minsuk’s bare wrist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 식탁 위에 놓인 강민숙의 맨 손목 안쪽에 새겨진 작고 기하학적인 원형 표식이 선명하게 드러난 클로즈업.\n\nLOCATION (lock): At the dining table inside the rooftop apartment’s compact living area, lit by daylight filtering through a small window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Beside the table at seated chest height, lower slightly at the end of the dolly-in and look down onto 강민숙's inner wrist resting across the tabletop. The wrist occupies most of the central frame without exceeding a natural human scale, while the small geometric circle remains sharp and the surrounding hand and table edge provide context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 식탁 상판 (The bare wrist rests on it); used as Simple surrounding plane that isolates the wrist and supplies a clear scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The sunlight admitted through the small living-room window gives the wrist natural but restrained illumination within the otherwise desaturated interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A small circular mark is permanently visible on the inside of Minsuk’s bare wrist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 식탁 위에 놓인 강민숙의 맨 손목 안쪽에 새겨진 작고 기하학적인 원형 표식이 선명하게 드러난 클로즈업.\n\nLOCATION (lock): At the dining table inside the rooftop apartment’s compact living area, lit by daylight filtering through a small window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Beside the table at seated chest height, lower slightly at the end of the dolly-in and look down onto 강민숙's inner wrist resting across the tabletop. The wrist occupies most of the central frame without exceeding a natural human scale, while the small geometric circle remains sharp and the surrounding hand and table edge provide context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 식탁 상판 (The bare wrist rests on it); used as Simple surrounding plane that isolates the wrist and supplies a clear scale reference.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The sunlight admitted through the small living-room window gives the wrist natural but restrained illumination within the otherwise desaturated interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A small circular mark is permanently visible on the inside of Minsuk’s bare wrist.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S5sh6__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S5sh6.png"
    },
    {
     "label": "CHARACTER REFERENCE — 강민숙: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761261>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B05.png"
    },
    {
     "label": "CHARACTER REFERENCE — 강민숙: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:761261>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 프레이밍과 맨 손목, 단순한 원형 표식을 잘 구현했으며 배경의 장소 디테일이 참조 이미지와 매우 정확하게 일치합니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "손목의 표식이 지나치게 복잡하게 묘사되었고, 지시되지 않은 소품(밥그릇)이 추가되었으며 팔목에 옷소매가 노출되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 프레이밍과 맨 손목, 단순한 원형 표식을 잘 구현했으며 배경의 장소 디테일이 참조 이미지와 매우 정확하게 일치합니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "손목의 표식이 지나치게 복잡하게 묘사되었고, 지시되지 않은 소품(밥그릇)이 추가되었으며 팔목에 옷소매가 노출되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "식탁에 자연스럽게 놓인 손목, 기하학적 원형 표식, 그리고 레퍼런스와 일치하는 의상 소매까지 지시사항을 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "손이 식탁에 완전히 밀착되지 않고 들려 있으며, 기하학적 형태가 부족한 단순한 원형 표식과 왜곡된 구도가 아쉽습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "식탁에 자연스럽게 놓인 손목, 기하학적 원형 표식, 그리고 레퍼런스와 일치하는 의상 소매까지 지시사항을 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "손이 식탁에 완전히 밀착되지 않고 들려 있으며, 기하학적 형태가 부족한 단순한 원형 표식과 왜곡된 구도가 아쉽습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 13
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "totals": {
   "A": 12,
   "B": 13
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "지시된 프레이밍과 맨 손목, 단순한 원형 표식을 잘 구현했으며 배경의 장소 디테일이 참조 이미지와 매우 정확하게 일치합니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "손목의 표식이 지나치게 복잡하게 묘사되었고, 지시되지 않은 소품(밥그릇)이 추가되었으며 팔목에 옷소매가 노출되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B05.png"
   },
   {
    "label": "CHARACTER REFERENCE — 강민숙: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761261>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트에 언급되지 않은 밥그릇과 젓가락이 배경의 식탁 위에 임의로 추가되었습니다.",
     "fix_en": "Remove the bowl and chopsticks from the background so the table surface is empty except for the arm."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the bowl and chopsticks from the background so the table surface is empty except for the arm.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "손목과 기하학적 표식을 정확히 클로즈업했으며, 단순한 배경 평면을 유지하라는 지시를 불필요한 사물 없이 완벽하게 따랐습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "전반적인 구도와 세부 묘사는 우수하나, 프롬프트나 레퍼런스에 없는 밥그릇이 추가되어 '단순한 주변 평면'이라는 의도를 약간 저해합니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "손목과 기하학적 표식을 정확히 클로즈업했으며, 단순한 배경 평면을 유지하라는 지시를 불필요한 사물 없이 완벽하게 따랐습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "전반적인 구도와 세부 묘사는 우수하나, 프롬프트나 레퍼런스에 없는 밥그릇이 추가되어 '단순한 주변 평면'이라는 의도를 약간 저해합니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 인서트 클로즈업 구도와 손목 안쪽의 기하학적 표식을 정확히 묘사했으며, 배경을 단순한 평면으로 잘 유지했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트에 지시되지 않은 임의의 사물(밥그릇과 젓가락)이 추가되어 하드 위반에 해당합니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 인서트 클로즈업 구도와 손목 안쪽의 기하학적 표식을 정확히 묘사했으며, 배경을 단순한 평면으로 잘 유지했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트에 지시되지 않은 임의의 사물(밥그릇과 젓가락)이 추가되어 하드 위반에 해당합니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 16
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S5sh6__bgfirst_bg.png",
   "bg_asset_id": "cdebd20d-acf3-47b7-a623-d8a99164370f",
   "bg_record_key": "S5sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B05",
   "choice": "Candidate E",
   "confident": true,
   "reason_ko": "\"식탁 위에 놓인\"이라는 지문에서 식탁(dining table)이 중심이 되는 거실 겸 주방 공간이 필요함을 알 수 있습니다.",
   "kept": "L06B05"
  }
 },
 "S5sh8::variants": {
  "author_fp": "a6ee87a47de80ccf",
  "author": {
   "variants": [
    {
     "approach_ko": "얕은 초점과 부드러운 측면광으로 어깨에 기대는 수리영의 미세한 표정과 접촉의 온기를 강조한다.",
     "prompt_en": "Use an intimate portrait-lens feel with very shallow focus, holding Su-riyeong’s half-closed eyes, faint smile, and the point where her nose presses into Minsuk’s shoulder in delicate clarity. Let Minsuk’s foreground profile and shoulder fall gently softer, creating layered closeness while the entry interior dissolves into restrained, desaturated context. Shape the existing window daylight softly across Su-riyeong’s face, preserving natural skin and fabric texture without polish or sentimentality."
    },
    {
     "approach_ko": "보다 가까운 관찰자적 시선과 넓은 초점층으로 두 사람의 얼굴·어깨 접촉을 담담하고 현실적으로 포착한다.",
     "prompt_en": "Photograph the close-up with a natural, slightly wider observational lens feel, keeping the spatial relationship between Minsuk’s foreground partial profile and Su-riyeong’s center-right face immediate and physically grounded. Maintain enough depth for both their contact and a faint trace of the entry area to remain legible, while still allowing the background to stay minimal. Use the soft desaturated daylight as broad ambient illumination, favoring quiet tonal restraint, realistic facial detail, and the unembellished tenderness of the held moment."
    }
   ]
  },
  "reused": false
 },
 "S5sh8": {
  "input_fingerprint": "174949dade111e1e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 외출 복장의 수리영이 강민숙의 어깨에 코를 박고 눈을 반쯤 감은 채 옅게 미소 짓는 구도.\n\nLOCATION (lock): Just inside the rooftop apartment’s entry area, where the daughter finishes putting on her shoes before leaving. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside 강민숙, finish the track in a close conversational frame facing 수리영 in three-quarter profile as she presses her nose into her mother's shoulder. Keep 수리영's softly smiling face in the center-right and 강민숙's shoulder and partial profile in the left foreground, preserving their contact rather than opening the composition.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, midground; 강민숙 in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 현관 근처 실내 공간 (수리영 is prepared to leave from this area); used as Soft, minimally visible spatial context behind the close mother-daughter contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight entering through the small living-room window is kept soft and desaturated, briefly easing the oppressive interior mood without introducing another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minsuk retains the small circular wrist mark that Su-riyeong has seen since childhood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락); 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 외출 복장의 수리영이 강민숙의 어깨에 코를 박고 눈을 반쯤 감은 채 옅게 미소 짓는 구도.\n\nLOCATION (lock): Just inside the rooftop apartment’s entry area, where the daughter finishes putting on her shoes before leaving. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside 강민숙, finish the track in a close conversational frame facing 수리영 in three-quarter profile as she presses her nose into her mother's shoulder. Keep 수리영's softly smiling face in the center-right and 강민숙's shoulder and partial profile in the left foreground, preserving their contact rather than opening the composition.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, midground; 강민숙 in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 현관 근처 실내 공간 (수리영 is prepared to leave from this area); used as Soft, minimally visible spatial context behind the close mother-daughter contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight entering through the small living-room window is kept soft and desaturated, briefly easing the oppressive interior mood without introducing another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minsuk retains the small circular wrist mark that Su-riyeong has seen since childhood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락); 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse an intimate portrait-lens feel with very shallow focus, holding Su-riyeong’s half-closed eyes, faint smile, and the point where her nose presses into Minsuk’s shoulder in delicate clarity. Let Minsuk’s foreground profile and shoulder fall gently softer, creating layered closeness while the entry interior dissolves into restrained, desaturated context. Shape the existing window daylight softly across Su-riyeong’s face, preserving natural skin and fabric texture without polish or sentimentality.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 외출 복장의 수리영이 강민숙의 어깨에 코를 박고 눈을 반쯤 감은 채 옅게 미소 짓는 구도.\n\nLOCATION (lock): Just inside the rooftop apartment’s entry area, where the daughter finishes putting on her shoes before leaving. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside 강민숙, finish the track in a close conversational frame facing 수리영 in three-quarter profile as she presses her nose into her mother's shoulder. Keep 수리영's softly smiling face in the center-right and 강민숙's shoulder and partial profile in the left foreground, preserving their contact rather than opening the composition.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, midground; 강민숙 in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 현관 근처 실내 공간 (수리영 is prepared to leave from this area); used as Soft, minimally visible spatial context behind the close mother-daughter contact.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight entering through the small living-room window is kept soft and desaturated, briefly easing the oppressive interior mood without introducing another light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minsuk retains the small circular wrist mark that Su-riyeong has seen since childhood.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락); 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph the close-up with a natural, slightly wider observational lens feel, keeping the spatial relationship between Minsuk’s foreground partial profile and Su-riyeong’s center-right face immediate and physically grounded. Maintain enough depth for both their contact and a faint trace of the entry area to remain legible, while still allowing the background to stay minimal. Use the soft desaturated daylight as broad ambient illumination, favoring quiet tonal restraint, realistic facial detail, and the unembellished tenderness of the held moment.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "레퍼런스 이미지의 주방 배경과 식탁 위에 놓인 팔의 포즈를 완벽하게 유지하여 장소 및 연속성 지침을 잘 따랐으나, 수리영의 비니가 누락되었습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "수리영의 복장은 잘 구현했으나, 레퍼런스의 배경을 무시하고 새로운 공간을 창조했으며 손목 안쪽의 표식이 손등으로 이동하는 심각한 연속성 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S5sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 강민숙: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:761261>"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영의 눈이 반쯤 감겨 있어야 하는데 크게 떠 있습니다.",
     "fix_en": "Lower Su-riyeong's eyelids so her eyes are half-closed."
    },
    {
     "issue_ko": "수리영이 캐릭터 레퍼런스의 비니를 착용하지 않아 머리카락이 노출되어 있습니다.",
     "fix_en": "Add the dark grey beanie to Su-riyeong's head, covering her upper hair and forehead."
    },
    {
     "issue_ko": "강민숙의 손목에 있는 마크가 이전 컷 레퍼런스에 나타난 기하학적 선과 점이 있는 타투가 아닌 단순한 원으로 그려져 있습니다.",
     "fix_en": "Change the simple circle mark on the wrist to match the precise geometric dotted-line tattoo shown in the previous shot still."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Lower Su-riyeong's eyelids so her eyes are half-closed.\n- Add the dark grey beanie to Su-riyeong's head, covering her upper hair and forehead.\n- Change the simple circle mark on the wrist to match the precise geometric dotted-line tattoo shown in the previous shot still.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도와 인물의 피부 및 의상 질감이 자연스럽게 연출되었으나, 수리영의 비니가 누락되고 손목 마크의 내부 디테일이 생략되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "수리영의 의상은 레퍼런스와 더 유사하나, 손목 마크가 지정된 원형이 아닌 형태이며 디지털 스티커를 합성한 것처럼 보여 재질감 규칙을 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도와 인물의 피부 및 의상 질감이 자연스럽게 연출되었으나, 수리영의 비니가 누락되고 손목 마크의 내부 디테일이 생략되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "수리영의 의상은 레퍼런스와 더 유사하나, 손목 마크가 지정된 원형이 아닌 형태이며 디지털 스티커를 합성한 것처럼 보여 재질감 규칙을 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "수리영의 비니 등 복장은 잘 구현되었으나, 전경의 흐릿한 팔 위에 선명한 그래픽 스티커처럼 합성된 문신이 물리적 피사계 심도와 질감 규칙을 위반했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "수리영의 비니가 누락되는 복장 오류가 있으나, 카메라의 피사계 심도와 피부 질감에 맞게 손목 문신이 자연스럽게 블러 처리되어 물리적 사실성 기준을 충족했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "수리영의 비니 등 복장은 잘 구현되었으나, 전경의 흐릿한 팔 위에 선명한 그래픽 스티커처럼 합성된 문신이 물리적 피사계 심도와 질감 규칙을 위반했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "수리영의 비니가 누락되는 복장 오류가 있으나, 카메라의 피사계 심도와 피부 질감에 맞게 손목 문신이 자연스럽게 블러 처리되어 물리적 사실성 기준을 충족했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 8
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S5sh6"
  }
 },
 "groupbg::해안도로": {
  "input_fingerprint": "8638f571c2b075a5",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "해안도로",
    "tags": [
     "S6sh1",
     "S9sh1"
    ]
   },
   "context_sig": "c31824e4d1186395"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n서해안 해안도로: 바다와 인접해 포장된 한산한 해안가 아스팔트 도로. 한쪽으로는 방파제와 바다가 보이고 다른 쪽으로는 구조물들이 이어진다. (특징: 포장된 아스팔트 도로; 도로 옆 콘크리트 방파제 구조물; 시야가 탁 트인 잿빛 바다 풍경)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 수리영 자전거를 타고 해안가 도로를 달린다.\n- 해안도로를 자전거를 타고 바삐 집으로 달려가는 수리영.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n서해안 해안도로: 바다와 인접해 포장된 한산한 해안가 아스팔트 도로. 한쪽으로는 방파제와 바다가 보이고 다른 쪽으로는 구조물들이 이어진다. (특징: 포장된 아스팔트 도로; 도로 옆 콘크리트 방파제 구조물; 시야가 탁 트인 잿빛 바다 풍경)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 수리영 자전거를 타고 해안가 도로를 달린다.\n- 해안도로를 자전거를 타고 바삐 집으로 달려가는 수리영.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_해안도로_7192d9.png",
  "asset_id": "fb6abebd-e002-4cb4-a2b4-f506b9009590",
  "input_asset_ids": [
   "62b6748e-d69b-40e5-b520-d6efedfebcfb"
  ],
  "origin_tag": "S6sh1",
  "place_text": "On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore.",
  "origin_inputs": {
   "place_text": "On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore.",
   "time_of_day_en": "day",
   "conti_asset_id": "62b6748e-d69b-40e5-b520-d6efedfebcfb"
  }
 },
 "S6sh1::bgfirst_bg": {
  "input_fingerprint": "bed3e836d7dd080b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서늘한 톤의 해안가 도로 위, 이어폰을 낀 수리영이 자전거 안장에 앉아 고개를 위로 들고 눈을 감은 정면 상반신, 달리는 추진력에 의해 머리카락이 진행 반대 방향으로 흩날리는 모습.\n\nLOCATION (lock): On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Centered at upper-torso height ahead of the bicycle, travel backward at 수리영's pace while the dolly-in narrows to her frontal upper body. Keep her face in the upper center, eyes closed and chin raised, with the earphones visible and her hair streaming backward relative to her forward movement while the road remains a stable scale reference beneath her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 riding toward camera in the middle-center of the frame, midground, moves toward forward camera axis; coastal road behind 수리영 in the lower-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 해안가 도로 (수리영 is riding along it during daytime); used as Moving background and lower-frame scale reference that communicates bicycle speed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Cool, desaturated daytime tonality preserves clear and lively visibility while retaining restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서늘한 톤의 해안가 도로 위, 이어폰을 낀 수리영이 자전거 안장에 앉아 고개를 위로 들고 눈을 감은 정면 상반신, 달리는 추진력에 의해 머리카락이 진행 반대 방향으로 흩날리는 모습.\n\nLOCATION (lock): On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Centered at upper-torso height ahead of the bicycle, travel backward at 수리영's pace while the dolly-in narrows to her frontal upper body. Keep her face in the upper center, eyes closed and chin raised, with the earphones visible and her hair streaming backward relative to her forward movement while the road remains a stable scale reference beneath her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 riding toward camera in the middle-center of the frame, midground, moves toward forward camera axis; coastal road behind 수리영 in the lower-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 해안가 도로 (수리영 is riding along it during daytime); used as Moving background and lower-frame scale reference that communicates bicycle speed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Cool, desaturated daytime tonality preserves clear and lively visibility while retaining restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S6sh1__bgfirst_bg.png",
  "asset_id": "8b97b61b-b61f-4e01-8109-d57d3f39754d",
  "input_asset_ids": [
   "62b6748e-d69b-40e5-b520-d6efedfebcfb",
   "fb6abebd-e002-4cb4-a2b4-f506b9009590"
  ]
 },
 "S6sh1": {
  "input_fingerprint": "8368a6ef6d5587f6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서늘한 톤의 해안가 도로 위, 이어폰을 낀 수리영이 자전거 안장에 앉아 고개를 위로 들고 눈을 감은 정면 상반신, 달리는 추진력에 의해 머리카락이 진행 반대 방향으로 흩날리는 모습.\n\nLOCATION (lock): On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Centered at upper-torso height ahead of the bicycle, travel backward at 수리영's pace while the dolly-in narrows to her frontal upper body. Keep her face in the upper center, eyes closed and chin raised, with the earphones visible and her hair streaming backward relative to her forward movement while the road remains a stable scale reference beneath her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 riding toward camera in the middle-center of the frame, midground, moves toward forward camera axis; coastal road behind 수리영 in the lower-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 해안가 도로 (수리영 is riding along it during daytime); used as Moving background and lower-frame scale reference that communicates bicycle speed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Cool, desaturated daytime tonality preserves clear and lively visibility while retaining restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is traveling with her bicycle and wears her earphones on the coastal road.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서늘한 톤의 해안가 도로 위, 이어폰을 낀 수리영이 자전거 안장에 앉아 고개를 위로 들고 눈을 감은 정면 상반신, 달리는 추진력에 의해 머리카락이 진행 반대 방향으로 흩날리는 모습.\n\nLOCATION (lock): On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Centered at upper-torso height ahead of the bicycle, travel backward at 수리영's pace while the dolly-in narrows to her frontal upper body. Keep her face in the upper center, eyes closed and chin raised, with the earphones visible and her hair streaming backward relative to her forward movement while the road remains a stable scale reference beneath her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 riding toward camera in the middle-center of the frame, midground, moves toward forward camera axis; coastal road behind 수리영 in the lower-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 해안가 도로 (수리영 is riding along it during daytime); used as Moving background and lower-frame scale reference that communicates bicycle speed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Cool, desaturated daytime tonality preserves clear and lively visibility while retaining restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is traveling with her bicycle and wears her earphones on the coastal road.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서늘한 톤의 해안가 도로 위, 이어폰을 낀 수리영이 자전거 안장에 앉아 고개를 위로 들고 눈을 감은 정면 상반신, 달리는 추진력에 의해 머리카락이 진행 반대 방향으로 흩날리는 모습.\n\nLOCATION (lock): On the open coastal roadway beside the sea, in daylight as the bicycle moves along the shore. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Centered at upper-torso height ahead of the bicycle, travel backward at 수리영's pace while the dolly-in narrows to her frontal upper body. Keep her face in the upper center, eyes closed and chin raised, with the earphones visible and her hair streaming backward relative to her forward movement while the road remains a stable scale reference beneath her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 riding toward camera in the middle-center of the frame, midground, moves toward forward camera axis; coastal road behind 수리영 in the lower-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 해안가 도로 (수리영 is riding along it during daytime); used as Moving background and lower-frame scale reference that communicates bicycle speed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Cool, desaturated daytime tonality preserves clear and lively visibility while retaining restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is traveling with her bicycle and wears her earphones on the coastal road.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S6sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S6sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_해안도로_7192d9.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 샷의 구도와 인물의 포즈는 잘 구현했으나, 레퍼런스의 의상을 무시하고 임의의 초록색 스웨터를 입혀 감점되었습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "요구된 서늘한 톤의 조명과 인물의 상반신 구도, 흩날리는 머리카락을 완벽히 구현했으며, 레퍼런스의 의상까지 정확하게 일치합니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 샷의 구도와 인물의 포즈는 잘 구현했으나, 레퍼런스의 의상을 무시하고 임의의 초록색 스웨터를 입혀 감점되었습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "요구된 서늘한 톤의 조명과 인물의 상반신 구도, 흩날리는 머리카락을 완벽히 구현했으며, 레퍼런스의 의상까지 정확하게 일치합니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 앵글, 포즈(눈을 감고 고개를 든 모습), 흩날리는 머리카락을 잘 연출했으며 레퍼런스의 의상(검은색 재킷)을 완벽하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "구도와 인물의 표정 연기는 훌륭하나, 캐릭터 레퍼런스에 주어진 명확한 의상(재킷)을 무시하고 임의의 스웨트셔츠를 입힌 명백한 오류가 있습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 앵글, 포즈(눈을 감고 고개를 든 모습), 흩날리는 머리카락을 잘 연출했으며 레퍼런스의 의상(검은색 재킷)을 완벽하게 재현했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "구도와 인물의 표정 연기는 훌륭하나, 캐릭터 레퍼런스에 주어진 명확한 의상(재킷)을 무시하고 임의의 스웨트셔츠를 입힌 명백한 오류가 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 11,
     "B": 17
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 11,
   "B": 17
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "지정된 샷의 구도와 인물의 포즈는 잘 구현했으나, 레퍼런스의 의상을 무시하고 임의의 초록색 스웨터를 입혀 감점되었습니다."
   },
   {
    "label": "B",
    "score": 9,
    "verdict_ko": "요구된 서늘한 톤의 조명과 인물의 상반신 구도, 흩날리는 머리카락을 완벽히 구현했으며, 레퍼런스의 의상까지 정확하게 일치합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_해안도로_7192d9.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": []
  },
  "composition_critique": {
   "issues": []
  },
  "fix_skipped": true,
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S6sh1__bgfirst_bg.png",
   "bg_asset_id": "8b97b61b-b61f-4e01-8109-d57d3f39754d",
   "bg_record_key": "S6sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "해안도로",
   "groupbg_asset_id": "fb6abebd-e002-4cb4-a2b4-f506b9009590"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "groupbg::보호원_견사": {
  "input_fingerprint": "af1f17cb7dfcfab7",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "보호원_견사",
    "tags": [
     "S7sh2",
     "S7sh5"
    ]
   },
   "context_sig": "a0214fced4d6011d"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n유기견 보호원: 여러 마리의 개들이 수용되어 있는 낡은 시설. 녹슨 철망으로 된 사육장과 미용 및 처치를 위한 작업 공간이 있다. (특징: 녹슨 철창으로 구획된 사육장들; 동물용 미용 도구와 약품이 놓인 작업대; 콘크리트 바닥과 거친 질감의 벽면)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#7. 유기견 보호원 / D\n- 수리영, 갇혀있는 개들 보면서 한숨 후~\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n유기견 보호원: 여러 마리의 개들이 수용되어 있는 낡은 시설. 녹슨 철망으로 된 사육장과 미용 및 처치를 위한 작업 공간이 있다. (특징: 녹슨 철창으로 구획된 사육장들; 동물용 미용 도구와 약품이 놓인 작업대; 콘크리트 바닥과 거친 질감의 벽면)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#7. 유기견 보호원 / D\n- 수리영, 갇혀있는 개들 보면서 한숨 후~\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_보호원_견사_275746.png",
  "asset_id": "cca6a8b2-a39b-494f-8743-9cc993efab5f",
  "input_asset_ids": [
   "57fe8a7b-4e59-4035-ac37-2902e010759c"
  ],
  "origin_tag": "S7sh2",
  "place_text": "Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog.",
  "origin_inputs": {
   "place_text": "Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog.",
   "time_of_day_en": "day",
   "conti_asset_id": "57fe8a7b-4e59-4035-ac37-2902e010759c"
  }
 },
 "S7sh2::bgfirst_bg": {
  "input_fingerprint": "a7a36c7ca25c6789",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 철장 바로 앞에서 수리영이 무표정하고 서늘한 눈빛으로 맹견의 눈을 똑바로 응시하는 안면 클로즈업.\n\nLOCATION (lock): Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At eye level directly in front of 수리영, complete the dolly-in and lock into a facial close-up with the cage front immediately beside the camera axis. Her unsmiling face occupies most of the frame and her eyeline passes just beyond the lens toward the excluded 로트와일러, while a narrow strip of cage structure prevents the shot from becoming a detached portrait.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 alone in close-up in the middle-center of the frame, midground, looks toward 로트와일러 beyond the lens axis.\n- KEY BACKGROUND ELEMENTS: 철장 전면 (The 로트와일러 is confined behind it but excluded from this frame) — The cage front runs beside the camera axis, with only a slim portion visible around 수리영's face; used as A narrow edge adjacent to the lens axis that carries 수리영's eyeline toward the off-frame dog.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light at the shelter is rendered with cool, desaturated low-key restraint across 수리영's fixed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 철장 바로 앞에서 수리영이 무표정하고 서늘한 눈빛으로 맹견의 눈을 똑바로 응시하는 안면 클로즈업.\n\nLOCATION (lock): Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At eye level directly in front of 수리영, complete the dolly-in and lock into a facial close-up with the cage front immediately beside the camera axis. Her unsmiling face occupies most of the frame and her eyeline passes just beyond the lens toward the excluded 로트와일러, while a narrow strip of cage structure prevents the shot from becoming a detached portrait.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 alone in close-up in the middle-center of the frame, midground, looks toward 로트와일러 beyond the lens axis.\n- KEY BACKGROUND ELEMENTS: 철장 전면 (The 로트와일러 is confined behind it but excluded from this frame) — The cage front runs beside the camera axis, with only a slim portion visible around 수리영's face; used as A narrow edge adjacent to the lens axis that carries 수리영's eyeline toward the off-frame dog.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light at the shelter is rendered with cool, desaturated low-key restraint across 수리영's fixed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S7sh2__bgfirst_bg.png",
  "asset_id": "b850756b-eece-4c44-85de-d083ed201389",
  "input_asset_ids": [
   "57fe8a7b-4e59-4035-ac37-2902e010759c",
   "cca6a8b2-a39b-494f-8743-9cc993efab5f"
  ]
 },
 "S7sh2": {
  "input_fingerprint": "3d7126fb87226327",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 철장 바로 앞에서 수리영이 무표정하고 서늘한 눈빛으로 맹견의 눈을 똑바로 응시하는 안면 클로즈업.\n\nLOCATION (lock): Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At eye level directly in front of 수리영, complete the dolly-in and lock into a facial close-up with the cage front immediately beside the camera axis. Her unsmiling face occupies most of the frame and her eyeline passes just beyond the lens toward the excluded 로트와일러, while a narrow strip of cage structure prevents the shot from becoming a detached portrait.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 alone in close-up in the middle-center of the frame, midground, looks toward 로트와일러 beyond the lens axis.\n- KEY BACKGROUND ELEMENTS: 철장 전면 (The 로트와일러 is confined behind it but excluded from this frame) — The cage front runs beside the camera axis, with only a slim portion visible around 수리영's face; used as A narrow edge adjacent to the lens axis that carries 수리영's eyeline toward the off-frame dog.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light at the shelter is rendered with cool, desaturated low-key restraint across 수리영's fixed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 철장 바로 앞에서 수리영이 무표정하고 서늘한 눈빛으로 맹견의 눈을 똑바로 응시하는 안면 클로즈업.\n\nLOCATION (lock): Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At eye level directly in front of 수리영, complete the dolly-in and lock into a facial close-up with the cage front immediately beside the camera axis. Her unsmiling face occupies most of the frame and her eyeline passes just beyond the lens toward the excluded 로트와일러, while a narrow strip of cage structure prevents the shot from becoming a detached portrait.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 alone in close-up in the middle-center of the frame, midground, looks toward 로트와일러 beyond the lens axis.\n- KEY BACKGROUND ELEMENTS: 철장 전면 (The 로트와일러 is confined behind it but excluded from this frame) — The cage front runs beside the camera axis, with only a slim portion visible around 수리영's face; used as A narrow edge adjacent to the lens axis that carries 수리영's eyeline toward the off-frame dog.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light at the shelter is rendered with cool, desaturated low-key restraint across 수리영's fixed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 철장 바로 앞에서 수리영이 무표정하고 서늘한 눈빛으로 맹견의 눈을 똑바로 응시하는 안면 클로즈업.\n\nLOCATION (lock): Directly in front of a dog enclosure at an animal shelter, face-to-face with the confined guard dog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At eye level directly in front of 수리영, complete the dolly-in and lock into a facial close-up with the cage front immediately beside the camera axis. Her unsmiling face occupies most of the frame and her eyeline passes just beyond the lens toward the excluded 로트와일러, while a narrow strip of cage structure prevents the shot from becoming a detached portrait.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 alone in close-up in the middle-center of the frame, midground, looks toward 로트와일러 beyond the lens axis.\n- KEY BACKGROUND ELEMENTS: 철장 전면 (The 로트와일러 is confined behind it but excluded from this frame) — The cage front runs beside the camera axis, with only a slim portion visible around 수리영's face; used as A narrow edge adjacent to the lens axis that carries 수리영's eyeline toward the off-frame dog.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light at the shelter is rendered with cool, desaturated low-key restraint across 수리영's fixed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S7sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S7sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_보호원_견사_275746.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 안면 클로즈업 구도, 렌즈 너머를 향한 정면 시선, 화면 가장자리의 철장 배치 및 캐릭터 레퍼런스(비니)를 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "프레이밍이 안면 클로즈업보다 너무 넓어 프레이밍 지시를 위반했으며, 시선 처리와 캐릭터 복장(비니 누락)이 레퍼런스와 일치하지 않습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 안면 클로즈업 구도, 렌즈 너머를 향한 정면 시선, 화면 가장자리의 철장 배치 및 캐릭터 레퍼런스(비니)를 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "프레이밍이 안면 클로즈업보다 너무 넓어 프레이밍 지시를 위반했으며, 시선 처리와 캐릭터 복장(비니 누락)이 레퍼런스와 일치하지 않습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 안면 클로즈업 앵글과 렌즈 너머를 똑바로 응시하는 시선, 프레임 가장자리의 철장 배치 등을 매우 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레임이 너무 넓어 안면 클로즈업 지시를 위반했으며, 시선 역시 카메라 렌즈 너머가 아닌 측면을 향하고 있습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 안면 클로즈업 앵글과 렌즈 너머를 똑바로 응시하는 시선, 프레임 가장자리의 철장 배치 등을 매우 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레임이 너무 넓어 안면 클로즈업 지시를 위반했으며, 시선 역시 카메라 렌즈 너머가 아닌 측면을 향하고 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 7,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "요구된 안면 클로즈업 구도, 렌즈 너머를 향한 정면 시선, 화면 가장자리의 철장 배치 및 캐릭터 레퍼런스(비니)를 정확하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "프레이밍이 안면 클로즈업보다 너무 넓어 프레이밍 지시를 위반했으며, 시선 처리와 캐릭터 복장(비니 누락)이 레퍼런스와 일치하지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_보호원_견사_275746.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영의 시선이 렌즈 너머의 대상을 향하지 않고 카메라 렌즈를 정면으로 똑바로 쳐다보고 있습니다.",
     "fix_en": "Redirect her eyes slightly so she is looking just past the camera lens instead of staring directly into it."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redirect her eyes slightly so she is looking just past the camera lens instead of staring directly into it.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 프레이밍과 렌즈 너머를 응시하는 시선 방향을 정확하게 구현함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "렌즈 축 너머를 향해야 할 시선이 측면으로 심하게 빗나감."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 프레이밍과 렌즈 너머를 응시하는 시선 방향을 정확하게 구현함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "렌즈 축 너머를 향해야 할 시선이 측면으로 심하게 빗나감."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "렌즈 너머를 똑바로 응시하는 시선과 프레임 측면의 철장 배치가 샷의 요구사항을 정확하게 충족합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "캐릭터가 렌즈 너머가 아닌 측면을 바라보고 있어 시선 처리 지침을 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "렌즈 너머를 똑바로 응시하는 시선과 프레임 측면의 철장 배치가 샷의 요구사항을 정확하게 충족합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "캐릭터가 렌즈 너머가 아닌 측면을 바라보고 있어 시선 처리 지침을 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 8
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S7sh2__bgfirst_bg.png",
   "bg_asset_id": "b850756b-eece-4c44-85de-d083ed201389",
   "bg_record_key": "S7sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "보호원_견사",
   "groupbg_asset_id": "cca6a8b2-a39b-494f-8743-9cc993efab5f"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S7sh5::variants": {
  "author_fp": "e571e91b23b5ab40",
  "author": {
   "variants": [
    {
     "approach_ko": "얕은 심도와 압축된 망원 계열의 시선으로 통화 중인 수리영의 표정과 고립감을 밀도 높게 포착한다.",
     "prompt_en": "Use a gently compressed, intimate lens feel with shallow depth of field, holding precise focus on her furrowed brow and lowered eyes while the shelter corner falls softly out of focus behind her. Let muted daytime ambience model the face with delicate side-to-side falloff, preserving subdued skin tones and a quiet, private tension."
    },
    {
     "approach_ko": "보통 화각과 비교적 깊은 심도로 수리영과 쉼터 모서리의 관계를 또렷하게 남겨 답 없는 통화의 답답함을 강조한다.",
     "prompt_en": "Use a natural, moderately broad perspective with comparatively deep focus, keeping the shelter corner legible as a close spatial boundary rather than dissolving it. Preserve open visual pressure on the left side of the composition while the subdued ambient daylight renders the frame in restrained, desaturated layers, emphasizing her isolation within the real shelter space."
    }
   ]
  },
  "reused": false
 },
 "S7sh5": {
  "input_fingerprint": "bebce802aef4353e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한구석에서 수리영이 핸드폰을 귀에 댄 채 미간에 주름을 잡고 바닥을 내려다보는 상반신.\n\nLOCATION (lock): In a quiet corner of the animal shelter, set apart from the occupied dog enclosures. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height in the shelter corner, the tracking move settles into a medium upper-body hold on 수리영 at a slight downward angle, inheriting the line of her lowered gaze. She occupies the right-center of frame with the phone pressed to her ear, while restrained headroom and a subtle downward tilt concentrate attention on her furrowed brow before the camera pans away toward the helper's call.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: shelter corner (Present; no further state specified); used as Provides a compressed background boundary behind 수리영, isolating the private call from the shelter activity; mobile phone (Held to her ear during an unanswered call attempt) — Its side and partial back face the camera while the speaker side is pressed to her ear; used as Held beside her face as the action-defining prop, with its narrow side visible to camera.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted daytime ambient light with low-key, desaturated contrast keeps the moment naturalistic and subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong has her phone in hand while attempting to reach Minsuk, who does not answer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한구석에서 수리영이 핸드폰을 귀에 댄 채 미간에 주름을 잡고 바닥을 내려다보는 상반신.\n\nLOCATION (lock): In a quiet corner of the animal shelter, set apart from the occupied dog enclosures. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height in the shelter corner, the tracking move settles into a medium upper-body hold on 수리영 at a slight downward angle, inheriting the line of her lowered gaze. She occupies the right-center of frame with the phone pressed to her ear, while restrained headroom and a subtle downward tilt concentrate attention on her furrowed brow before the camera pans away toward the helper's call.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: shelter corner (Present; no further state specified); used as Provides a compressed background boundary behind 수리영, isolating the private call from the shelter activity; mobile phone (Held to her ear during an unanswered call attempt) — Its side and partial back face the camera while the speaker side is pressed to her ear; used as Held beside her face as the action-defining prop, with its narrow side visible to camera.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted daytime ambient light with low-key, desaturated contrast keeps the moment naturalistic and subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong has her phone in hand while attempting to reach Minsuk, who does not answer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a gently compressed, intimate lens feel with shallow depth of field, holding precise focus on her furrowed brow and lowered eyes while the shelter corner falls softly out of focus behind her. Let muted daytime ambience model the face with delicate side-to-side falloff, preserving subdued skin tones and a quiet, private tension.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한구석에서 수리영이 핸드폰을 귀에 댄 채 미간에 주름을 잡고 바닥을 내려다보는 상반신.\n\nLOCATION (lock): In a quiet corner of the animal shelter, set apart from the occupied dog enclosures. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height in the shelter corner, the tracking move settles into a medium upper-body hold on 수리영 at a slight downward angle, inheriting the line of her lowered gaze. She occupies the right-center of frame with the phone pressed to her ear, while restrained headroom and a subtle downward tilt concentrate attention on her furrowed brow before the camera pans away toward the helper's call.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: shelter corner (Present; no further state specified); used as Provides a compressed background boundary behind 수리영, isolating the private call from the shelter activity; mobile phone (Held to her ear during an unanswered call attempt) — Its side and partial back face the camera while the speaker side is pressed to her ear; used as Held beside her face as the action-defining prop, with its narrow side visible to camera.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted daytime ambient light with low-key, desaturated contrast keeps the moment naturalistic and subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong has her phone in hand while attempting to reach Minsuk, who does not answer.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a natural, moderately broad perspective with comparatively deep focus, keeping the shelter corner legible as a close spatial boundary rather than dissolving it. Preserve open visual pressure on the left side of the composition while the subdued ambient daylight renders the frame in restrained, desaturated layers, emphasizing her isolation within the real shelter space.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 7,
   "A": 6
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "레퍼런스의 장소(콘크리트 바닥)와 캐릭터의 외형을 잘 유지하며 지시된 구도와 동작을 정확히 구현했습니다."
   },
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "프레임 배치와 표정 연기는 우수하나, 바닥이 흙으로 묘사되어 장소 일관성이 떨어지고 손가락 형태가 다소 어색합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S7sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "휴대폰의 화면과 수화기 스피커 부분이 귀가 아닌 바깥쪽(카메라 방향)을 향하고 있어 거꾸로 들고 있습니다.",
     "fix_en": "Flip the mobile phone so its screen and earpiece speaker are pressed against her ear, exposing the back of the phone to the camera."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Flip the mobile phone so its screen and earpiece speaker are pressed against her ear, exposing the back of the phone to the camera.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지정된 미디엄 샷 프레이밍, 배경, 인물의 표정 및 핸드폰을 귀에 댄 자세를 완벽하고 자연스럽게 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "오른쪽 귀에 핸드폰을 대고 있음에도 왼손에 비정상적으로 큰 두 번째 핸드폰이 생성된 치명적인 오류(Hard Violation)가 있습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지정된 미디엄 샷 프레이밍, 배경, 인물의 표정 및 핸드폰을 귀에 댄 자세를 완벽하고 자연스럽게 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "오른쪽 귀에 핸드폰을 대고 있음에도 왼손에 비정상적으로 큰 두 번째 핸드폰이 생성된 치명적인 오류(Hard Violation)가 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "핸드폰 두 개가 동시에 생성되는 치명적인 객체 중복 오류(Hard Violation)가 발생하여 실격입니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "프레이밍, 미간을 찌푸린 채 바닥을 내려다보는 표정, 핸드폰을 귀에 댄 동작 및 레퍼런스와 일치하는 배경까지 지시사항을 완벽하게 구현했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "핸드폰 두 개가 동시에 생성되는 치명적인 객체 중복 오류(Hard Violation)가 발생하여 실격입니다."
     },
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "프레이밍, 미간을 찌푸린 채 바닥을 내려다보는 표정, 핸드폰을 귀에 댄 동작 및 레퍼런스와 일치하는 배경까지 지시사항을 완벽하게 구현했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S7sh2"
  }
 },
 "S8sh1::variants": {
  "author_fp": "b70783626f8f2533",
  "author": {
   "variants": [
    {
     "approach_ko": "노면 가까이에서 정면으로 밀착 추적해 페달링의 속도와 전진감을 강조한다.",
     "prompt_en": "Photograph from a low, close frontal tracking position, with the bicycle and rider dominating the frame and the road dropping away beneath them. Use an expansive lens feel that heightens forward momentum; keep the face readable while the trailing fabric and peripheral neon reflections carry restrained directional motion blur. Let the red glow skim across real fabric, metal, and skin against a deep, desaturated background."
    },
    {
     "approach_ko": "멀리 떨어진 정면 망원 시점으로 도로의 깊이를 압축해 인물을 고립시키고 긴장감을 만든다.",
     "prompt_en": "Photograph from a distant, eye-level frontal position with a compressed lens feel, holding the rider centrally within the dark commercial-road corridor. Layer the evening traffic, roadside stalls, and neon-lit depth tightly behind her without allowing them to compete with her face and pedaling posture. Keep the bicycle and figure crisply resolved while restrained red neon blooms softly through the low-key dusk atmosphere."
    }
   ]
  },
  "reused": false
 },
 "S8sh1": {
  "input_fingerprint": "c5d25e6e0225a37a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 붉은 네온사인이 번지는 어두운 번화가 도로 한가운데, 수리영이 자전거 안장에 앉아 한쪽 무릎을 굽혀 페달을 밟고 옷자락이 뒤로 휘날리는 정면 구도.\n\nLOCATION (lock): On a neon-lit commercial street near the harbor, approaching the large supermarket through evening traffic and roadside stalls. The shot takes place here — the attached STORYBOARD SKETCH fixes the staging, camera and figure placement of this exact place. No location photograph is attached — build the location itself strictly from the location text above and the shot text, inventing nothing beyond them.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk illumination mixes restrained red neon accents with dark, desaturated street tones and naturalistic low-key contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong continues traveling on the same bicycle and retains her phone after the unanswered call.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 붉은 네온사인이 번지는 어두운 번화가 도로 한가운데, 수리영이 자전거 안장에 앉아 한쪽 무릎을 굽혀 페달을 밟고 옷자락이 뒤로 휘날리는 정면 구도.\n\nLOCATION (lock): On a neon-lit commercial street near the harbor, approaching the large supermarket through evening traffic and roadside stalls. The shot takes place here — the attached STORYBOARD SKETCH fixes the staging, camera and figure placement of this exact place. No location photograph is attached — build the location itself strictly from the location text above and the shot text, inventing nothing beyond them.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk illumination mixes restrained red neon accents with dark, desaturated street tones and naturalistic low-key contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong continues traveling on the same bicycle and retains her phone after the unanswered call.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph from a low, close frontal tracking position, with the bicycle and rider dominating the frame and the road dropping away beneath them. Use an expansive lens feel that heightens forward momentum; keep the face readable while the trailing fabric and peripheral neon reflections carry restrained directional motion blur. Let the red glow skim across real fabric, metal, and skin against a deep, desaturated background.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 붉은 네온사인이 번지는 어두운 번화가 도로 한가운데, 수리영이 자전거 안장에 앉아 한쪽 무릎을 굽혀 페달을 밟고 옷자락이 뒤로 휘날리는 정면 구도.\n\nLOCATION (lock): On a neon-lit commercial street near the harbor, approaching the large supermarket through evening traffic and roadside stalls. The shot takes place here — the attached STORYBOARD SKETCH fixes the staging, camera and figure placement of this exact place. No location photograph is attached — build the location itself strictly from the location text above and the shot text, inventing nothing beyond them.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk illumination mixes restrained red neon accents with dark, desaturated street tones and naturalistic low-key contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong continues traveling on the same bicycle and retains her phone after the unanswered call.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph from a distant, eye-level frontal position with a compressed lens feel, holding the rider centrally within the dark commercial-road corridor. Layer the evening traffic, roadside stalls, and neon-lit depth tightly behind her without allowing them to compete with her face and pedaling posture. Keep the bicycle and figure crisply resolved while restrained red neon blooms softly through the low-key dusk atmosphere.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 8,
   "A": 6
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "스토리보드에 제시된 좌우 배경의 배치(좌측 차량, 우측 노점)를 정확히 준수하였으며, 붉은 네온사인의 조명과 흩날리는 옷자락의 묘사가 우수합니다."
   },
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "인물의 정면 구도와 동작은 잘 표현되었으나, 스토리보드의 핵심 지정 사항인 차량(좌)과 노점(우)의 위치가 반대로 뒤집혀 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "STORYBOARD SKETCH — a bare thin-line staging guide, a REFERENCE ONLY: take from it ONLY the camera framing, figure placement, heading and size/depth order. It carries ZERO visual style — every texture, material, light and all realism come from the text and the photographic references. Never let any line-drawing quality leak into the output.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/lane_sketch_S8sh1.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "우측 간판과 좌측 자동차 번호판 등에 금지된 텍스트가 포함되어 있습니다.",
     "fix_en": "Remove all text, letters, and numbers from the neon signs, awnings, and vehicle license plates."
    },
    {
     "issue_ko": "샷 텍스트에 수리영 외의 인물이 명시되지 않았음에도 배경에 다수의 행인들이 묘사되었습니다.",
     "fix_en": "Remove all background pedestrians and vendors from the scene, leaving Su-riyeong as the only person in the image."
    },
    {
     "issue_ko": "수리영이 휴대폰을 들고 있어야 한다는 Carried State 지시와 달리 두 손이 모두 비어 있습니다.",
     "fix_en": "Add a mobile phone into one of Su-riyeong's hands as she grips the handlebar."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove all text, letters, and numbers from the neon signs, awnings, and vehicle license plates.\n- Remove all background pedestrians and vendors from the scene, leaving Su-riyeong as the only person in the image.\n- Add a mobile phone into one of Su-riyeong's hands as she grips the handlebar.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "자전거를 타며 한 손에 스마트폰을 쥐고 있는 디테일을 성공적으로 구현하여 프롬프트의 지시를 잘 따랐습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "구도와 인물의 묘사는 훌륭하나, 전화기를 들고 있어야 한다는 구체적인 소지품 유지 조건을 완전히 누락했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "자전거를 타며 한 손에 스마트폰을 쥐고 있는 디테일을 성공적으로 구현하여 프롬프트의 지시를 잘 따랐습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "구도와 인물의 묘사는 훌륭하나, 전화기를 들고 있어야 한다는 구체적인 소지품 유지 조건을 완전히 누락했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "텍스트 금지 지시를 훌륭하게 준수했으며, 요구된 휴대폰을 손에 들고 있는 디테일을 정확히 표현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "우측 간판에 명확한 텍스트가 생성되어 금지 규칙을 위반했으며, 지시된 휴대폰 소지 상태도 누락되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "텍스트 금지 지시를 훌륭하게 준수했으며, 요구된 휴대폰을 손에 들고 있는 디테일을 정확히 표현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "우측 간판에 명확한 텍스트가 생성되어 금지 규칙을 위반했으며, 지시된 휴대폰 소지 상태도 누락되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "lane(map_marker): 스케치+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S8sh4::bgfirst_bg": {
  "input_fingerprint": "8e19ceb5e2bf8208",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어둑해진 마트 입구 밖, 수리영이 자전거 옆에 선 채 어두운 핸드폰 화면을 굳은 표정으로 들여다보는 상반신.\n\nLOCATION (lock): Immediately outside the supermarket entrance at dusk, beside the parked bicycle.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the entrance at chest height, the camera holds at a medium-close distance in front of 수리영, offset just enough for a three-quarter view rather than a direct portrait. Her upper body occupies the left-center, the dark phone screen sits below her fixed expression, and part of the bicycle remains visible at the right edge as a grounded scale reference.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: supermarket entrance (No open or closed state specified) — The exterior-facing side of the entrance is seen behind and slightly beside her; used as Frames the exterior location behind 수리영 and marks that she has just exited the store; bicycle (Stationary beside 수리영) — One side of the bicycle faces camera near the right edge; used as Remains partially visible beside her, anchoring her intended departure home; mobile phone screen (Screen appears dark) — The front screen is angled upward toward 수리영 and remains sufficiently visible to camera as a dark display; used as Shares focus with her face in the lower central portion of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Darkening dusk and available exterior illumination are rendered in restrained, desaturated low-key contrast without embellishing an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어둑해진 마트 입구 밖, 수리영이 자전거 옆에 선 채 어두운 핸드폰 화면을 굳은 표정으로 들여다보는 상반신.\n\nLOCATION (lock): Immediately outside the supermarket entrance at dusk, beside the parked bicycle.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the entrance at chest height, the camera holds at a medium-close distance in front of 수리영, offset just enough for a three-quarter view rather than a direct portrait. Her upper body occupies the left-center, the dark phone screen sits below her fixed expression, and part of the bicycle remains visible at the right edge as a grounded scale reference.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: supermarket entrance (No open or closed state specified) — The exterior-facing side of the entrance is seen behind and slightly beside her; used as Frames the exterior location behind 수리영 and marks that she has just exited the store; bicycle (Stationary beside 수리영) — One side of the bicycle faces camera near the right edge; used as Remains partially visible beside her, anchoring her intended departure home; mobile phone screen (Screen appears dark) — The front screen is angled upward toward 수리영 and remains sufficiently visible to camera as a dark display; used as Shares focus with her face in the lower central portion of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Darkening dusk and available exterior illumination are rendered in restrained, desaturated low-key contrast without embellishing an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S8sh4__bgfirst_bg.png",
  "asset_id": "a4b61ffe-09f6-49c1-aed5-cb784f85444a",
  "input_asset_ids": [
   "0b6bd164-1b82-44a4-9b5d-6e77d6c0fb40",
   "c2e803ac-08f2-4adb-99eb-cb3af6964a7e"
  ]
 },
 "S8sh4": {
  "input_fingerprint": "a40bfa0007fb8beb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어둑해진 마트 입구 밖, 수리영이 자전거 옆에 선 채 어두운 핸드폰 화면을 굳은 표정으로 들여다보는 상반신.\n\nLOCATION (lock): Immediately outside the supermarket entrance at dusk, beside the parked bicycle. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the entrance at chest height, the camera holds at a medium-close distance in front of 수리영, offset just enough for a three-quarter view rather than a direct portrait. Her upper body occupies the left-center, the dark phone screen sits below her fixed expression, and part of the bicycle remains visible at the right edge as a grounded scale reference.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: supermarket entrance (No open or closed state specified) — The exterior-facing side of the entrance is seen behind and slightly beside her; used as Frames the exterior location behind 수리영 and marks that she has just exited the store; bicycle (Stationary beside 수리영) — One side of the bicycle faces camera near the right edge; used as Remains partially visible beside her, anchoring her intended departure home; mobile phone screen (Screen appears dark) — The front screen is angled upward toward 수리영 and remains sufficiently visible to camera as a dark display; used as Shares focus with her face in the lower central portion of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Darkening dusk and available exterior illumination are rendered in restrained, desaturated low-key contrast without embellishing an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her bicycle remains parked at the mart while she again carries and checks the phone used to call Minsuk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어둑해진 마트 입구 밖, 수리영이 자전거 옆에 선 채 어두운 핸드폰 화면을 굳은 표정으로 들여다보는 상반신.\n\nLOCATION (lock): Immediately outside the supermarket entrance at dusk, beside the parked bicycle. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the entrance at chest height, the camera holds at a medium-close distance in front of 수리영, offset just enough for a three-quarter view rather than a direct portrait. Her upper body occupies the left-center, the dark phone screen sits below her fixed expression, and part of the bicycle remains visible at the right edge as a grounded scale reference.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: supermarket entrance (No open or closed state specified) — The exterior-facing side of the entrance is seen behind and slightly beside her; used as Frames the exterior location behind 수리영 and marks that she has just exited the store; bicycle (Stationary beside 수리영) — One side of the bicycle faces camera near the right edge; used as Remains partially visible beside her, anchoring her intended departure home; mobile phone screen (Screen appears dark) — The front screen is angled upward toward 수리영 and remains sufficiently visible to camera as a dark display; used as Shares focus with her face in the lower central portion of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Darkening dusk and available exterior illumination are rendered in restrained, desaturated low-key contrast without embellishing an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her bicycle remains parked at the mart while she again carries and checks the phone used to call Minsuk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어둑해진 마트 입구 밖, 수리영이 자전거 옆에 선 채 어두운 핸드폰 화면을 굳은 표정으로 들여다보는 상반신.\n\nLOCATION (lock): Immediately outside the supermarket entrance at dusk, beside the parked bicycle. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the entrance at chest height, the camera holds at a medium-close distance in front of 수리영, offset just enough for a three-quarter view rather than a direct portrait. Her upper body occupies the left-center, the dark phone screen sits below her fixed expression, and part of the bicycle remains visible at the right edge as a grounded scale reference.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: supermarket entrance (No open or closed state specified) — The exterior-facing side of the entrance is seen behind and slightly beside her; used as Frames the exterior location behind 수리영 and marks that she has just exited the store; bicycle (Stationary beside 수리영) — One side of the bicycle faces camera near the right edge; used as Remains partially visible beside her, anchoring her intended departure home; mobile phone screen (Screen appears dark) — The front screen is angled upward toward 수리영 and remains sufficiently visible to camera as a dark display; used as Shares focus with her face in the lower central portion of the composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Darkening dusk and available exterior illumination are rendered in restrained, desaturated low-key contrast without embellishing an additional source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her bicycle remains parked at the mart while she again carries and checks the phone used to call Minsuk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S8sh4__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S8sh4.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L11B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "우측 배경의 자전거가 여러 대 겹쳐진 형태로 왜곡되어 생성된 치명적인 형태 오류(Hard Violation)가 있습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "제시된 구도와 인물의 위치, 어두운 핸드폰 화면을 응시하는 모습 및 우측 자전거의 배치를 정확히 구현했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "우측 배경의 자전거가 여러 대 겹쳐진 형태로 왜곡되어 생성된 치명적인 형태 오류(Hard Violation)가 있습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "제시된 구도와 인물의 위치, 어두운 핸드폰 화면을 응시하는 모습 및 우측 자전거의 배치를 정확히 구현했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "마트 입구의 디테일과 자전거 등 장소 묘사가 레퍼런스와 매우 흡사하나, 명시된 지시와 달리 핸드폰 화면이 카메라를 향하지 않고 뒷면이 노출되었습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "레퍼런스에 없는 실외 냉동고와 두 대의 자전거가 임의로 추가되어 장소 일관성을 어겼으며, A와 마찬가지로 핸드폰 뒷면이 보입니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "마트 입구의 디테일과 자전거 등 장소 묘사가 레퍼런스와 매우 흡사하나, 명시된 지시와 달리 핸드폰 화면이 카메라를 향하지 않고 뒷면이 노출되었습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "레퍼런스에 없는 실외 냉동고와 두 대의 자전거가 임의로 추가되어 장소 일관성을 어겼으며, A와 마찬가지로 핸드폰 뒷면이 보입니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 8,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "우측 배경의 자전거가 여러 대 겹쳐진 형태로 왜곡되어 생성된 치명적인 형태 오류(Hard Violation)가 있습니다."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "제시된 구도와 인물의 위치, 어두운 핸드폰 화면을 응시하는 모습 및 우측 자전거의 배치를 정확히 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L11B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라에 어두운 핸드폰 화면이 보여야 한다는 지시와 달리, 핸드폰을 세워 들고 있어 뒷면과 측면만 보입니다.",
     "fix_en": "Tilt the phone to a flatter, more horizontal angle in her hand so the dark screen surface becomes visible to the camera."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Tilt the phone to a flatter, more horizontal angle in her hand so the dark screen surface becomes visible to the camera.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "인물과 배경의 구도가 정확하며, 핸드폰을 쥐고 바라보는 방향이 지시사항과 물리적 현실성에 부합합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물이 핸드폰의 뒷면을 바라보고 화면이 카메라를 향해 있는 심각한 소품 방향 오류가 발생했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "인물과 배경의 구도가 정확하며, 핸드폰을 쥐고 바라보는 방향이 지시사항과 물리적 현실성에 부합합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물이 핸드폰의 뒷면을 바라보고 화면이 카메라를 향해 있는 심각한 소품 방향 오류가 발생했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 상반신 구도와 조명을 훌륭히 구현했으며, 핸드폰 후면(카메라 렌즈)이 카메라를 향해 있어 인물이 화면을 응시하는 소품의 방향성이 사실적입니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "구도와 배경은 정확하나, 핸드폰의 화면(전면)이 카메라를 향하고 있어 인물이 기기의 뒷면을 들여다보는 소품 방향 오류가 발생했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 상반신 구도와 조명을 훌륭히 구현했으며, 핸드폰 후면(카메라 렌즈)이 카메라를 향해 있어 인물이 화면을 응시하는 소품의 방향성이 사실적입니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "구도와 배경은 정확하나, 핸드폰의 화면(전면)이 카메라를 향하고 있어 인물이 기기의 뒷면을 들여다보는 소품 방향 오류가 발생했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 8
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S8sh4__bgfirst_bg.png",
   "bg_asset_id": "a4b61ffe-09f6-49c1-aed5-cb784f85444a",
   "bg_record_key": "S8sh4::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L11B01",
    "B": "L11B02",
    "C": "L11B03",
    "D": "L11B04"
   },
   "assigned": "L11B02",
   "choice": "Candidate B",
   "confident": true,
   "reason_ko": "\"마트 입구 밖\", \"자전거 옆\", \"핸드폰 화면\"이라는 지문에 맞춰 마트 외부와 자전거, 핸드폰이 모두 보이는 현재 배정된 이미지가 적합합니다.",
   "kept": "L11B02"
  }
 },
 "S9sh1::bgfirst_bg": {
  "input_fingerprint": "66235a880bc40a81",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 달빛이 내리는 칠흑 같은 해안도로 위, 한쪽 무릎을 굽혀 페달을 밟아 내리는 중간 자세로 안장에 앉아 상체를 깊숙이 숙이고 옷자락이 뒤로 강하게 펄럭이는 수리영의 자전거 측면 실루엣.\n\nLOCATION (lock): On a dark coastal roadway beside the night sea, heading urgently toward home.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At bicycle-axle height on 수리영's left, the camera tracks parallel in a perpendicular profile, holding her full bicycle silhouette across the middle of frame. Her torso pitches low over the handlebars, one knee drives downward, and the coastal road streams horizontally behind her while her clothing trails opposite the direction of travel.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, moves toward screen right along the coastal road.\n- KEY BACKGROUND ELEMENTS: coastal road (No additional surface state specified); used as Supplies the horizontal movement axis beneath and behind the bicycle; bicycle (Moving rapidly along the road) — Its left side is presented squarely to camera in full profile; used as Defines the complete profile silhouette and active pedaling rhythm.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moonlight shapes a stark side silhouette within a near-black, desaturated night palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 달빛이 내리는 칠흑 같은 해안도로 위, 한쪽 무릎을 굽혀 페달을 밟아 내리는 중간 자세로 안장에 앉아 상체를 깊숙이 숙이고 옷자락이 뒤로 강하게 펄럭이는 수리영의 자전거 측면 실루엣.\n\nLOCATION (lock): On a dark coastal roadway beside the night sea, heading urgently toward home.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At bicycle-axle height on 수리영's left, the camera tracks parallel in a perpendicular profile, holding her full bicycle silhouette across the middle of frame. Her torso pitches low over the handlebars, one knee drives downward, and the coastal road streams horizontally behind her while her clothing trails opposite the direction of travel.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, moves toward screen right along the coastal road.\n- KEY BACKGROUND ELEMENTS: coastal road (No additional surface state specified); used as Supplies the horizontal movement axis beneath and behind the bicycle; bicycle (Moving rapidly along the road) — Its left side is presented squarely to camera in full profile; used as Defines the complete profile silhouette and active pedaling rhythm.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moonlight shapes a stark side silhouette within a near-black, desaturated night palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S9sh1__bgfirst_bg.png",
  "asset_id": "34d19f5a-0e56-4d9b-b2ad-a5e7a0b82361",
  "input_asset_ids": [
   "b38e90af-d125-44ab-bd04-92cfd776e851",
   "fb6abebd-e002-4cb4-a2b4-f506b9009590"
  ]
 },
 "S9sh1": {
  "input_fingerprint": "f62c2d58815265e4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 달빛이 내리는 칠흑 같은 해안도로 위, 한쪽 무릎을 굽혀 페달을 밟아 내리는 중간 자세로 안장에 앉아 상체를 깊숙이 숙이고 옷자락이 뒤로 강하게 펄럭이는 수리영의 자전거 측면 실루엣.\n\nLOCATION (lock): On a dark coastal roadway beside the night sea, heading urgently toward home. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At bicycle-axle height on 수리영's left, the camera tracks parallel in a perpendicular profile, holding her full bicycle silhouette across the middle of frame. Her torso pitches low over the handlebars, one knee drives downward, and the coastal road streams horizontally behind her while her clothing trails opposite the direction of travel.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, moves toward screen right along the coastal road.\n- KEY BACKGROUND ELEMENTS: coastal road (No additional surface state specified); used as Supplies the horizontal movement axis beneath and behind the bicycle; bicycle (Moving rapidly along the road) — Its left side is presented squarely to camera in full profile; used as Defines the complete profile silhouette and active pedaling rhythm.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moonlight shapes a stark side silhouette within a near-black, desaturated night palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong continues home on the same bicycle, with her phone still in her possession.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 달빛이 내리는 칠흑 같은 해안도로 위, 한쪽 무릎을 굽혀 페달을 밟아 내리는 중간 자세로 안장에 앉아 상체를 깊숙이 숙이고 옷자락이 뒤로 강하게 펄럭이는 수리영의 자전거 측면 실루엣.\n\nLOCATION (lock): On a dark coastal roadway beside the night sea, heading urgently toward home. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At bicycle-axle height on 수리영's left, the camera tracks parallel in a perpendicular profile, holding her full bicycle silhouette across the middle of frame. Her torso pitches low over the handlebars, one knee drives downward, and the coastal road streams horizontally behind her while her clothing trails opposite the direction of travel.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, moves toward screen right along the coastal road.\n- KEY BACKGROUND ELEMENTS: coastal road (No additional surface state specified); used as Supplies the horizontal movement axis beneath and behind the bicycle; bicycle (Moving rapidly along the road) — Its left side is presented squarely to camera in full profile; used as Defines the complete profile silhouette and active pedaling rhythm.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moonlight shapes a stark side silhouette within a near-black, desaturated night palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong continues home on the same bicycle, with her phone still in her possession.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 달빛이 내리는 칠흑 같은 해안도로 위, 한쪽 무릎을 굽혀 페달을 밟아 내리는 중간 자세로 안장에 앉아 상체를 깊숙이 숙이고 옷자락이 뒤로 강하게 펄럭이는 수리영의 자전거 측면 실루엣.\n\nLOCATION (lock): On a dark coastal roadway beside the night sea, heading urgently toward home. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At bicycle-axle height on 수리영's left, the camera tracks parallel in a perpendicular profile, holding her full bicycle silhouette across the middle of frame. Her torso pitches low over the handlebars, one knee drives downward, and the coastal road streams horizontally behind her while her clothing trails opposite the direction of travel.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, moves toward screen right along the coastal road.\n- KEY BACKGROUND ELEMENTS: coastal road (No additional surface state specified); used as Supplies the horizontal movement axis beneath and behind the bicycle; bicycle (Moving rapidly along the road) — Its left side is presented squarely to camera in full profile; used as Defines the complete profile silhouette and active pedaling rhythm.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moonlight shapes a stark side silhouette within a near-black, desaturated night palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong continues home on the same bicycle, with her phone still in her possession.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S9sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S9sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_해안도로_7192d9.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물을 완전한 평면 형태의 검은색 실루엣으로 처리하여 인체 표현 지침을 위반했고 자전거의 바퀴살과 소지품이 누락되었습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "카메라 각도와 샷 크기를 정확히 지켰으며, 입체감이 살아있는 실루엣 연출, 정확한 배경 묘사 및 소지품(휴대폰) 유지까지 훌륭하게 구현했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물을 완전한 평면 형태의 검은색 실루엣으로 처리하여 인체 표현 지침을 위반했고 자전거의 바퀴살과 소지품이 누락되었습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "카메라 각도와 샷 크기를 정확히 지켰으며, 입체감이 살아있는 실루엣 연출, 정확한 배경 묘사 및 소지품(휴대폰) 유지까지 훌륭하게 구현했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "트래킹 샷의 수평적 구도를 올바르게 구현했으며, 단순한 검은 실루엣이 아닌 인물의 디테일과 소지한 핸드폰을 유지하여 지침을 준수함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물을 완전히 평면적인 검은 실루엣으로 뭉개어 인물 표현 지침을 어겼으며, 도로 방향과 수직으로 주행하는 불가능한 구도를 연출함."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "트래킹 샷의 수평적 구도를 올바르게 구현했으며, 단순한 검은 실루엣이 아닌 인물의 디테일과 소지한 핸드폰을 유지하여 지침을 준수함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물을 완전히 평면적인 검은 실루엣으로 뭉개어 인물 표현 지침을 어겼으며, 도로 방향과 수직으로 주행하는 불가능한 구도를 연출함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 6,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "인물을 완전한 평면 형태의 검은색 실루엣으로 처리하여 인체 표현 지침을 위반했고 자전거의 바퀴살과 소지품이 누락되었습니다."
   },
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "카메라 각도와 샷 크기를 정확히 지켰으며, 입체감이 살아있는 실루엣 연출, 정확한 배경 묘사 및 소지품(휴대폰) 유지까지 훌륭하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_해안도로_7192d9.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "피사체가 화면 오른쪽이 아닌 왼쪽을 향해 이동하고 있습니다.",
     "fix_en": "Orient the subject and bicycle to face and move toward screen right."
    },
    {
     "issue_ko": "캐릭터의 상체가 깊숙이 숙여지지 않고 다소 세워져 있습니다.",
     "fix_en": "Pitch the character's torso deeply forward over the handlebars."
    },
    {
     "issue_ko": "캐릭터의 손이 자전거 핸들이 아닌 허공을 잡고 있습니다.",
     "fix_en": "Lower the character's hands so they firmly grip the actual bicycle handlebars."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reorient the bicycle and rider into a square left-side profile moving toward screen right, with her gaze and handlebars aimed right and her trailing clothing streaming toward screen left.",
     "issue_ko": "자전거와 수리영이 화면 왼쪽을 향해 이동하며 오른쪽 측면을 보이고 있어, 화면 오른쪽으로 달리면서 왼쪽 측면을 정면 프로필로 보여야 하는 이동축과 측면 방향이 반대다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Orient the subject and bicycle to face and move toward screen right.\n- Pitch the character's torso deeply forward over the handlebars.\n- Lower the character's hands so they firmly grip the actual bicycle handlebars.\n- Reorient the bicycle and rider into a square left-side profile moving toward screen right, with her gaze and handlebars aimed right and her trailing clothing streaming toward screen left.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "요청된 앵글, 조명, 인물의 자세 및 의상 펄럭임이 정확하게 구현되었으며 자전거 구조가 물리적으로 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "대부분의 지시사항을 잘 따랐으나, 핸들바 부분과 손의 형태가 다소 어색하게 생성되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "요청된 앵글, 조명, 인물의 자세 및 의상 펄럭임이 정확하게 구현되었으며 자전거 구조가 물리적으로 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "대부분의 지시사항을 잘 따랐으나, 핸들바 부분과 손의 형태가 다소 어색하게 생성되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "배경의 수평 모션 블러를 통해 카메라 트래킹과 '도로가 수평으로 흐르는' 묘사를 완벽히 구현했으며, 페달을 밟아 내리는 중간 자세도 정확합니다."
     },
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "배경이 정지되어 있어 피사체와 평행하게 이동하며 '도로가 수평으로 흐르는' 트래킹 샷의 요구사항을 충족하지 못했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "배경의 수평 모션 블러를 통해 카메라 트래킹과 '도로가 수평으로 흐르는' 묘사를 완벽히 구현했으며, 페달을 밟아 내리는 중간 자세도 정확합니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "배경이 정지되어 있어 피사체와 평행하게 이동하며 '도로가 수평으로 흐르는' 트래킹 샷의 요구사항을 충족하지 못했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 15
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S9sh1__bgfirst_bg.png",
   "bg_asset_id": "34d19f5a-0e56-4d9b-b2ad-a5e7a0b82361",
   "bg_record_key": "S9sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "해안도로",
   "groupbg_asset_id": "fb6abebd-e002-4cb4-a2b4-f506b9009590"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S10sh5::bgfirst_bg": {
  "input_fingerprint": "167268660f4956a4",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 미세하게 벌어진 철문 틈새로, 얼굴이 보이지 않는 한국인 신원미상 남성의 검은 어깨선이 안으로 반쯤 들어간 정지 순간.\n\nLOCATION (lock): At the exterior threshold of the rooftop room, where its rusted metal entrance door has opened only a narrow crack.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Immediately outside the doorway at shoulder height, the camera peers through the narrow opening from an oblique side angle and holds the exact instant when only half of 한국인 신원미상 남성's shoulder line has crossed the threshold. The door edge divides the close frame vertically, concealing his face and leaving his shoulder and upper back compressed within the opening.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 한국인 신원미상 남성 in the middle-center of the frame, midground, moves toward interior of the rooftop room; partially opened iron door in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: iron entrance door (Slightly open) — Its edge and part of the exterior-facing plane angle across the foreground while the opening reveals the interior threshold; used as Creates a narrow through-object frame that hides the intruder's face and bisects his shoulder; doorway opening (Narrow due to the slightly opened door); used as Defines the compressed passage through which the upper body is entering.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key night contrast reduces the intruder to a dark shoulder contour while preserving the hard geometry of the threshold.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 미세하게 벌어진 철문 틈새로, 얼굴이 보이지 않는 한국인 신원미상 남성의 검은 어깨선이 안으로 반쯤 들어간 정지 순간.\n\nLOCATION (lock): At the exterior threshold of the rooftop room, where its rusted metal entrance door has opened only a narrow crack.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Immediately outside the doorway at shoulder height, the camera peers through the narrow opening from an oblique side angle and holds the exact instant when only half of 한국인 신원미상 남성's shoulder line has crossed the threshold. The door edge divides the close frame vertically, concealing his face and leaving his shoulder and upper back compressed within the opening.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 한국인 신원미상 남성 in the middle-center of the frame, midground, moves toward interior of the rooftop room; partially opened iron door in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: iron entrance door (Slightly open) — Its edge and part of the exterior-facing plane angle across the foreground while the opening reveals the interior threshold; used as Creates a narrow through-object frame that hides the intruder's face and bisects his shoulder; doorway opening (Narrow due to the slightly opened door); used as Defines the compressed passage through which the upper body is entering.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key night contrast reduces the intruder to a dark shoulder contour while preserving the hard geometry of the threshold.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S10sh5__bgfirst_bg.png",
  "asset_id": "737503d6-3761-4248-ae53-02c2c7939d70",
  "input_asset_ids": [
   "36803ea9-13b3-4610-a77f-ae9423738d49",
   "253e2196-609a-49b0-89ba-f8e4e8b6ada6",
   "d4d740a2-c21c-4a72-bc83-4cc34149b304"
  ]
 },
 "S10sh5": {
  "input_fingerprint": "e52b342584082b8c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미세하게 벌어진 철문 틈새로, 얼굴이 보이지 않는 한국인 신원미상 남성의 검은 어깨선이 안으로 반쯤 들어간 정지 순간.\n\nLOCATION (lock): At the exterior threshold of the rooftop room, where its rusted metal entrance door has opened only a narrow crack. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Immediately outside the doorway at shoulder height, the camera peers through the narrow opening from an oblique side angle and holds the exact instant when only half of 한국인 신원미상 남성's shoulder line has crossed the threshold. The door edge divides the close frame vertically, concealing his face and leaving his shoulder and upper back compressed within the opening.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 한국인 신원미상 남성 in the middle-center of the frame, midground, moves toward interior of the rooftop room; partially opened iron door in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: iron entrance door (Slightly open) — Its edge and part of the exterior-facing plane angle across the foreground while the opening reveals the interior threshold; used as Creates a narrow through-object frame that hides the intruder's face and bisects his shoulder; doorway opening (Narrow due to the slightly opened door); used as Defines the compressed passage through which the upper body is entering.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key night contrast reduces the intruder to a dark shoulder contour while preserving the hard geometry of the threshold.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The intruder has left the rooftop room’s iron entrance door slightly open as he passes inside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 한국인 신원미상 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미세하게 벌어진 철문 틈새로, 얼굴이 보이지 않는 한국인 신원미상 남성의 검은 어깨선이 안으로 반쯤 들어간 정지 순간.\n\nLOCATION (lock): At the exterior threshold of the rooftop room, where its rusted metal entrance door has opened only a narrow crack. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Immediately outside the doorway at shoulder height, the camera peers through the narrow opening from an oblique side angle and holds the exact instant when only half of 한국인 신원미상 남성's shoulder line has crossed the threshold. The door edge divides the close frame vertically, concealing his face and leaving his shoulder and upper back compressed within the opening.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 한국인 신원미상 남성 in the middle-center of the frame, midground, moves toward interior of the rooftop room; partially opened iron door in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: iron entrance door (Slightly open) — Its edge and part of the exterior-facing plane angle across the foreground while the opening reveals the interior threshold; used as Creates a narrow through-object frame that hides the intruder's face and bisects his shoulder; doorway opening (Narrow due to the slightly opened door); used as Defines the compressed passage through which the upper body is entering.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key night contrast reduces the intruder to a dark shoulder contour while preserving the hard geometry of the threshold.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The intruder has left the rooftop room’s iron entrance door slightly open as he passes inside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 한국인 신원미상 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 미세하게 벌어진 철문 틈새로, 얼굴이 보이지 않는 한국인 신원미상 남성의 검은 어깨선이 안으로 반쯤 들어간 정지 순간.\n\nLOCATION (lock): At the exterior threshold of the rooftop room, where its rusted metal entrance door has opened only a narrow crack. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Immediately outside the doorway at shoulder height, the camera peers through the narrow opening from an oblique side angle and holds the exact instant when only half of 한국인 신원미상 남성's shoulder line has crossed the threshold. The door edge divides the close frame vertically, concealing his face and leaving his shoulder and upper back compressed within the opening.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 한국인 신원미상 남성 in the middle-center of the frame, midground, moves toward interior of the rooftop room; partially opened iron door in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: iron entrance door (Slightly open) — Its edge and part of the exterior-facing plane angle across the foreground while the opening reveals the interior threshold; used as Creates a narrow through-object frame that hides the intruder's face and bisects his shoulder; doorway opening (Narrow due to the slightly opened door); used as Defines the compressed passage through which the upper body is entering.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key night contrast reduces the intruder to a dark shoulder contour while preserving the hard geometry of the threshold.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The intruder has left the rooftop room’s iron entrance door slightly open as he passes inside.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 한국인 신원미상 남성 (한국인 성인 남성 얼굴, 자연스러운 피부색, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S10sh5__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S10sh5.png"
    },
    {
     "label": "CHARACTER REFERENCE — 한국인 신원미상 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:824125>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L04B03.png"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/background_chain/seed_bg_rooftop_villa_sel.png"
    },
    {
     "label": "CHARACTER REFERENCE — 한국인 신원미상 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:824125>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "클로즈업 앵글과 좁은 문틈으로 어깨선만 반쯤 보이는 프레이밍 지시를 매우 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 클로즈업 샷 대신 너무 넓은 화각으로 인물의 하반신과 주변 환경이 과도하게 노출되어 프레이밍 규정을 어겼습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "클로즈업 앵글과 좁은 문틈으로 어깨선만 반쯤 보이는 프레이밍 지시를 매우 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 클로즈업 샷 대신 너무 넓은 화각으로 인물의 하반신과 주변 환경이 과도하게 노출되어 프레이밍 규정을 어겼습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 클로즈업 샷과 카메라 위치를 정확히 구현하여 문틈 사이로 보이는 피사체의 어깨선과 뒷모습만을 완벽하게 포착했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "프레임이 너무 넓어 피사체의 전신이 거의 다 노출되었으며, 요구된 클로즈업 스케일과 카메라 위치 기준을 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 클로즈업 샷과 카메라 위치를 정확히 구현하여 문틈 사이로 보이는 피사체의 어깨선과 뒷모습만을 완벽하게 포착했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "프레임이 너무 넓어 피사체의 전신이 거의 다 노출되었으며, 요구된 클로즈업 스케일과 카메라 위치 기준을 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 7
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 15,
   "B": 7
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "클로즈업 앵글과 좁은 문틈으로 어깨선만 반쯤 보이는 프레이밍 지시를 매우 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "요구된 클로즈업 샷 대신 너무 넓은 화각으로 인물의 하반신과 주변 환경이 과도하게 노출되어 프레이밍 규정을 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L04B03.png"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/background_chain/seed_bg_rooftop_villa_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 한국인 신원미상 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:824125>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "문 가장자리에 완전히 가려져야 할 남성의 뒤통수와 귀가 열린 문틈 사이로 보임.",
     "fix_en": "Hide the head and ear completely behind the left door edge so only the shoulder and upper back are visible in the narrow opening."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Close the iron door to the narrow-crack position shown in the shot background, restoring its foreground edge to the original screen position and compressing the visible doorway opening.",
     "issue_ko": "철문이 기준 배경보다 훨씬 넓게 열려 있어 좁은 틈새 구도가 깨지고, 전경의 문 가장자리도 화면에서 지나치게 왼쪽에 놓여 있다."
    },
    {
     "fix_en": "Reposition the man farther behind the door edge so that only half of his shoulder line and a compressed portion of his upper back are visible within the narrow opening, with his head fully concealed.",
     "issue_ko": "남성의 뒤통수와 상반신이 너무 많이 드러나 있어, 얼굴을 가린 채 어깨선 절반만 문턱을 넘은 순간이라는 배치와 맞지 않는다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Hide the head and ear completely behind the left door edge so only the shoulder and upper back are visible in the narrow opening.\n- Close the iron door to the narrow-crack position shown in the shot background, restoring its foreground edge to the original screen position and compressing the visible doorway opening.\n- Reposition the man farther behind the door edge so that only half of his shoulder line and a compressed portion of his upper back are visible within the narrow opening, with his head fully concealed.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 배경과 구도를 정확히 유지하면서, 레퍼런스의 검은 코트 질감과 견장 디테일을 더 사실적으로 구현했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "배경과 구도는 잘 유지되었으나, 코트 어깨의 견장 형태가 레퍼런스에 비해 다소 어색하게 렌더링되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 배경과 구도를 정확히 유지하면서, 레퍼런스의 검은 코트 질감과 견장 디테일을 더 사실적으로 구현했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "배경과 구도는 잘 유지되었으나, 코트 어깨의 견장 형태가 레퍼런스에 비해 다소 어색하게 렌더링되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 샷의 구도와 원본 배경을 완벽하게 유지하면서, 레퍼런스 인물의 의상(트렌치코트의 견장 디테일)과 실루엣을 스케치에 맞게 자연스럽게 합성함."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구사항을 잘 준수하였으나, 인물의 머리카락 형태와 코트 어깨선 디테일이 A에 비해 레퍼런스와의 일치도가 다소 떨어짐."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 샷의 구도와 원본 배경을 완벽하게 유지하면서, 레퍼런스 인물의 의상(트렌치코트의 견장 디테일)과 실루엣을 스케치에 맞게 자연스럽게 합성함."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구사항을 잘 준수하였으나, 인물의 머리카락 형태와 코트 어깨선 디테일이 A에 비해 레퍼런스와의 일치도가 다소 떨어짐."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S10sh5__bgfirst_bg.png",
   "bg_asset_id": "737503d6-3761-4248-ae53-02c2c7939d70",
   "bg_record_key": "S10sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready",
  "plate_select": {
   "candidates": {
    "A": "L04B01",
    "B": "L04B02",
    "C": "L04B03"
   },
   "assigned": "L04B03",
   "choice": "Candidate C",
   "confident": true,
   "reason_ko": "원문의 '미세하게 벌어진 철문 틈새로'라는 묘사에 따라, 옥상 구조물 중 철문을 가장 가까이에서 비추고 있는 후보 C가 적합함.",
   "kept": "L04B03"
  }
 },
 "S10sh6::variants": {
  "author_fp": "9d32cdcefcec65ba",
  "author": {
   "variants": [
    {
     "approach_ko": "창과 실루엣을 냉정한 기하학적 평면으로 포착하는 선명하고 객관적인 접근.",
     "prompt_en": "Treat the window as a severe geometric proscenium, with the exterior frame and opaque pane rendered in crisp, even focus. Keep the image rigidly frontal and visually flattened, using dense blacks against restrained steel-gray highlights so the distorted human contours read as a stark, clinical tableau. Preserve subtle physical texture in the glass and worn frame without softening their enclosing geometry."
    },
    {
     "approach_ko": "불투명 유리의 질감과 산란광을 강조해 공격의 형상을 불안하고 촉각적으로 드러내는 접근.",
     "prompt_en": "Favor the tactile obstruction of the opaque glass: hold its surface texture sharply while the figures beyond dissolve into broad, wavering tonal masses with faint edge bloom from the severe light behind them. Let the window frame fall slightly darker and less prominent toward its edges, drawing attention into the murky collision of forms within the pane. Use muted charcoal tones and restrained halation to create an oppressive, voyeuristic sense of distance without losing the figures’ human proportions."
    }
   ]
  },
  "reused": false
 },
 "S10sh6": {
  "input_fingerprint": "3baf5c762604c18a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 옥탑방의 작은 불투명 유리창 너머로, 거대한 검은 그림자가 쓰러진 강민숙의 어깨 쪽으로 몸을 굽힌 기괴한 실루엣 구도.\n\nLOCATION (lock): Outside the rooftop room’s small window on the villa roof, looking through the opaque glass at silhouettes inside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The camera cuts to a fixed position close and perpendicular to the small window, widening only within its plane so both figures remain enclosed by the window boundaries. 거대한 검은 그림자 rises across the upper-right and bends toward 강민숙's shoulder in the lower-left, their forms visible only as distorted silhouettes through the obscuring glass.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 거대한 검은 그림자 in the upper-right of the frame, background, looks toward 강민숙's shoulder; 강민숙 in the lower-left of the frame, background; small obscuring window in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: small obscuring glass window (Intact; transparency is limited) — The camera faces the pane squarely; the figures beyond are visible only as indistinct silhouettes with interior detail obscured; used as Acts as the entire observational barrier and confines the attack to a flattened silhouette tableau; window frame (No further state specified) — Its front plane faces camera nearly squarely and bounds the figures on all sides; used as Provides the enclosing border around both silhouettes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Severe night contrast behind the obscuring pane creates a grim, desaturated silhouette image without identifying a specific interior source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The iron entrance remains ajar behind the intruder. Minsuk still bears her small wrist mark as he bends over her and tears into her shoulder.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 옥탑방의 작은 불투명 유리창 너머로, 거대한 검은 그림자가 쓰러진 강민숙의 어깨 쪽으로 몸을 굽힌 기괴한 실루엣 구도.\n\nLOCATION (lock): Outside the rooftop room’s small window on the villa roof, looking through the opaque glass at silhouettes inside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The camera cuts to a fixed position close and perpendicular to the small window, widening only within its plane so both figures remain enclosed by the window boundaries. 거대한 검은 그림자 rises across the upper-right and bends toward 강민숙's shoulder in the lower-left, their forms visible only as distorted silhouettes through the obscuring glass.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 거대한 검은 그림자 in the upper-right of the frame, background, looks toward 강민숙's shoulder; 강민숙 in the lower-left of the frame, background; small obscuring window in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: small obscuring glass window (Intact; transparency is limited) — The camera faces the pane squarely; the figures beyond are visible only as indistinct silhouettes with interior detail obscured; used as Acts as the entire observational barrier and confines the attack to a flattened silhouette tableau; window frame (No further state specified) — Its front plane faces camera nearly squarely and bounds the figures on all sides; used as Provides the enclosing border around both silhouettes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Severe night contrast behind the obscuring pane creates a grim, desaturated silhouette image without identifying a specific interior source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The iron entrance remains ajar behind the intruder. Minsuk still bears her small wrist mark as he bends over her and tears into her shoulder.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nTreat the window as a severe geometric proscenium, with the exterior frame and opaque pane rendered in crisp, even focus. Keep the image rigidly frontal and visually flattened, using dense blacks against restrained steel-gray highlights so the distorted human contours read as a stark, clinical tableau. Preserve subtle physical texture in the glass and worn frame without softening their enclosing geometry.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 옥탑방의 작은 불투명 유리창 너머로, 거대한 검은 그림자가 쓰러진 강민숙의 어깨 쪽으로 몸을 굽힌 기괴한 실루엣 구도.\n\nLOCATION (lock): Outside the rooftop room’s small window on the villa roof, looking through the opaque glass at silhouettes inside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The camera cuts to a fixed position close and perpendicular to the small window, widening only within its plane so both figures remain enclosed by the window boundaries. 거대한 검은 그림자 rises across the upper-right and bends toward 강민숙's shoulder in the lower-left, their forms visible only as distorted silhouettes through the obscuring glass.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 거대한 검은 그림자 in the upper-right of the frame, background, looks toward 강민숙's shoulder; 강민숙 in the lower-left of the frame, background; small obscuring window in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: small obscuring glass window (Intact; transparency is limited) — The camera faces the pane squarely; the figures beyond are visible only as indistinct silhouettes with interior detail obscured; used as Acts as the entire observational barrier and confines the attack to a flattened silhouette tableau; window frame (No further state specified) — Its front plane faces camera nearly squarely and bounds the figures on all sides; used as Provides the enclosing border around both silhouettes.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Severe night contrast behind the obscuring pane creates a grim, desaturated silhouette image without identifying a specific interior source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The iron entrance remains ajar behind the intruder. Minsuk still bears her small wrist mark as he bends over her and tears into her shoulder.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nFavor the tactile obstruction of the opaque glass: hold its surface texture sharply while the figures beyond dissolve into broad, wavering tonal masses with faint edge bloom from the severe light behind them. Let the window frame fall slightly darker and less prominent toward its edges, drawing attention into the murky collision of forms within the pane. Use muted charcoal tones and restrained halation to create an oppressive, voyeuristic sense of distance without losing the figures’ human proportions.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 9,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "프롬프트가 요구한 불투명 유리창의 질감과 그 너머로 보이는 기괴한 실루엣의 형태를 완벽하게 구현했으며, 와이드 샷의 구도와 주변 배경도 레퍼런스와 일치합니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "유리창이 불투명하지 않아 인물의 이목구비와 내부 벽면 등 디테일이 너무 선명하게 드러나며, 실루엣 구도라는 핵심 지시사항을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S10sh5_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "레퍼런스 이미지에 명확히 보이는 창문의 수직 쇠창살(방범창)이 이 샷에서는 누락되었습니다.",
     "fix_en": "Add the vertical metal security bars over the exterior of the window frame exactly as they appear in the reference image."
    },
    {
     "issue_ko": "침입자 뒤로 열려 있어야 할 철제 출입문이 보이지 않습니다.",
     "fix_en": "Show the rusty iron entrance door remaining ajar in the background behind the intruder."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Add the vertical metal security bars over the exterior of the window frame exactly as they appear in the reference image.\n- Show the rusty iron entrance door remaining ajar in the background behind the intruder.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "레퍼런스에 제시된 창문의 쇠창살과 내부에 열려 있는 녹슨 철문을 정확히 반영하며, 불투명 유리를 통한 실루엣 구도를 훌륭하게 연출했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "불투명 유리 너머의 실루엣 연출은 양호하나, 레퍼런스의 핵심인 창살이 누락되었고 열린 철문 상태가 반영되지 않았습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "레퍼런스에 제시된 창문의 쇠창살과 내부에 열려 있는 녹슨 철문을 정확히 반영하며, 불투명 유리를 통한 실루엣 구도를 훌륭하게 연출했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "불투명 유리 너머의 실루엣 연출은 양호하나, 레퍼런스의 핵심인 창살이 누락되었고 열린 철문 상태가 반영되지 않았습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "레퍼런스 이미지에 있는 창문의 쇠창살과 내부로 열린 녹슨 문의 공간적 위치를 정확히 유지했으며, 지시된 실루엣 구도도 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "실루엣 구도는 나쁘지 않으나, 레퍼런스에 명확히 존재하는 창문의 쇠창살을 누락하여 장소의 일관성을 크게 훼손했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "레퍼런스 이미지에 있는 창문의 쇠창살과 내부로 열린 녹슨 문의 공간적 위치를 정확히 유지했으며, 지시된 실루엣 구도도 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "실루엣 구도는 나쁘지 않으나, 레퍼런스에 명확히 존재하는 창문의 쇠창살을 누락하여 장소의 일관성을 크게 훼손했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh5"
  },
  "lane_policy": "ab_select_bypass:bg_only:share_plan_prev_bgonly",
  "plate_select": {
   "candidates": {
    "A": "L04B01",
    "B": "L04B02",
    "C": "L04B03"
   },
   "assigned": "L04B03",
   "choice": "Candidate C",
   "confident": true,
   "reason_ko": "지문에서 '불투명 유리창 너머로... 실루엣 구도'라고 명시되어 있으므로, 옥탑방 외부에서 불투명 창문을 직접 조명하고 있는 현재 배정된 후보가 실루엣을 투사하기에 가장 적합한 앵글입니다.",
   "kept": "L04B03"
  }
 },
 "S11sh4::bgfirst_bg": {
  "input_fingerprint": "3fba2ac0cb95dd29",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 옥상 바닥에 두 발을 딛고 선 수리영의 어깨가 경직된 채, 입을 살짝 벌리고 정면을 응시하는 상반신.\n\nLOCATION (lock): On the villa rooftop at the top of the exterior concrete stairs, facing the rooftop room’s entrance.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At upper-torso height on the rooftop, the static camera faces 수리영 along her fixed eyeline and holds her in a centered medium frame. Her planted stance is implied by locked shoulders and an arrested torso, while her slightly open mouth and gaze just past the lens carry the cut toward the entrance rather than inviting direct address.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, looks toward rooftop-room entrance beyond the camera axis.\n- KEY BACKGROUND ELEMENTS: rooftop area (No additional state specified); used as Provides sparse rooftop context while keeping all compositional pressure on 수리영's arrested reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural night ambience is kept dark and desaturated, with restrained contrast shaping her rigid expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 옥상 바닥에 두 발을 딛고 선 수리영의 어깨가 경직된 채, 입을 살짝 벌리고 정면을 응시하는 상반신.\n\nLOCATION (lock): On the villa rooftop at the top of the exterior concrete stairs, facing the rooftop room’s entrance.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At upper-torso height on the rooftop, the static camera faces 수리영 along her fixed eyeline and holds her in a centered medium frame. Her planted stance is implied by locked shoulders and an arrested torso, while her slightly open mouth and gaze just past the lens carry the cut toward the entrance rather than inviting direct address.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, looks toward rooftop-room entrance beyond the camera axis.\n- KEY BACKGROUND ELEMENTS: rooftop area (No additional state specified); used as Provides sparse rooftop context while keeping all compositional pressure on 수리영's arrested reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural night ambience is kept dark and desaturated, with restrained contrast shaping her rigid expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S11sh4__bgfirst_bg.png",
  "asset_id": "5832631d-83c3-4726-9969-907241b12132",
  "input_asset_ids": [
   "aac96b55-2262-44ea-9bbd-676c64f48b05",
   "a46a208b-f121-4a2c-94eb-07c800f5d739"
  ]
 },
 "S11sh4": {
  "input_fingerprint": "a07f9b8f883a4d77",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 옥상 바닥에 두 발을 딛고 선 수리영의 어깨가 경직된 채, 입을 살짝 벌리고 정면을 응시하는 상반신.\n\nLOCATION (lock): On the villa rooftop at the top of the exterior concrete stairs, facing the rooftop room’s entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At upper-torso height on the rooftop, the static camera faces 수리영 along her fixed eyeline and holds her in a centered medium frame. Her planted stance is implied by locked shoulders and an arrested torso, while her slightly open mouth and gaze just past the lens carry the cut toward the entrance rather than inviting direct address.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, looks toward rooftop-room entrance beyond the camera axis.\n- KEY BACKGROUND ELEMENTS: rooftop area (No additional state specified); used as Provides sparse rooftop context while keeping all compositional pressure on 수리영's arrested reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural night ambience is kept dark and desaturated, with restrained contrast shaping her rigid expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s bicycle remains hastily parked in the rear of the yard, while the rooftop entrance door is still half open.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 옥상 바닥에 두 발을 딛고 선 수리영의 어깨가 경직된 채, 입을 살짝 벌리고 정면을 응시하는 상반신.\n\nLOCATION (lock): On the villa rooftop at the top of the exterior concrete stairs, facing the rooftop room’s entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At upper-torso height on the rooftop, the static camera faces 수리영 along her fixed eyeline and holds her in a centered medium frame. Her planted stance is implied by locked shoulders and an arrested torso, while her slightly open mouth and gaze just past the lens carry the cut toward the entrance rather than inviting direct address.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, looks toward rooftop-room entrance beyond the camera axis.\n- KEY BACKGROUND ELEMENTS: rooftop area (No additional state specified); used as Provides sparse rooftop context while keeping all compositional pressure on 수리영's arrested reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural night ambience is kept dark and desaturated, with restrained contrast shaping her rigid expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s bicycle remains hastily parked in the rear of the yard, while the rooftop entrance door is still half open.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 옥상 바닥에 두 발을 딛고 선 수리영의 어깨가 경직된 채, 입을 살짝 벌리고 정면을 응시하는 상반신.\n\nLOCATION (lock): On the villa rooftop at the top of the exterior concrete stairs, facing the rooftop room’s entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At upper-torso height on the rooftop, the static camera faces 수리영 along her fixed eyeline and holds her in a centered medium frame. Her planted stance is implied by locked shoulders and an arrested torso, while her slightly open mouth and gaze just past the lens carry the cut toward the entrance rather than inviting direct address.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, looks toward rooftop-room entrance beyond the camera axis.\n- KEY BACKGROUND ELEMENTS: rooftop area (No additional state specified); used as Provides sparse rooftop context while keeping all compositional pressure on 수리영's arrested reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural night ambience is kept dark and desaturated, with restrained contrast shaping her rigid expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s bicycle remains hastily parked in the rear of the yard, while the rooftop entrance door is still half open.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S11sh4__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S11sh4.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L05B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 '상반신(미디엄 샷)' 프레이밍과 긴 검은 머리카락, 살짝 벌린 입 등의 세부 묘사를 완벽하게 구현하여 지시문을 정확히 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지시된 '상반신(미디엄 샷)'이 아닌 전신이 다 보이는 와이드 샷으로 연출되어 프레이밍 우선순위 규칙을 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 '상반신(미디엄 샷)' 프레이밍과 긴 검은 머리카락, 살짝 벌린 입 등의 세부 묘사를 완벽하게 구현하여 지시문을 정확히 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지시된 '상반신(미디엄 샷)'이 아닌 전신이 다 보이는 와이드 샷으로 연출되어 프레이밍 우선순위 규칙을 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 '상반신(미디엄 샷)' 프레이밍과 시선 처리를 정확히 구현했으며 배경의 자전거 디테일도 잘 살렸습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "요구된 상반신(미디엄 샷)이 아닌 전신을 보여주는 풀 샷으로 렌더링하여 프레이밍 지시를 어겼습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 '상반신(미디엄 샷)' 프레이밍과 시선 처리를 정확히 구현했으며 배경의 자전거 디테일도 잘 살렸습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 상반신(미디엄 샷)이 아닌 전신을 보여주는 풀 샷으로 렌더링하여 프레이밍 지시를 어겼습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 14,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 '상반신(미디엄 샷)' 프레이밍과 긴 검은 머리카락, 살짝 벌린 입 등의 세부 묘사를 완벽하게 구현하여 지시문을 정확히 따랐습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "지시된 '상반신(미디엄 샷)'이 아닌 전신이 다 보이는 와이드 샷으로 연출되어 프레이밍 우선순위 규칙을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L05B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "스케치의 배경 선을 무시해야 한다는 규칙을 어기고, 원본 배경에 없는 흰색 물탱크가 임의로 추가되었습니다.",
     "fix_en": "Remove the white water tank from the left side to strictly preserve the original background."
    },
    {
     "issue_ko": "인물의 고개가 오른쪽으로 돌아가 있어, '정면을 응시하는'이라는 지문 및 머리가 정면을 향한 레이아웃 스케치와 일치하지 않습니다.",
     "fix_en": "Turn the character's head to face perfectly straight forward toward the camera."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Restore the exact camera position, axis, framing, and perspective of the shot-background reference without changing any scene content.",
     "issue_ko": "선택 이미지의 카메라 시점과 원근이 기준 배경에서 달라져 옥상방, 계단, 벽의 화면상 위치와 크기가 변했습니다."
    },
    {
     "fix_en": "Scale Su-riyeong up and shift her slightly right and upward so she matches the sketch's centered medium-shot placement and screen size.",
     "issue_ko": "수리영이 레이아웃 스케치보다 작고 낮으며 화면 중심에서 왼쪽으로 치우쳐 있습니다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the white water tank from the left side to strictly preserve the original background.\n- Turn the character's head to face perfectly straight forward toward the camera.\n- Restore the exact camera position, axis, framing, and perspective of the shot-background reference without changing any scene content.\n- Scale Su-riyeong up and shift her slightly right and upward so she matches the sketch's centered medium-shot placement and screen size.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "중앙 미디엄 프레임과 렌즈 바로 너머를 향한 정면 응시, 경직된 상반신이 가장 정확하며 문도 반쯤 열려 있지만 입이 거의 닫혀 있고 배경 물탱크가 누락됐다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "입을 살짝 벌린 연기는 맞지만 시선이 카메라 축에서 지나치게 벗어나고, 원본에 없던 회색 물탱크를 추가해 고정 배경 보존 조건을 크게 위반했다."
     }
    ],
    "winner": "B"
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "중앙 미디엄 프레임과 렌즈 바로 너머를 향한 정면 응시, 경직된 상반신이 가장 정확하며 문도 반쯤 열려 있지만 입이 거의 닫혀 있고 배경 물탱크가 누락됐다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "입을 살짝 벌린 연기는 맞지만 시선이 카메라 축에서 지나치게 벗어나고, 원본에 없던 회색 물탱크를 추가해 고정 배경 보존 조건을 크게 위반했다."
     }
    ]
   },
   "reverse_raw": {
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "중앙 미디엄 프레임과 정면에 가까운 경직된 상반신은 충실하지만, 입이 거의 닫혀 있고 시선이 렌즈를 향하며 배경의 기존 물탱크가 사라졌다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "살짝 벌어진 입과 렌즈를 비껴 입구로 향하는 시선은 더 정확하지만, 참조에 없는 두 번째 회색 물탱크를 발명해 배경 고정 및 무관한 객체 금지를 위반했다."
     }
    ],
    "winner": "A"
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "중앙 미디엄 프레임과 정면에 가까운 경직된 상반신은 충실하지만, 입이 거의 닫혀 있고 시선이 렌즈를 향하며 배경의 기존 물탱크가 사라졌다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "살짝 벌어진 입과 렌즈를 비껴 입구로 향하는 시선은 더 정확하지만, 참조에 없는 두 번째 회색 물탱크를 발명해 배경 고정 및 무관한 객체 금지를 위반했다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 13
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S11sh4__bgfirst_bg.png",
   "bg_asset_id": "5832631d-83c3-4726-9969-907241b12132",
   "bg_record_key": "S11sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L05B01",
    "B": "L05B02"
   },
   "assigned": "L05B01",
   "choice": "Candidate A",
   "confident": true,
   "reason_ko": "지문에서 '옥상 바닥에 두 발을 딛고'라고 명시되어 있으므로 옥상 공간을 보여주는 Candidate A가 적합합니다.",
   "kept": "L05B01"
  }
 },
 "S11sh5::variants": {
  "author_fp": "bcdac569a827ed42",
  "author": {
   "variants": [
    {
     "approach_ko": "정면에 가까운 절제된 원근과 깊은 초점으로 문과 암흑의 형태를 냉정하게 응시한다.",
     "prompt_en": "Use a restrained, rectilinear lens feel with deep focus and an unwavering eye-level viewpoint. Keep the doorway geometry crisp and nearly frontal while allowing the oblique door slab and its edge to articulate the asymmetrical opening. Let faint night ambience graze the rust and worn metal with muted cool highlights, falling off abruptly into an interior that remains completely unreadable."
    },
    {
     "approach_ko": "압축된 원근과 얕은 초점 이행으로 반쯤 열린 문틈의 암흑에 시선을 강하게 가둔다.",
     "prompt_en": "Use a subtly compressed lens feel, holding the prescribed wide framing while visually crowding the narrow rooftop margins around the entrance. Place critical focus on the half-open door edge and the mouth of the gap, with the nearest margins falling gently soft so the eye is pulled straight into the black interior. Underexpose the surrounding structure into heavy desaturated shadow, preserving only sparse directional sheen across the corroded metal."
    }
   ]
  },
  "reused": false
 },
 "S11sh5": {
  "input_fingerprint": "e5a1c067e355dc71",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수리영의 시야에 맺힌, 녹슨 옥탑방 철문이 반쯤 휑하게 열려 안쪽의 짙은 암흑만을 담아내는 스산한 시점 쇼트.\n\nLOCATION (lock): Directly before the rooftop room’s half-open metal entrance door, looking from the exposed roof into the dark interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From 수리영's eye height and approximate position, the static camera looks directly toward the rooftop-room entrance from several meters away; 수리영 is completely excluded as this is her optical viewpoint. The half-open iron door sits near center with its angled slab on one side and the dark interior occupying the revealed gap, leaving surrounding rooftop space only as a narrow margin.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: half-open iron door in the middle-center of the frame, midground; dark room interior in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rusty iron entrance door (Half open and rusty) — The exterior plane is seen obliquely, while the door edge points into the visible entrance gap; used as Primary focal anchor of the subjective reveal, with its opening forming an asymmetrical dark void; room interior beyond the entrance (Dark; no interior detail is visible); used as Appears through the door gap as the endpoint of 수리영's stare.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Night ambience leaves the entrance in low-key, desaturated contrast and the interior in deep darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rusty iron entrance door remains half open, exposing the dark interior.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수리영의 시야에 맺힌, 녹슨 옥탑방 철문이 반쯤 휑하게 열려 안쪽의 짙은 암흑만을 담아내는 스산한 시점 쇼트.\n\nLOCATION (lock): Directly before the rooftop room’s half-open metal entrance door, looking from the exposed roof into the dark interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From 수리영's eye height and approximate position, the static camera looks directly toward the rooftop-room entrance from several meters away; 수리영 is completely excluded as this is her optical viewpoint. The half-open iron door sits near center with its angled slab on one side and the dark interior occupying the revealed gap, leaving surrounding rooftop space only as a narrow margin.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: half-open iron door in the middle-center of the frame, midground; dark room interior in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rusty iron entrance door (Half open and rusty) — The exterior plane is seen obliquely, while the door edge points into the visible entrance gap; used as Primary focal anchor of the subjective reveal, with its opening forming an asymmetrical dark void; room interior beyond the entrance (Dark; no interior detail is visible); used as Appears through the door gap as the endpoint of 수리영's stare.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Night ambience leaves the entrance in low-key, desaturated contrast and the interior in deep darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rusty iron entrance door remains half open, exposing the dark interior.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained, rectilinear lens feel with deep focus and an unwavering eye-level viewpoint. Keep the doorway geometry crisp and nearly frontal while allowing the oblique door slab and its edge to articulate the asymmetrical opening. Let faint night ambience graze the rust and worn metal with muted cool highlights, falling off abruptly into an interior that remains completely unreadable.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수리영의 시야에 맺힌, 녹슨 옥탑방 철문이 반쯤 휑하게 열려 안쪽의 짙은 암흑만을 담아내는 스산한 시점 쇼트.\n\nLOCATION (lock): Directly before the rooftop room’s half-open metal entrance door, looking from the exposed roof into the dark interior. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From 수리영's eye height and approximate position, the static camera looks directly toward the rooftop-room entrance from several meters away; 수리영 is completely excluded as this is her optical viewpoint. The half-open iron door sits near center with its angled slab on one side and the dark interior occupying the revealed gap, leaving surrounding rooftop space only as a narrow margin.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: half-open iron door in the middle-center of the frame, midground; dark room interior in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rusty iron entrance door (Half open and rusty) — The exterior plane is seen obliquely, while the door edge points into the visible entrance gap; used as Primary focal anchor of the subjective reveal, with its opening forming an asymmetrical dark void; room interior beyond the entrance (Dark; no interior detail is visible); used as Appears through the door gap as the endpoint of 수리영's stare.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Night ambience leaves the entrance in low-key, desaturated contrast and the interior in deep darkness.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rusty iron entrance door remains half open, exposing the dark interior.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a subtly compressed lens feel, holding the prescribed wide framing while visually crowding the narrow rooftop margins around the entrance. Place critical focus on the half-open door edge and the mouth of the gap, with the nearest margins falling gently soft so the eye is pulled straight into the black interior. Underexpose the surrounding structure into heavy desaturated shadow, preserving only sparse directional sheen across the corroded metal.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 3,
   "B": 8
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "레퍼런스 이미지에 확립된 장소의 건축 구조(흰색 패널 외벽, 창문)와 철문의 색상(녹색)을 전혀 일치시키지 못했습니다."
   },
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "레퍼런스의 배경 요소(흰색 벽, 창문, 녹색 철문)를 정확히 유지하면서, 요구된 1인칭 시점과 반쯤 열린 문틈의 어두운 구도를 훌륭하게 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S11sh4_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "문이 왼쪽 벽에 연결되어 열려 있는 상태인데, 손잡이가 경첩이 있어야 할 왼쪽 가장자리에 달려 있어 물리적인 구조가 어긋납니다.",
     "fix_en": "Move the doorknob from the left edge of the door to its right edge."
    },
    {
     "issue_ko": "레퍼런스 이미지에서 출입구 우측 상단 벽면에 부착되어 있던 조명 기구가 누락되었습니다.",
     "fix_en": "Add the wall-mounted light fixture above the right side of the doorway, matching the reference."
    },
    {
     "issue_ko": "레퍼런스 이미지에는 없었던 전등 스위치가 오른쪽 콘크리트 벽면에 임의로 추가되었습니다.",
     "fix_en": "Remove the light switch from the right concrete wall."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reframe tighter around the half-open door and dark entrance gap so they dominate the middle-center and middle-right, leaving only narrow rooftop margins while preserving the same eye-level frontal axis.",
     "issue_ko": "문 왼쪽의 창문 벽과 오른쪽 벽이 화면을 지나치게 넓게 차지해, 출입구 주변 옥상 공간이 좁은 여백으로만 남아야 한다는 구도와 다릅니다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Move the doorknob from the left edge of the door to its right edge.\n- Add the wall-mounted light fixture above the right side of the doorway, matching the reference.\n- Remove the light switch from the right concrete wall.\n- Reframe tighter around the half-open door and dark entrance gap so they dominate the middle-center and middle-right, leaving only narrow rooftop margins while preserving the same eye-level frontal axis.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "레퍼런스에 등장하는 우측 벽의 스위치, 문고리, 상단 조명의 방향 등 공간의 세부 요소를 훌륭하게 유지하며 시점 쇼트를 정확히 연출했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "우측 콘크리트 벽에 레퍼런스에 없는 조명 기구가 임의로 추가되었으며, 원래 있던 스위치와 문고리가 누락되어 장소 일관성을 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "레퍼런스에 등장하는 우측 벽의 스위치, 문고리, 상단 조명의 방향 등 공간의 세부 요소를 훌륭하게 유지하며 시점 쇼트를 정확히 연출했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "우측 콘크리트 벽에 레퍼런스에 없는 조명 기구가 임의로 추가되었으며, 원래 있던 스위치와 문고리가 누락되어 장소 일관성을 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "B는 레퍼런스에 있는 우측 콘크리트 벽면의 스위치와 문의 손잡이 등 고정된 디테일을 정확히 유지하여 완벽한 장소 일치도를 보여줍니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "A는 우측 벽면에 레퍼런스에 없는 조명 기구를 새로 추가하고 문의 손잡이를 누락하여 장소 일치도 조건에서 감점되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "B는 레퍼런스에 있는 우측 콘크리트 벽면의 스위치와 문의 손잡이 등 고정된 디테일을 정확히 유지하여 완벽한 장소 일치도를 보여줍니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "A는 우측 벽면에 레퍼런스에 없는 조명 기구를 새로 추가하고 문의 손잡이를 누락하여 장소 일치도 조건에서 감점되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 9
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S11sh4"
  },
  "lane_policy": "ab_select_bypass:bg_only:share_plan_prev_bgonly",
  "plate_select": {
   "candidates": {
    "A": "L05B01",
    "B": "L05B02"
   },
   "assigned": "L05B01",
   "choice": "A",
   "confident": true,
   "reason_ko": "\"녹슨 옥탑방 철문이 반쯤 휑하게 열려\"라는 묘사가 옥탑방 외부에서 반쯤 열린 철문을 보여주는 A와 정확히 일치합니다.",
   "chosen": "L05B01"
  }
 },
 "S12sh9::variants": {
  "author_fp": "a83408aca541ca80",
  "author": {
   "variants": [
    {
     "approach_ko": "깊은 초점과 절제된 구도로 바닥 전체를 냉정한 증거 지형처럼 포착한다.",
     "prompt_en": "Use a restrained, mildly wide lens feel with near edge-to-edge focus, turning the floor into a stark evidentiary field. Preserve the precise central and lower-left placement while letting the irregular spacing of the stains guide the eye deeper into the room; soft, low ambient spill reveals the pool’s realistic gloss without sensational highlights."
    },
    {
     "approach_ko": "얕은 초점과 촉각적인 표면 묘사로 응고된 피의 무게와 불길한 발견감을 강조한다.",
     "prompt_en": "Use a tighter, compressed lens feel with a narrow plane of focus concentrated on the central pool, allowing the peripheral stains and far floor to fall into gentle softness while remaining clearly legible in their prescribed positions. Emphasize the liquid’s thickness, uneven edge, and subdued crimson reflections against the desaturated floor, creating an intimate, ominous sense of the camera creeping toward the evidence."
    }
   ]
  },
  "reused": false
 },
 "S12sh9": {
  "input_fingerprint": "c62f473a44bff7d7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수리영 방의 바닥 여기저기에 붉은 핏자국이 흩뿌려져 있고 커다란 혈흔 웅덩이가 고인 바닥 앵글.\n\nLOCATION (lock): On the floor of the daughter’s bedroom inside the rooftop apartment, between the doorway and the curtained window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just above floor level inside 수리영's bedroom, the camera aims nearly perpendicular to the floor from the doorway and begins a controlled dolly toward the room's center. Scattered stains break across the outer frame while the large pool lies centrally, with enough surrounding floor retained to show scale and the route that will lead toward the curtain.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: large pool of blood in the middle-center of the frame, midground; scattered bloodstains in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: bedroom floor (Marked by scattered bloodstains and pooled blood); used as Forms the near-overhead field across which the dolly will later follow the evidence; large pool of blood (Pooled on the floor); used as Serves as the central discovery point while remaining surrounded by visible floor for spatial context; scattered bloodstains (Distributed across multiple areas of the floor); used as Creates an irregular trail across the composition and foreshadows the footprints toward the curtain.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key ambient spill appropriate to the bedroom keeps the blood's restrained crimson distinct against a dark, desaturated field.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood spatters, a large pool of blood, and bloody footprints leading to the curtain remain across Su-riyeong’s room. Behind the curtain, the already dead Minsuk must already be seated with her head hanging, her shoulder torn, her wrist mark intact, and the crumpled photograph clenched in her hand.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수리영 방의 바닥 여기저기에 붉은 핏자국이 흩뿌려져 있고 커다란 혈흔 웅덩이가 고인 바닥 앵글.\n\nLOCATION (lock): On the floor of the daughter’s bedroom inside the rooftop apartment, between the doorway and the curtained window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just above floor level inside 수리영's bedroom, the camera aims nearly perpendicular to the floor from the doorway and begins a controlled dolly toward the room's center. Scattered stains break across the outer frame while the large pool lies centrally, with enough surrounding floor retained to show scale and the route that will lead toward the curtain.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: large pool of blood in the middle-center of the frame, midground; scattered bloodstains in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: bedroom floor (Marked by scattered bloodstains and pooled blood); used as Forms the near-overhead field across which the dolly will later follow the evidence; large pool of blood (Pooled on the floor); used as Serves as the central discovery point while remaining surrounded by visible floor for spatial context; scattered bloodstains (Distributed across multiple areas of the floor); used as Creates an irregular trail across the composition and foreshadows the footprints toward the curtain.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key ambient spill appropriate to the bedroom keeps the blood's restrained crimson distinct against a dark, desaturated field.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood spatters, a large pool of blood, and bloody footprints leading to the curtain remain across Su-riyeong’s room. Behind the curtain, the already dead Minsuk must already be seated with her head hanging, her shoulder torn, her wrist mark intact, and the crumpled photograph clenched in her hand.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained, mildly wide lens feel with near edge-to-edge focus, turning the floor into a stark evidentiary field. Preserve the precise central and lower-left placement while letting the irregular spacing of the stains guide the eye deeper into the room; soft, low ambient spill reveals the pool’s realistic gloss without sensational highlights.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 수리영 방의 바닥 여기저기에 붉은 핏자국이 흩뿌려져 있고 커다란 혈흔 웅덩이가 고인 바닥 앵글.\n\nLOCATION (lock): On the floor of the daughter’s bedroom inside the rooftop apartment, between the doorway and the curtained window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just above floor level inside 수리영's bedroom, the camera aims nearly perpendicular to the floor from the doorway and begins a controlled dolly toward the room's center. Scattered stains break across the outer frame while the large pool lies centrally, with enough surrounding floor retained to show scale and the route that will lead toward the curtain.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: large pool of blood in the middle-center of the frame, midground; scattered bloodstains in the lower-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: bedroom floor (Marked by scattered bloodstains and pooled blood); used as Forms the near-overhead field across which the dolly will later follow the evidence; large pool of blood (Pooled on the floor); used as Serves as the central discovery point while remaining surrounded by visible floor for spatial context; scattered bloodstains (Distributed across multiple areas of the floor); used as Creates an irregular trail across the composition and foreshadows the footprints toward the curtain.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key ambient spill appropriate to the bedroom keeps the blood's restrained crimson distinct against a dark, desaturated field.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Blood spatters, a large pool of blood, and bloody footprints leading to the curtain remain across Su-riyeong’s room. Behind the curtain, the already dead Minsuk must already be seated with her head hanging, her shoulder torn, her wrist mark intact, and the crumpled photograph clenched in her hand.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a tighter, compressed lens feel with a narrow plane of focus concentrated on the central pool, allowing the peripheral stains and far floor to fall into gentle softness while remaining clearly legible in their prescribed positions. Emphasize the liquid’s thickness, uneven edge, and subdued crimson reflections against the desaturated floor, creating an intimate, ominous sense of the camera creeping toward the evidence.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 4,
   "B": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "레퍼런스 이미지와 일치하지 않는 가구가 배치되어 있으며, 지문에서 요구된 커튼으로 향하는 핏자국 발자국이 누락되었습니다."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "레퍼런스의 공간 구조(우측 문, 좌측 책상)를 잘 재현하였고, 바닥의 혈흔 웅덩이와 커튼으로 향하는 발자국 디테일을 충실히 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B08.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "바닥에 거의 수직으로 향하는 '인서트 클로즈업(insert close-up)' 앵글이 아니라 방 안쪽 벽과 가구가 모두 보이는 넓은 구도입니다.",
     "fix_en": "Change the camera to a top-down insert close-up angled nearly perpendicular to the floor, removing the background walls and furniture from the frame."
    },
    {
     "issue_ko": "맞은편 벽에 레퍼런스 이미지에 있던 창문이 없고, 공중에 짧은 커튼 조각만 허공에 떠 있습니다.",
     "fix_en": "Add the window to the far wall between the curtains and extend the curtains down to match the reference location photograph."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Tilt the camera steeply downward so its optical axis is nearly perpendicular to the floor, creating the specified near-overhead view while keeping it just above floor level.",
     "issue_ko": "카메라가 바닥을 거의 수직으로 내려다봐야 하지만, 선택 이미지는 뒷벽을 향한 낮은 수평 시점이라 바닥이 비스듬히 펼쳐진다."
    },
    {
     "fix_en": "Tighten the framing to an insert close-up dominated by the central pool and surrounding floor, minimizing the visible wall, doorway, and furniture.",
     "issue_ko": "지정된 디테일 인서트 클로즈업보다 화각이 넓어 뒷벽, 문, 가구가 크게 포함되고 중앙 혈흔 웅덩이의 화면 비중이 부족하다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the camera to a top-down insert close-up angled nearly perpendicular to the floor, removing the background walls and furniture from the frame.\n- Add the window to the far wall between the curtains and extend the curtains down to match the reference location photograph.\n- Tilt the camera steeply downward so its optical axis is nearly perpendicular to the floor, creating the specified near-overhead view while keeping it just above floor level.\n- Tighten the framing to an insert close-up dominated by the central pool and surrounding floor, minimizing the visible wall, doorway, and furniture.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 인서트 클로즈업보다 화각이 넓게 연출되었으나, 레퍼런스의 공간 구조(책상, 문, 창문 위치)와 야간 시간대, 핏자국의 구도를 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "화각은 더 타이트하지만, 창문이 바닥에 붙어 있는 물리적으로 불가능한 구조 왜곡(1순위 위반)이 발생했으며 야간 설정도 무시되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 인서트 클로즈업보다 화각이 넓게 연출되었으나, 레퍼런스의 공간 구조(책상, 문, 창문 위치)와 야간 시간대, 핏자국의 구도를 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "화각은 더 타이트하지만, 창문이 바닥에 붙어 있는 물리적으로 불가능한 구조 왜곡(1순위 위반)이 발생했으며 야간 설정도 무시되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "레퍼런스와 달리 창문이 바닥에 위치하는 심각한 건축적 오류가 있으며, 야간 설정에도 불구하고 창밖이 밝게 묘사되어 위치 및 시간대 조건을 위반했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 인서트 클로즈업 스케일보다는 화각이 넓지만, 레퍼런스의 방 구조(책상, 문, 커튼 위치)와 야간 조명, 혈흔의 정확한 배치를 훌륭하게 재현했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "레퍼런스와 달리 창문이 바닥에 위치하는 심각한 건축적 오류가 있으며, 야간 설정에도 불구하고 창밖이 밝게 묘사되어 위치 및 시간대 조건을 위반했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 인서트 클로즈업 스케일보다는 화각이 넓지만, 레퍼런스의 방 구조(책상, 문, 커튼 위치)와 야간 조명, 혈흔의 정확한 배치를 훌륭하게 재현했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 7
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B07",
   "choice": "Candidate G",
   "confident": true,
   "reason_ko": "지문에서 '수리영 방의 바닥'을 묘사하고 있으며, 현재 할당된 후보 G가 딸의 방 바닥 공간을 정확히 보여주고 있으므로 그대로 유지합니다.",
   "kept": "L06B07"
  }
 },
 "S12sh13::variants": {
  "author_fp": "50707365cb26886f",
  "author": {
   "variants": [
    {
     "approach_ko": "깊은 초점과 절제된 정면성으로 시신과 피 묻은 바닥을 냉정하게 기록하는 정적 구도.",
     "prompt_en": "Use a restrained, observational wide-angle feel with deep focus, holding the entire collapsed figure, displaced curtain edge, and bloodstained floor in crisp spatial relation. Keep the geometry still and clinically legible, with generous dark floor and room space surrounding her to intensify the cold finality. Let the low-key ambience fall evenly into desaturated shadow, preserving only subdued glossy crimson in the blood."
    },
    {
     "approach_ko": "커튼 가장자리를 가까운 전경층으로 삼고 민숙에게 시선을 집중시키는 은밀하고 불안한 발견의 구도.",
     "prompt_en": "Build a layered reveal with the pulled-aside curtain close to the camera as a soft, dark foreground boundary, while Minsuk remains fully visible and sharply resolved in the midground. Use a subtly compressed lens feel so her slack folded posture carries more visual weight against the dim bedroom depth without changing the wide framing. Allow the ambient light to skim the fabric folds, bowed head, torn shoulder area, marked wrist, crumpled photograph, and wet blood selectively, with the surrounding room receding into soft cold shadow."
    }
   ]
  },
  "reused": false
 },
 "S12sh13": {
  "input_fingerprint": "b924944513ee60e6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 바닥에 주저앉은 강민숙이 고개를 푹 숙인 채 미동 없이 축 늘어진 서늘한 구도.\n\nLOCATION (lock): Behind the curtain in the daughter’s bedroom, where the mother’s body is slumped on the floor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At knee height near the curtain, the completed pan pauses several feet from 강민숙 and angles slightly downward to preserve her entire seated figure. She occupies the lower-center in a three-quarter view, head and torso folded forward without movement, while the displaced curtain remains at one edge as the boundary that had concealed her.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 강민숙 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: curtain (Pulled aside) — The curtain has been displaced to one side, opening the camera's view toward 강민숙; used as Frames one side of the revealed body and records the barrier that was just pulled aside; bedroom floor (Bloodstained); used as Grounds her seated collapse and carries visible evidence from the preceding discovery.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key room ambience renders the body in cold, desaturated tones with restrained crimson confined to the visible blood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the bedroom floor, nighttime interior lighting, scattered blood, and pooled stain from the reference. Add the seated dead woman in the exact bloodied floor area, while excluding any unrelated furniture or figures outside this bedroom frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): She is seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward. One pale hand remains closed around a badly crumpled photograph, while the source does not specify the exact placement of her legs or other arm.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minsuk remains dead, seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward, with her shoulder and collarbone torn open. Her marked wrist remains visible, and one pale hand remains closed around a badly crumpled photograph.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 바닥에 주저앉은 강민숙이 고개를 푹 숙인 채 미동 없이 축 늘어진 서늘한 구도.\n\nLOCATION (lock): Behind the curtain in the daughter’s bedroom, where the mother’s body is slumped on the floor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At knee height near the curtain, the completed pan pauses several feet from 강민숙 and angles slightly downward to preserve her entire seated figure. She occupies the lower-center in a three-quarter view, head and torso folded forward without movement, while the displaced curtain remains at one edge as the boundary that had concealed her.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 강민숙 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: curtain (Pulled aside) — The curtain has been displaced to one side, opening the camera's view toward 강민숙; used as Frames one side of the revealed body and records the barrier that was just pulled aside; bedroom floor (Bloodstained); used as Grounds her seated collapse and carries visible evidence from the preceding discovery.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key room ambience renders the body in cold, desaturated tones with restrained crimson confined to the visible blood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the bedroom floor, nighttime interior lighting, scattered blood, and pooled stain from the reference. Add the seated dead woman in the exact bloodied floor area, while excluding any unrelated furniture or figures outside this bedroom frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): She is seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward. One pale hand remains closed around a badly crumpled photograph, while the source does not specify the exact placement of her legs or other arm.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minsuk remains dead, seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward, with her shoulder and collarbone torn open. Her marked wrist remains visible, and one pale hand remains closed around a badly crumpled photograph.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained, observational wide-angle feel with deep focus, holding the entire collapsed figure, displaced curtain edge, and bloodstained floor in crisp spatial relation. Keep the geometry still and clinically legible, with generous dark floor and room space surrounding her to intensify the cold finality. Let the low-key ambience fall evenly into desaturated shadow, preserving only subdued glossy crimson in the blood.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 바닥에 주저앉은 강민숙이 고개를 푹 숙인 채 미동 없이 축 늘어진 서늘한 구도.\n\nLOCATION (lock): Behind the curtain in the daughter’s bedroom, where the mother’s body is slumped on the floor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At knee height near the curtain, the completed pan pauses several feet from 강민숙 and angles slightly downward to preserve her entire seated figure. She occupies the lower-center in a three-quarter view, head and torso folded forward without movement, while the displaced curtain remains at one edge as the boundary that had concealed her.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 강민숙 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: curtain (Pulled aside) — The curtain has been displaced to one side, opening the camera's view toward 강민숙; used as Frames one side of the revealed body and records the barrier that was just pulled aside; bedroom floor (Bloodstained); used as Grounds her seated collapse and carries visible evidence from the preceding discovery.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key room ambience renders the body in cold, desaturated tones with restrained crimson confined to the visible blood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the bedroom floor, nighttime interior lighting, scattered blood, and pooled stain from the reference. Add the seated dead woman in the exact bloodied floor area, while excluding any unrelated furniture or figures outside this bedroom frame.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): She is seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward. One pale hand remains closed around a badly crumpled photograph, while the source does not specify the exact placement of her legs or other arm.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minsuk remains dead, seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward, with her shoulder and collarbone torn open. Her marked wrist remains visible, and one pale hand remains closed around a badly crumpled photograph.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 강민숙 (한국인 40대 초반 여성 얼굴, 부드러운 중년 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nBuild a layered reveal with the pulled-aside curtain close to the camera as a soft, dark foreground boundary, while Minsuk remains fully visible and sharply resolved in the midground. Use a subtly compressed lens feel so her slack folded posture carries more visual weight against the dim bedroom depth without changing the wide framing. Allow the ambient light to skim the fabric folds, bowed head, torn shoulder area, marked wrist, crumpled photograph, and wet blood selectively, with the surrounding room receding into soft cold shadow.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "레퍼런스의 바닥 혈흔을 정확히 유지했으며, 찢어진 어깨, 손목 자국, 구겨진 사진 등 요구된 디테일과 축 늘어진 시신의 자세를 충실히 구현했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "이전 샷의 필수 요소인 바닥의 혈흔이 완전히 누락되었고, 찢어진 어깨와 손목 자국 등의 지시사항도 반영되지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S12sh9_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 강민숙: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1275879>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 양쪽에 커튼이 배치되어 있으나, 프롬프트는 커튼이 한쪽 가장자리에만 밀려 있어야 한다고 명시했습니다.",
     "fix_en": "Remove the curtain from the right edge of the frame so it only appears on one side."
    },
    {
     "issue_ko": "어깨와 쇄골 부위가 찢겨 열려 있어야(torn open) 하지만, 스웨터에 구멍만 나 있을 뿐 노출된 피부에는 상처가 없습니다.",
     "fix_en": "Add a severe, bloody flesh wound to the exposed skin on her shoulder and collarbone."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Move the pulled-aside curtain entirely to one edge of the frame so it forms a single foreground boundary, leaving the opposite edge unobstructed.",
     "issue_ko": "커튼이 한쪽 가장자리만 경계로 남아야 하는데, 양쪽 가장자리에서 인물을 둘러싸는 대칭 프레임처럼 배치되어 있다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the curtain from the right edge of the frame so it only appears on one side.\n- Add a severe, bloody flesh wound to the exposed skin on her shoulder and collarbone.\n- Move the pulled-aside curtain entirely to one edge of the frame so it forms a single foreground boundary, leaving the opposite edge unobstructed.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 상처 난 손목 디테일이 정확히 묘사되었으며, 전체적인 구도와 늘어진 자세가 지침을 잘 따랐습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "구도와 자세는 훌륭하나, 프롬프트에 명시된 '상처 난 손목(marked wrist)' 디테일이 누락되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 상처 난 손목 디테일이 정확히 묘사되었으며, 전체적인 구도와 늘어진 자세가 지침을 잘 따랐습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "구도와 자세는 훌륭하나, 프롬프트에 명시된 '상처 난 손목(marked wrist)' 디테일이 누락되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 앵글과 찢어진 어깨, 구겨진 사진 등 디테일을 잘 구현했으며, 커튼을 한쪽에만 배치하여 공간 레퍼런스에 부합합니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "손목의 자국 등은 표현되었으나, 우측에 프롬프트와 레퍼런스에 없는 커튼이 추가되어 '한쪽 가장자리의 커튼' 구도 지시를 어겼습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 앵글과 찢어진 어깨, 구겨진 사진 등 디테일을 잘 구현했으며, 커튼을 한쪽에만 배치하여 공간 레퍼런스에 부합합니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "손목의 자국 등은 표현되었으나, 우측에 프롬프트와 레퍼런스에 없는 커튼이 추가되어 '한쪽 가장자리의 커튼' 구도 지시를 어겼습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 12
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S12sh9"
  }
 },
 "S12sh19::variants": {
  "author_fp": "c2f0708b80799bf1",
  "author": {
   "variants": [
    {
     "approach_ko": "손과 사진의 표면을 스치는 얕은 초점의 촉각적이고 불안한 접근.",
     "prompt_en": "Use an intimate, slightly raking close-up with very shallow depth of field, letting the nearest finger edges soften while the sharpest plane catches the photograph’s crushed central folds and the pressure of the closed grip. Shape the established low-key illumination to graze across skin and paper, emphasizing tactile ridges, creases, and compressed depth with restrained desaturated contrast."
    },
    {
     "approach_ko": "손아귀와 사진의 구조를 또렷하게 드러내는 냉정하고 정적인 관찰.",
     "prompt_en": "Photograph the composition with a restrained, near-planar perspective and broader depth of field, holding the enclosing fingers and overlapping paper folds in firm simultaneous clarity. Keep the image still and clinically observational, with soft, even falloff from the established desaturated illumination and controlled shadows that reveal the geometry of the grip without making the pictured content legible."
    }
   ]
  },
  "reused": false
 },
 "S12sh19": {
  "input_fingerprint": "af5927f4b18f33dc",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 숨진 강민숙의 창백한 손아귀 안에 심하게 구겨진 사진 한 장이 쥐어져 있는 클로즈업.\n\nLOCATION (lock): Beside the body behind the curtain in the daughter’s bedroom, focused on the dead woman’s clenched hand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height beside 강민숙, the completed dolly-in settles on her closed fingers and the photograph they enclose, excluding nearly all of her body. Her pale fingers occupy the frame edges while the severely crumpled photograph covers roughly the central third, preserving enough of the hand to establish its human scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: crumpled photograph (Severely crumpled and held inside the closed hand) — The camera-facing image side is crushed into overlapping folds, leaving its pictured content largely unreadable; used as Central focal clue enclosed by the dead woman's fingers.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination isolates the pale hand and compressed photograph without introducing an unestablished light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dead Minsuk remains seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward, with the crumpled photograph locked in her pale hand and the torn shoulder wound unchanged. Her longstanding wrist mark remains present and turns red when Su-riyeong removes the photograph.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 숨진 강민숙의 창백한 손아귀 안에 심하게 구겨진 사진 한 장이 쥐어져 있는 클로즈업.\n\nLOCATION (lock): Beside the body behind the curtain in the daughter’s bedroom, focused on the dead woman’s clenched hand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height beside 강민숙, the completed dolly-in settles on her closed fingers and the photograph they enclose, excluding nearly all of her body. Her pale fingers occupy the frame edges while the severely crumpled photograph covers roughly the central third, preserving enough of the hand to establish its human scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: crumpled photograph (Severely crumpled and held inside the closed hand) — The camera-facing image side is crushed into overlapping folds, leaving its pictured content largely unreadable; used as Central focal clue enclosed by the dead woman's fingers.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination isolates the pale hand and compressed photograph without introducing an unestablished light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dead Minsuk remains seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward, with the crumpled photograph locked in her pale hand and the torn shoulder wound unchanged. Her longstanding wrist mark remains present and turns red when Su-riyeong removes the photograph.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse an intimate, slightly raking close-up with very shallow depth of field, letting the nearest finger edges soften while the sharpest plane catches the photograph’s crushed central folds and the pressure of the closed grip. Shape the established low-key illumination to graze across skin and paper, emphasizing tactile ridges, creases, and compressed depth with restrained desaturated contrast.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 숨진 강민숙의 창백한 손아귀 안에 심하게 구겨진 사진 한 장이 쥐어져 있는 클로즈업.\n\nLOCATION (lock): Beside the body behind the curtain in the daughter’s bedroom, focused on the dead woman’s clenched hand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height beside 강민숙, the completed dolly-in settles on her closed fingers and the photograph they enclose, excluding nearly all of her body. Her pale fingers occupy the frame edges while the severely crumpled photograph covers roughly the central third, preserving enough of the hand to establish its human scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: crumpled photograph (Severely crumpled and held inside the closed hand) — The camera-facing image side is crushed into overlapping folds, leaving its pictured content largely unreadable; used as Central focal clue enclosed by the dead woman's fingers.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination isolates the pale hand and compressed photograph without introducing an unestablished light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dead Minsuk remains seated collapsed on the floor behind the curtain with her upper body hanging slack and her head bowed deeply forward, with the crumpled photograph locked in her pale hand and the torn shoulder wound unchanged. Her longstanding wrist mark remains present and turns red when Su-riyeong removes the photograph.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph the composition with a restrained, near-planar perspective and broader depth of field, holding the enclosing fingers and overlapping paper folds in firm simultaneous clarity. Keep the image still and clinically observational, with soft, even falloff from the established desaturated illumination and controlled shadows that reveal the geometry of the grip without making the pictured content legible.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 8,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "레퍼런스와 완벽하게 일치하는 단일 손의 자세, 배경(바닥과 피, 커튼), 그리고 심하게 구겨진 사진을 클로즈업 앵글로 정확히 연출했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "레퍼런스와 달리 두 손이 사진을 함께 쥐고 있는 구도로 연출되어 이전 샷의 자세 및 설정과 심각하게 모순됩니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S12sh13_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 샷의 고정된 자세를 유지해야 하나, 레퍼런스 이미지에서 손등이 위를 향한 채 엎어져 있던 손의 자세와 달리 손바닥이 위를 향하고 있어 지시를 위반했습니다.",
     "fix_en": "Rotate the hand so it rests palm-down on the floor with the back of the hand facing upward, exactly matching the dead woman's locked pose in the reference."
    },
    {
     "issue_ko": "프롬프트에 명시된 '오래된 손목 자국(longstanding wrist mark)'이 프레임 안의 노출된 손목 부위에 묘사되지 않았습니다.",
     "fix_en": "Add a clearly visible, longstanding mark or scar on the exposed skin of the wrist."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Rotate and reposition the hand to preserve the exact palm-down/side-on orientation and floor contact established in the previous shot, while keeping the photograph clenched in place.",
     "issue_ko": "손이 이전 쇼트에서 고정된 손등·측면 위주의 바닥 접촉 자세가 아니라 손바닥이 위로 드러난 방향으로 뒤집혀 있어, 움직일 수 없는 시신의 손 자세와 방향이 이어지지 않는다."
    },
    {
     "fix_en": "Push the camera closer and reframe so the enclosing fingers reach the frame edges while the crumpled photograph remains in the central third.",
     "issue_ko": "프레이밍이 다소 넓어 손가락들이 프레임 가장자리를 차지하지 못하고 왼쪽과 위쪽에 불필요한 배경 공간이 많이 남아 있다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Rotate the hand so it rests palm-down on the floor with the back of the hand facing upward, exactly matching the dead woman's locked pose in the reference.\n- Add a clearly visible, longstanding mark or scar on the exposed skin of the wrist.\n- Rotate and reposition the hand to preserve the exact palm-down/side-on orientation and floor contact established in the previous shot, while keeping the photograph clenched in place.\n- Push the camera closer and reframe so the enclosing fingers reach the frame edges while the crumpled photograph remains in the central third.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 구도와 조명을 잘 따랐으며, 특히 프롬프트에 명시된 손목 자국을 명확하게 묘사하여 디테일에서 앞섭니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "전반적인 구성과 분위기는 훌륭하게 구현되었으나, 텍스트에서 요구한 손목 자국이 누락되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 구도와 조명을 잘 따랐으며, 특히 프롬프트에 명시된 손목 자국을 명확하게 묘사하여 디테일에서 앞섭니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "전반적인 구성과 분위기는 훌륭하게 구현되었으나, 텍스트에서 요구한 손목 자국이 누락되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "클로즈업 구도와 조명 지침을 잘 따랐으며, 프롬프트에 명시된 손목의 흉터 자국을 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지정된 구도와 배경 요소는 잘 반영했으나, 명시된 손목 자국(wrist mark)이 뚜렷하게 나타나지 않았습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "클로즈업 구도와 조명 지침을 잘 따랐으며, 프롬프트에 명시된 손목의 흉터 자국을 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 구도와 배경 요소는 잘 반영했으나, 명시된 손목 자국(wrist mark)이 뚜렷하게 나타나지 않았습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S12sh13"
  },
  "lane_policy": "share_plan_prev_bgonly",
  "locked_char_refs_excluded": [
   "강민숙(C04)"
  ],
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B07",
   "choice": "Candidate G",
   "confident": true,
   "reason_ko": "딸의 침실(침대, 책상, 커튼이 있는 방)이 배경이므로, 현재 할당된 후보가 적합합니다.",
   "kept": "L06B07"
  }
 },
 "S13sh2::bgfirst_bg": {
  "input_fingerprint": "7dd3f663f9632bc9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 붉고 푸른 사이렌 불빛이 번지는 경찰차 문을 연 채 한쪽 발을 아스팔트에 내딛고 있는 혜수의 측면.\n\nLOCATION (lock): On the street outside the multi-unit villa, beside a police car flashing red and blue lights.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From hip height beside the police vehicle, the static camera holds perpendicular to 혜수's exit path as she plants one foot on the asphalt. Her profile occupies one side of the medium frame while the open door remains beside her, with the active roof lights spreading red and blue across both figure and vehicle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: police vehicle (Door open, siren and roof lights active) — The vehicle's door side faces the camera, with the open door revealing 혜수's exit path; used as Retained beside 혜수 as a spatial frame and the grounded source of the colored pulses; asphalt (Visible beneath the open vehicle door); used as Ground plane beneath 혜수's newly planted foot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Active police lights cast alternating red and blue pulses through an otherwise dark, desaturated nighttime exterior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 붉고 푸른 사이렌 불빛이 번지는 경찰차 문을 연 채 한쪽 발을 아스팔트에 내딛고 있는 혜수의 측면.\n\nLOCATION (lock): On the street outside the multi-unit villa, beside a police car flashing red and blue lights.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From hip height beside the police vehicle, the static camera holds perpendicular to 혜수's exit path as she plants one foot on the asphalt. Her profile occupies one side of the medium frame while the open door remains beside her, with the active roof lights spreading red and blue across both figure and vehicle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: police vehicle (Door open, siren and roof lights active) — The vehicle's door side faces the camera, with the open door revealing 혜수's exit path; used as Retained beside 혜수 as a spatial frame and the grounded source of the colored pulses; asphalt (Visible beneath the open vehicle door); used as Ground plane beneath 혜수's newly planted foot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Active police lights cast alternating red and blue pulses through an otherwise dark, desaturated nighttime exterior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S13sh2__bgfirst_bg.png",
  "asset_id": "2d8d9fa6-e8c7-4369-8fac-f81823c817d5",
  "input_asset_ids": [
   "b368b77c-b3d4-444d-9764-3822608f1f61",
   "27ba994a-0cf5-4a3f-982f-d13502b6b4e2"
  ]
 },
 "S13sh2": {
  "input_fingerprint": "73a276f3a5375374",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉고 푸른 사이렌 불빛이 번지는 경찰차 문을 연 채 한쪽 발을 아스팔트에 내딛고 있는 혜수의 측면.\n\nLOCATION (lock): On the street outside the multi-unit villa, beside a police car flashing red and blue lights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From hip height beside the police vehicle, the static camera holds perpendicular to 혜수's exit path as she plants one foot on the asphalt. Her profile occupies one side of the medium frame while the open door remains beside her, with the active roof lights spreading red and blue across both figure and vehicle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: police vehicle (Door open, siren and roof lights active) — The vehicle's door side faces the camera, with the open door revealing 혜수's exit path; used as Retained beside 혜수 as a spatial frame and the grounded source of the colored pulses; asphalt (Visible beneath the open vehicle door); used as Ground plane beneath 혜수's newly planted foot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Active police lights cast alternating red and blue pulses through an otherwise dark, desaturated nighttime exterior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉고 푸른 사이렌 불빛이 번지는 경찰차 문을 연 채 한쪽 발을 아스팔트에 내딛고 있는 혜수의 측면.\n\nLOCATION (lock): On the street outside the multi-unit villa, beside a police car flashing red and blue lights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From hip height beside the police vehicle, the static camera holds perpendicular to 혜수's exit path as she plants one foot on the asphalt. Her profile occupies one side of the medium frame while the open door remains beside her, with the active roof lights spreading red and blue across both figure and vehicle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: police vehicle (Door open, siren and roof lights active) — The vehicle's door side faces the camera, with the open door revealing 혜수's exit path; used as Retained beside 혜수 as a spatial frame and the grounded source of the colored pulses; asphalt (Visible beneath the open vehicle door); used as Ground plane beneath 혜수's newly planted foot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Active police lights cast alternating red and blue pulses through an otherwise dark, desaturated nighttime exterior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 붉고 푸른 사이렌 불빛이 번지는 경찰차 문을 연 채 한쪽 발을 아스팔트에 내딛고 있는 혜수의 측면.\n\nLOCATION (lock): On the street outside the multi-unit villa, beside a police car flashing red and blue lights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From hip height beside the police vehicle, the static camera holds perpendicular to 혜수's exit path as she plants one foot on the asphalt. Her profile occupies one side of the medium frame while the open door remains beside her, with the active roof lights spreading red and blue across both figure and vehicle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: police vehicle (Door open, siren and roof lights active) — The vehicle's door side faces the camera, with the open door revealing 혜수's exit path; used as Retained beside 혜수 as a spatial frame and the grounded source of the colored pulses; asphalt (Visible beneath the open vehicle door); used as Ground plane beneath 혜수's newly planted foot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Active police lights cast alternating red and blue pulses through an otherwise dark, desaturated nighttime exterior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S13sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S13sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L05B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 로케이션 레퍼런스(계단, 자전거 보관소, 대문 등)를 완벽하게 구현했으며 물리적으로 올바른 차량 구조와 조명을 보여주지만, 미디엄 샷 지정보다 프레임이 다소 넓습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "차에서 내리는 동작의 찰나는 잘 포착했으나, 경찰차 뒷문이 반대로 열리는 구조적 오류(수어사이드 도어)가 있고 지정된 장소의 건축물 및 구조물 디테일을 거의 일치시키지 못했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 로케이션 레퍼런스(계단, 자전거 보관소, 대문 등)를 완벽하게 구현했으며 물리적으로 올바른 차량 구조와 조명을 보여주지만, 미디엄 샷 지정보다 프레임이 다소 넓습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "차에서 내리는 동작의 찰나는 잘 포착했으나, 경찰차 뒷문이 반대로 열리는 구조적 오류(수어사이드 도어)가 있고 지정된 장소의 건축물 및 구조물 디테일을 거의 일치시키지 못했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "차 문을 열고 하차하며 한쪽 발을 아스팔트에 딛는 찰나의 동작과 인물 배치를 지시대로 정확히 구현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "배경은 원본과 완벽히 일치하나, 이미 차에서 내려 걷고 있어 '한쪽 발을 내딛는' 하차 순간의 핵심 액션을 놓쳤습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "차 문을 열고 하차하며 한쪽 발을 아스팔트에 딛는 찰나의 동작과 인물 배치를 지시대로 정확히 구현했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "배경은 원본과 완벽히 일치하나, 이미 차에서 내려 걷고 있어 '한쪽 발을 내딛는' 하차 순간의 핵심 액션을 놓쳤습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 11
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "totals": {
   "A": 13,
   "B": 11
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "지정된 로케이션 레퍼런스(계단, 자전거 보관소, 대문 등)를 완벽하게 구현했으며 물리적으로 올바른 차량 구조와 조명을 보여주지만, 미디엄 샷 지정보다 프레임이 다소 넓습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "차에서 내리는 동작의 찰나는 잘 포착했으나, 경찰차 뒷문이 반대로 열리는 구조적 오류(수어사이드 도어)가 있고 지정된 장소의 건축물 및 구조물 디테일을 거의 일치시키지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L05B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:900092>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "인물이 레이아웃 스케치처럼 차체와 열린 문 사이에서 내리고 있지 않고, 차에서 완전히 벗어나 앞을 향해 걷고 있습니다.",
     "fix_en": "Move the character back into the space between the car chassis and the open door, posing her mid-exit with one foot planting on the asphalt as sketched."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Re-stage Hyesu at the open doorway with only her rear foot planted on the asphalt and her forward foot still lifted mid-exit, preserving her right-facing profile and everything else.",
     "issue_ko": "혜수가 이미 두 발을 모두 아스팔트에 딛고 차에서 멀어지며 걷고 있어, 스케치의 한 발만 막 내딛고 다른 발은 들린 하차 순간과 다릅니다."
    },
    {
     "fix_en": "Restore the exact original background framing and aspect ratio without the top-and-bottom crop, preserving all subject staging relative to that frame.",
     "issue_ko": "선택 이미지가 원본 배경보다 위아래가 잘린 더 넓은 화면비로 재구성되어, 고정된 카메라 프레이밍이 달라졌습니다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Move the character back into the space between the car chassis and the open door, posing her mid-exit with one foot planting on the asphalt as sketched.\n- Re-stage Hyesu at the open doorway with only her rear foot planted on the asphalt and her forward foot still lifted mid-exit, preserving her right-facing profile and everything else.\n- Restore the exact original background framing and aspect ratio without the top-and-bottom crop, preserving all subject staging relative to that frame.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 배경과 인물의 외형을 잘 유지했으며, 사이렌 조명의 색감과 스케치의 전반적인 구도를 자연스럽게 구현했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "열린 경찰차 문 아래쪽 공간에 인물의 세 번째 발이 잘못 생성되어 물리적으로 불가능한 신체 구조를 보여줍니다 (Hard Violation)."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 배경과 인물의 외형을 잘 유지했으며, 사이렌 조명의 색감과 스케치의 전반적인 구도를 자연스럽게 구현했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "열린 경찰차 문 아래쪽 공간에 인물의 세 번째 발이 잘못 생성되어 물리적으로 불가능한 신체 구조를 보여줍니다 (Hard Violation)."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "스케치된 위치에 인물을 자연스럽게 배치하였으며, 지시된 조명과 배경, 캐릭터의 디테일을 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 뒤편 차량 하단에 세 번째 다리가 생성되는 치명적인 신체 구조 오류가 있어 감점되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "스케치된 위치에 인물을 자연스럽게 배치하였으며, 지시된 조명과 배경, 캐릭터의 디테일을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물의 뒤편 차량 하단에 세 번째 다리가 생성되는 치명적인 신체 구조 오류가 있어 감점되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 5,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S13sh2__bgfirst_bg.png",
   "bg_asset_id": "2d8d9fa6-e8c7-4369-8fac-f81823c817d5",
   "bg_record_key": "S13sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L05B01",
    "B": "L05B02"
   },
   "assigned": "L05B02",
   "choice": "Candidate B",
   "confident": true,
   "reason_ko": "경찰차 문을 열고 아스팔트에 발을 내딛는 장면이므로, 경찰차가 주차된 빌라 앞 진입로가 보이는 Candidate B가 적합합니다.",
   "kept": "L05B02"
  }
 },
 "S13sh5::variants": {
  "author_fp": "7ef263331ac56cde",
  "author": {
   "variants": [
    {
     "approach_ko": "극도로 얕은 심도로 수리영의 절망적인 옆얼굴에 밀착하고 혜수는 부드럽게 흐려진 채 시선 관계만 남긴다.",
     "prompt_en": "Use an intimate, compressed close-up with extremely shallow depth of field, holding critical focus on Suriyoung’s raised side profile while Hyesu remains softly recognizable at the opposite edge. Let the intermittent siren color skim across Suriyoung’s pallor in restrained, shifting bands, with the exterior falling into subdued nighttime softness so her desperate performance dominates."
    },
    {
     "approach_ko": "두 인물의 옆얼굴과 상하 시선축을 함께 선명하게 살려 공간적 압박과 대치감을 강조한다.",
     "prompt_en": "Favor a more dimensional close-up with enough depth to keep both opposing profiles legible, emphasizing the steep upward-downward eyeline and the narrow spatial tension between them. Shape the intermittent red and blue illumination as soft separation along their facial contours, retaining restrained exterior detail and desaturated textures for a stark, observational sense of shock."
    }
   ]
  },
  "reused": false
 },
 "S13sh5": {
  "input_fingerprint": "1ac236b8f7895537",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 혜수의 얼굴을 올려다보고 있는 수리영의 절망적인 얼굴 측면.\n\nLOCATION (lock): In the villa’s exterior approach near the courtyard and rooftop stairway, where the shaken daughter faces the arriving officer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The dolly-in ends at 수리영's head height in a tight side-profile close-up, angled upward so her raised face holds the foreground while 혜수 remains above her eyeline at the opposite edge. 수리영's face takes most of the frame, with only enough of 혜수 retained to make the desperate upward exchange physically legible.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, foreground, looks toward 혜수 above her eyeline; 혜수 in the upper-right of the frame, midground, looks toward 수리영 below her eyeline.\n- KEY BACKGROUND ELEMENTS: police siren illumination (Active in the exterior); used as Soft, intermittent background color separating the two profiles without displacing their eyeline exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Red and blue siren pulses intermittently cross the otherwise desaturated close-up, sharpening 수리영's pallor and 혜수's shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the police-car-side pavement, flashing red-and-blue light, and nighttime exterior look from the reference. Exclude the prior open-car-door action and the officer stepping out; frame the distressed young woman looking upward toward the officer.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 혜수의 얼굴을 올려다보고 있는 수리영의 절망적인 얼굴 측면.\n\nLOCATION (lock): In the villa’s exterior approach near the courtyard and rooftop stairway, where the shaken daughter faces the arriving officer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The dolly-in ends at 수리영's head height in a tight side-profile close-up, angled upward so her raised face holds the foreground while 혜수 remains above her eyeline at the opposite edge. 수리영's face takes most of the frame, with only enough of 혜수 retained to make the desperate upward exchange physically legible.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, foreground, looks toward 혜수 above her eyeline; 혜수 in the upper-right of the frame, midground, looks toward 수리영 below her eyeline.\n- KEY BACKGROUND ELEMENTS: police siren illumination (Active in the exterior); used as Soft, intermittent background color separating the two profiles without displacing their eyeline exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Red and blue siren pulses intermittently cross the otherwise desaturated close-up, sharpening 수리영's pallor and 혜수's shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the police-car-side pavement, flashing red-and-blue light, and nighttime exterior look from the reference. Exclude the prior open-car-door action and the officer stepping out; frame the distressed young woman looking upward toward the officer.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse an intimate, compressed close-up with extremely shallow depth of field, holding critical focus on Suriyoung’s raised side profile while Hyesu remains softly recognizable at the opposite edge. Let the intermittent siren color skim across Suriyoung’s pallor in restrained, shifting bands, with the exterior falling into subdued nighttime softness so her desperate performance dominates.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 혜수의 얼굴을 올려다보고 있는 수리영의 절망적인 얼굴 측면.\n\nLOCATION (lock): In the villa’s exterior approach near the courtyard and rooftop stairway, where the shaken daughter faces the arriving officer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The dolly-in ends at 수리영's head height in a tight side-profile close-up, angled upward so her raised face holds the foreground while 혜수 remains above her eyeline at the opposite edge. 수리영's face takes most of the frame, with only enough of 혜수 retained to make the desperate upward exchange physically legible.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, foreground, looks toward 혜수 above her eyeline; 혜수 in the upper-right of the frame, midground, looks toward 수리영 below her eyeline.\n- KEY BACKGROUND ELEMENTS: police siren illumination (Active in the exterior); used as Soft, intermittent background color separating the two profiles without displacing their eyeline exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Red and blue siren pulses intermittently cross the otherwise desaturated close-up, sharpening 수리영's pallor and 혜수's shock.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the police-car-side pavement, flashing red-and-blue light, and nighttime exterior look from the reference. Exclude the prior open-car-door action and the officer stepping out; frame the distressed young woman looking upward toward the officer.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nFavor a more dimensional close-up with enough depth to keep both opposing profiles legible, emphasizing the steep upward-downward eyeline and the narrow spatial tension between them. Shape the intermittent red and blue illumination as soft separation along their facial contours, retaining restrained exterior detail and desaturated textures for a stark, observational sense of shock.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 9,
   "A": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 9,
    "verdict_ko": "지시된 타이트한 클로즈업 샷의 비율을 잘 따랐으며, 특히 레퍼런스 이미지의 핵심 배경인 계단 구조물을 정확히 포함하여 장소적 연속성(Location lock)을 완벽히 구현했습니다."
   },
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "인물들의 배치와 감정선은 좋으나, 프레이밍이 지시된 것보다 넓어 두 인물이 비슷한 비중을 차지하며 레퍼런스의 특징적인 계단 배경이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S13sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "왼쪽의 수리영이 캐릭터 레퍼런스 이미지에서 착용하고 있는 어두운 회색 비니를 쓰고 있지 않습니다.",
     "fix_en": "Add the dark grey beanie from the character reference to Suri-young's head."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Add the dark grey beanie from the character reference to Suri-young's head.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지정된 클로즈업 앵글과 시선 교차를 정확히 구현했으며, 이전 샷 레퍼런스에 나타난 혜수의 복장(비니 없음) 연속성을 충실히 유지했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "구도와 조명은 훌륭하나, 이전 샷 레퍼런스에서 비니를 쓰지 않았던 혜수에게 임의로 비니를 추가하여 이전 샷과의 연속성 규칙을 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지정된 클로즈업 앵글과 시선 교차를 정확히 구현했으며, 이전 샷 레퍼런스에 나타난 혜수의 복장(비니 없음) 연속성을 충실히 유지했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "구도와 조명은 훌륭하나, 이전 샷 레퍼런스에서 비니를 쓰지 않았던 혜수에게 임의로 비니를 추가하여 이전 샷과의 연속성 규칙을 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도와 조명은 훌륭하나, 이전 샷 레퍼런스에 없던 비니를 경찰관(우측)에게 씌워 복장 연속성 오류가 발생했습니다."
     },
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "지시된 클로즈업 구도와 시선 교환을 완벽히 구현했으며, 이전 샷의 경찰관 복장 및 헤어스타일 연속성도 정확하게 유지했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "구도와 조명은 훌륭하나, 이전 샷 레퍼런스에 없던 비니를 경찰관(우측)에게 씌워 복장 연속성 오류가 발생했습니다."
     },
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지시된 클로즈업 구도와 시선 교환을 완벽히 구현했으며, 이전 샷의 경찰관 복장 및 헤어스타일 연속성도 정확하게 유지했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 14
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S13sh2"
  }
 },
 "S14sh5::bgfirst_bg": {
  "input_fingerprint": "e0fbb85afdb8312e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 깨끗한 방 한가운데 선 수리영의 얼굴이 파리하게 질린 채 입술이 살짝 벌어진 정면.\n\nLOCATION (lock): In the center of the unnaturally clean daughter’s bedroom inside the rooftop apartment, with all blood and the body gone.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 수리영's eye level, the forward move settles into a centered, straight-on close-up as she freezes in the cleared room. Her pale face occupies most of the frame, while the orderly room remains narrowly visible around her as negative evidence rather than competing detail.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: bedroom floor (Clean, with no bloodstains visible); used as Visible around 수리영 as cleared negative space where the blood evidence had been; bed (Orderly, with no red mark visible on the wall above it) — Its side and upper area are visible behind 수리영; used as Narrowly retained behind 수리영 to show the missing supernatural mark; area behind the curtain (Empty; 강민숙's body is gone); used as Background area confirming that the body is absent.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination renders 수리영's face pallid against the unnaturally clean room without specifying an unestablished source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 깨끗한 방 한가운데 선 수리영의 얼굴이 파리하게 질린 채 입술이 살짝 벌어진 정면.\n\nLOCATION (lock): In the center of the unnaturally clean daughter’s bedroom inside the rooftop apartment, with all blood and the body gone.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 수리영's eye level, the forward move settles into a centered, straight-on close-up as she freezes in the cleared room. Her pale face occupies most of the frame, while the orderly room remains narrowly visible around her as negative evidence rather than competing detail.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: bedroom floor (Clean, with no bloodstains visible); used as Visible around 수리영 as cleared negative space where the blood evidence had been; bed (Orderly, with no red mark visible on the wall above it) — Its side and upper area are visible behind 수리영; used as Narrowly retained behind 수리영 to show the missing supernatural mark; area behind the curtain (Empty; 강민숙's body is gone); used as Background area confirming that the body is absent.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination renders 수리영's face pallid against the unnaturally clean room without specifying an unestablished source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S14sh5__bgfirst_bg.png",
  "asset_id": "b284ce1e-244d-42f2-baa3-db43b6676d8c",
  "input_asset_ids": [
   "4ffcd298-4e22-46ea-b996-0afd0de7f224",
   "5c2e0c8a-0cb4-4739-bf68-e458d9307cec"
  ]
 },
 "S14sh5": {
  "input_fingerprint": "94fd43ce56010d42",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 깨끗한 방 한가운데 선 수리영의 얼굴이 파리하게 질린 채 입술이 살짝 벌어진 정면.\n\nLOCATION (lock): In the center of the unnaturally clean daughter’s bedroom inside the rooftop apartment, with all blood and the body gone. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 수리영's eye level, the forward move settles into a centered, straight-on close-up as she freezes in the cleared room. Her pale face occupies most of the frame, while the orderly room remains narrowly visible around her as negative evidence rather than competing detail.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: bedroom floor (Clean, with no bloodstains visible); used as Visible around 수리영 as cleared negative space where the blood evidence had been; bed (Orderly, with no red mark visible on the wall above it) — Its side and upper area are visible behind 수리영; used as Narrowly retained behind 수리영 to show the missing supernatural mark; area behind the curtain (Empty; 강민숙's body is gone); used as Background area confirming that the body is absent.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination renders 수리영's face pallid against the unnaturally clean room without specifying an unestablished source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s room is now unnaturally clean: the blood, bloody footprints, red wall mark, and Minsuk’s body have all vanished.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 깨끗한 방 한가운데 선 수리영의 얼굴이 파리하게 질린 채 입술이 살짝 벌어진 정면.\n\nLOCATION (lock): In the center of the unnaturally clean daughter’s bedroom inside the rooftop apartment, with all blood and the body gone. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 수리영's eye level, the forward move settles into a centered, straight-on close-up as she freezes in the cleared room. Her pale face occupies most of the frame, while the orderly room remains narrowly visible around her as negative evidence rather than competing detail.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: bedroom floor (Clean, with no bloodstains visible); used as Visible around 수리영 as cleared negative space where the blood evidence had been; bed (Orderly, with no red mark visible on the wall above it) — Its side and upper area are visible behind 수리영; used as Narrowly retained behind 수리영 to show the missing supernatural mark; area behind the curtain (Empty; 강민숙's body is gone); used as Background area confirming that the body is absent.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination renders 수리영's face pallid against the unnaturally clean room without specifying an unestablished source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s room is now unnaturally clean: the blood, bloody footprints, red wall mark, and Minsuk’s body have all vanished.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 깨끗한 방 한가운데 선 수리영의 얼굴이 파리하게 질린 채 입술이 살짝 벌어진 정면.\n\nLOCATION (lock): In the center of the unnaturally clean daughter’s bedroom inside the rooftop apartment, with all blood and the body gone. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 수리영's eye level, the forward move settles into a centered, straight-on close-up as she freezes in the cleared room. Her pale face occupies most of the frame, while the orderly room remains narrowly visible around her as negative evidence rather than competing detail.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: bedroom floor (Clean, with no bloodstains visible); used as Visible around 수리영 as cleared negative space where the blood evidence had been; bed (Orderly, with no red mark visible on the wall above it) — Its side and upper area are visible behind 수리영; used as Narrowly retained behind 수리영 to show the missing supernatural mark; area behind the curtain (Empty; 강민숙's body is gone); used as Background area confirming that the body is absent.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination renders 수리영's face pallid against the unnaturally clean room without specifying an unestablished source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s room is now unnaturally clean: the blood, bloody footprints, red wall mark, and Minsuk’s body have all vanished.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S14sh5__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S14sh5.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B08.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지정된 클로즈업 앵글, 창백한 안색과 살짝 벌어진 입술, 그리고 원본 레퍼런스 방의 구조를 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물의 표정과 구도는 좋으나, 방 오른쪽에 존재하지 않는 침대가 추가되어 공간 구조를 심각하게 훼손했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지정된 클로즈업 앵글, 창백한 안색과 살짝 벌어진 입술, 그리고 원본 레퍼런스 방의 구조를 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물의 표정과 구도는 좋으나, 방 오른쪽에 존재하지 않는 침대가 추가되어 공간 구조를 심각하게 훼손했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 클로즈업 앵글과 창백한 표정을 잘 표현했으며, 레퍼런스의 방 구조와 가구 배치를 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물의 구도와 표정은 준수하나, 우측 배경에 레퍼런스에 없는 두 번째 침대를 생성하여 장소 일관성을 크게 훼손했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 클로즈업 앵글과 창백한 표정을 잘 표현했으며, 레퍼런스의 방 구조와 가구 배치를 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물의 구도와 표정은 준수하나, 우측 배경에 레퍼런스에 없는 두 번째 침대를 생성하여 장소 일관성을 크게 훼손했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 16,
     "B": 8
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 16,
   "B": 8
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "지정된 클로즈업 앵글, 창백한 안색과 살짝 벌어진 입술, 그리고 원본 레퍼런스 방의 구조를 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "인물의 표정과 구도는 좋으나, 방 오른쪽에 존재하지 않는 침대가 추가되어 공간 구조를 심각하게 훼손했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B08.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "배경 원본에 있던 책상 앞 나무 의자가 사라졌으며, 책상 위의 책과 물건들이 원본과 다르게 변형되었습니다.",
     "fix_en": "Restore the wooden chair at the desk and the exact arrangement of books and items on the desk to match the background reference perfectly."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Restore the wooden chair at the desk and the exact arrangement of books and items on the desk to match the background reference perfectly.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "정면 클로즈업, 창백한 얼굴과 살짝 벌어진 입술, 중앙 배치가 정확하며 B보다 원본 방의 책상·책·가구 구성을 더 충실히 유지했다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "인물의 구도와 연기는 정확하지만 책상 위 책을 가로로 쌓고 의자·책상 배치를 바꾸어 고정 배경을 그대로 유지하라는 핵심 조건을 위반했다."
     }
    ],
    "winner": "A"
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "정면 클로즈업, 창백한 얼굴과 살짝 벌어진 입술, 중앙 배치가 정확하며 B보다 원본 방의 책상·책·가구 구성을 더 충실히 유지했다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "인물의 구도와 연기는 정확하지만 책상 위 책을 가로로 쌓고 의자·책상 배치를 바꾸어 고정 배경을 그대로 유지하라는 핵심 조건을 위반했다."
     }
    ]
   },
   "reverse_raw": {
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "정중앙 정면 클로즈업, 창백한 얼굴과 살짝 벌어진 입술, 깨끗한 침실의 좁은 배경 노출을 정확히 구현하며 원본 방의 원근과 배치도 가장 충실하다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "표정과 클로즈업은 충실하지만 책상·의자와 책들이 원본보다 왼쪽으로 재배치·과다 노출되어 고정 배경을 그대로 유지하라는 조건을 위반한다."
     }
    ],
    "winner": "B"
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "정중앙 정면 클로즈업, 창백한 얼굴과 살짝 벌어진 입술, 깨끗한 침실의 좁은 배경 노출을 정확히 구현하며 원본 방의 원근과 배치도 가장 충실하다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "표정과 클로즈업은 충실하지만 책상·의자와 책들이 원본보다 왼쪽으로 재배치·과다 노출되어 고정 배경을 그대로 유지하라는 조건을 위반한다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 18,
     "B": 14
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S14sh5__bgfirst_bg.png",
   "bg_asset_id": "b284ce1e-244d-42f2-baa3-db43b6676d8c",
   "bg_record_key": "S14sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B08",
   "choice": "Candidate H",
   "confident": true,
   "reason_ko": "\"깨끗한 방 한가운데\"라는 묘사에 맞게 잘 정돈된 침실 공간이 배경이 되어야 합니다.",
   "kept": "L06B08"
  }
 },
 "S14sh9::bgfirst_bg": {
  "input_fingerprint": "c2b43d43a0d1aa07",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고개 숙인 수리영의 어깨 위로 혜수의 손이 따뜻하게 얹혀진 클로즈업.\n\nLOCATION (lock): Inside the rooftop apartment near the daughter’s bedroom and dining area, where the officer comforts her after the search.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From shoulder height just behind and to the side of 수리영, the completed dolly-in crops tightly across her bowed shoulder toward 혜수's resting hand. The sloping shoulder fills the lower half while 혜수's hand sits near center, with both faces largely excluded so the reassuring contact becomes the sole emotional action.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained, desaturated low-key illumination keeps the hand and bowed shoulder gently separated while preserving the scene's oppressive unease.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고개 숙인 수리영의 어깨 위로 혜수의 손이 따뜻하게 얹혀진 클로즈업.\n\nLOCATION (lock): Inside the rooftop apartment near the daughter’s bedroom and dining area, where the officer comforts her after the search.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From shoulder height just behind and to the side of 수리영, the completed dolly-in crops tightly across her bowed shoulder toward 혜수's resting hand. The sloping shoulder fills the lower half while 혜수's hand sits near center, with both faces largely excluded so the reassuring contact becomes the sole emotional action.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained, desaturated low-key illumination keeps the hand and bowed shoulder gently separated while preserving the scene's oppressive unease.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S14sh9__bgfirst_bg.png",
  "asset_id": "bc78e2c3-59af-4221-92fc-682d22276ece",
  "input_asset_ids": [
   "21d37fca-2d34-4b52-a0a3-4c0272c72cf2",
   "5c2e0c8a-0cb4-4739-bf68-e458d9307cec"
  ]
 },
 "S14sh9": {
  "input_fingerprint": "0a4f517be6730428",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개 숙인 수리영의 어깨 위로 혜수의 손이 따뜻하게 얹혀진 클로즈업.\n\nLOCATION (lock): Inside the rooftop apartment near the daughter’s bedroom and dining area, where the officer comforts her after the search. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From shoulder height just behind and to the side of 수리영, the completed dolly-in crops tightly across her bowed shoulder toward 혜수's resting hand. The sloping shoulder fills the lower half while 혜수's hand sits near center, with both faces largely excluded so the reassuring contact becomes the sole emotional action.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained, desaturated low-key illumination keeps the hand and bowed shoulder gently separated while preserving the scene's oppressive unease.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains unnaturally clean, with the two teacups gone; the backpack and doll in Su-riyeong’s room are subtly displaced from their earlier positions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개 숙인 수리영의 어깨 위로 혜수의 손이 따뜻하게 얹혀진 클로즈업.\n\nLOCATION (lock): Inside the rooftop apartment near the daughter’s bedroom and dining area, where the officer comforts her after the search. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From shoulder height just behind and to the side of 수리영, the completed dolly-in crops tightly across her bowed shoulder toward 혜수's resting hand. The sloping shoulder fills the lower half while 혜수's hand sits near center, with both faces largely excluded so the reassuring contact becomes the sole emotional action.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained, desaturated low-key illumination keeps the hand and bowed shoulder gently separated while preserving the scene's oppressive unease.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains unnaturally clean, with the two teacups gone; the backpack and doll in Su-riyeong’s room are subtly displaced from their earlier positions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 고개 숙인 수리영의 어깨 위로 혜수의 손이 따뜻하게 얹혀진 클로즈업.\n\nLOCATION (lock): Inside the rooftop apartment near the daughter’s bedroom and dining area, where the officer comforts her after the search. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From shoulder height just behind and to the side of 수리영, the completed dolly-in crops tightly across her bowed shoulder toward 혜수's resting hand. The sloping shoulder fills the lower half while 혜수's hand sits near center, with both faces largely excluded so the reassuring contact becomes the sole emotional action.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained, desaturated low-key illumination keeps the hand and bowed shoulder gently separated while preserving the scene's oppressive unease.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains unnaturally clean, with the two teacups gone; the backpack and doll in Su-riyeong’s room are subtly displaced from their earlier positions.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S14sh9__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S14sh9.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B08.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 타이트한 클로즈업 프레이밍을 무시하고 화면을 넓게 잡아 혜수의 얼굴 전체와 방 내부를 크게 노출하여 카메라 및 프레임 지침을 심각하게 위반했습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "얼굴을 대부분 배제하고 수리영의 어깨와 그 위에 얹힌 혜수의 손을 화면 중심에 꽉 차게 담아내어, 요구된 클로즈업 프레이밍과 감정적 연출 지침을 완벽하게 충족했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 타이트한 클로즈업 프레이밍을 무시하고 화면을 넓게 잡아 혜수의 얼굴 전체와 방 내부를 크게 노출하여 카메라 및 프레임 지침을 심각하게 위반했습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "얼굴을 대부분 배제하고 수리영의 어깨와 그 위에 얹힌 혜수의 손을 화면 중심에 꽉 차게 담아내어, 요구된 클로즈업 프레이밍과 감정적 연출 지침을 완벽하게 충족했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "프롬프트가 지시한 대로 두 인물의 얼굴을 대부분 배제하고 수리영의 어깨와 혜수의 손에 집중하는 완벽한 클로즈업 앵글과 샷 스케일을 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "얼굴을 배제하고 어깨와 손에 집중해야 하는 타이트한 클로즈업 지시를 무시하고, 넓은 화각으로 방 전체와 혜수의 얼굴까지 노출하여 프레임 구성을 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "프롬프트가 지시한 대로 두 인물의 얼굴을 대부분 배제하고 수리영의 어깨와 혜수의 손에 집중하는 완벽한 클로즈업 앵글과 샷 스케일을 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "얼굴을 배제하고 어깨와 손에 집중해야 하는 타이트한 클로즈업 지시를 무시하고, 넓은 화각으로 방 전체와 혜수의 얼굴까지 노출하여 프레임 구성을 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 19
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 8,
   "B": 19
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "지시된 타이트한 클로즈업 프레이밍을 무시하고 화면을 넓게 잡아 혜수의 얼굴 전체와 방 내부를 크게 노출하여 카메라 및 프레임 지침을 심각하게 위반했습니다."
   },
   {
    "label": "B",
    "score": 9,
    "verdict_ko": "얼굴을 대부분 배제하고 수리영의 어깨와 그 위에 얹힌 혜수의 손을 화면 중심에 꽉 차게 담아내어, 요구된 클로즈업 프레이밍과 감정적 연출 지침을 완벽하게 충족했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B08.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   },
   {
    "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:900092>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "어깨에 얹힌 혜수의 손에 엄지를 제외하고도 5개의 손가락이 그려져 있어 해부학적으로 잘못되었습니다.",
     "fix_en": "Remove the lowest extra finger resting on the shoulder so the hand displays a normal four fingers."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the lowest extra finger resting on the shoulder so the hand displays a normal four fingers.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지정된 클로즈업 앵글과 인물의 배치, 억눌린 조명 분위기를 완벽하게 구현했으며, 프레임의 핵심인 손의 해부학적 디테일이 매우 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "프레이밍과 톤은 프롬프트의 요구사항을 잘 충족하지만, 피사체의 중심인 손의 검지손가락 마디와 피부 텍스처에서 미세한 해부학적 왜곡이 관찰됩니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지정된 클로즈업 앵글과 인물의 배치, 억눌린 조명 분위기를 완벽하게 구현했으며, 프레임의 핵심인 손의 해부학적 디테일이 매우 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "프레이밍과 톤은 프롬프트의 요구사항을 잘 충족하지만, 피사체의 중심인 손의 검지손가락 마디와 피부 텍스처에서 미세한 해부학적 왜곡이 관찰됩니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지시된 앵글과 샷 크기를 완벽하게 구현했으며, 인물의 자세, 배경, 그리고 손의 해부학적 구조가 모두 자연스럽게 연출되었습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "어깨 위에 올려진 손의 손가락이 6개(보이는 것만 5개)로 묘사되어 물리적으로 불가능한 신체 구조를 보여주는 하드 위반입니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "지시된 앵글과 샷 크기를 완벽하게 구현했으며, 인물의 자세, 배경, 그리고 손의 해부학적 구조가 모두 자연스럽게 연출되었습니다."
     },
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "어깨 위에 올려진 손의 손가락이 6개(보이는 것만 5개)로 묘사되어 물리적으로 불가능한 신체 구조를 보여주는 하드 위반입니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 19
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S14sh9__bgfirst_bg.png",
   "bg_asset_id": "bc78e2c3-59af-4221-92fc-682d22276ece",
   "bg_record_key": "S14sh9::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B08",
   "choice": "Candidate H",
   "confident": true,
   "reason_ko": "클로즈업 숏으로 배경이 구체적으로 명시되지 않았으므로, 현재 배정된 침실(Candidate H)을 유지함.",
   "kept": "L06B08"
  }
 },
 "groupbg::해안_정류장": {
  "input_fingerprint": "193651275afa99bc",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "해안_정류장",
    "tags": [
     "S15sh1",
     "S15sh4",
     "S15sh6"
    ]
   },
   "context_sig": "f627edeacff0d2f4"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On a bench at an outdoor coastal bus stop, facing the dark night sea.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n해안가 버스정류장과 인도: 바다가 정면으로 보이는 해안도로 변의 버스정류장. 낡은 벤치와 정류장 구조물이 서 있다. (특징: 투명 아크릴과 철골로 된 정류장 부스; 플라스틱 의자와 낡은 정류장 표지판; 배경으로 보이는 방파제와 바다)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#15. 해안가. 버스정류장 부근 / N\n- 바다를 바라보며 멍한 채로 앉아있는 수리영\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On a bench at an outdoor coastal bus stop, facing the dark night sea.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n해안가 버스정류장과 인도: 바다가 정면으로 보이는 해안도로 변의 버스정류장. 낡은 벤치와 정류장 구조물이 서 있다. (특징: 투명 아크릴과 철골로 된 정류장 부스; 플라스틱 의자와 낡은 정류장 표지판; 배경으로 보이는 방파제와 바다)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#15. 해안가. 버스정류장 부근 / N\n- 바다를 바라보며 멍한 채로 앉아있는 수리영\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_해안_정류장_4344c8.png",
  "asset_id": "8e219d20-8f4f-477e-bd00-b0635c7afb5f",
  "input_asset_ids": [
   "1c0a0efd-bcaa-4629-9905-850e46ea4b65"
  ],
  "origin_tag": "S15sh1",
  "place_text": "On a bench at an outdoor coastal bus stop, facing the dark night sea.",
  "origin_inputs": {
   "place_text": "On a bench at an outdoor coastal bus stop, facing the dark night sea.",
   "time_of_day_en": "night",
   "conti_asset_id": "1c0a0efd-bcaa-4629-9905-850e46ea4b65"
  }
 },
 "S15sh1::bgfirst_bg": {
  "input_fingerprint": "1bb08209aab9c529",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 밤바다를 배경으로 버스정류장 벤치에 앉아 멍하니 앞을 보는 수리영의 측면.\n\nLOCATION (lock): On a bench at an outdoor coastal bus stop, facing the dark night sea.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level beside the bus-stop bench, the static medium composition looks across 수리영's profile toward the night sea. She sits off-center with her folded posture contained by the bench, while the sea occupies the open background in the direction of her unfocused gaze.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward night sea; night sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: bus-stop bench (Occupied by 수리영) — The bench is seen from the side, aligned beneath 수리영's seated profile; used as Supports 수리영's seated profile and establishes the bus-stop location; sea (Visible at night beyond the bench); used as Open background beyond 수리영 and the destination of her distant stare.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key nighttime ambience and a dark, desaturated palette hold 수리영 against the sea without assigning an unestablished artificial source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 밤바다를 배경으로 버스정류장 벤치에 앉아 멍하니 앞을 보는 수리영의 측면.\n\nLOCATION (lock): On a bench at an outdoor coastal bus stop, facing the dark night sea.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level beside the bus-stop bench, the static medium composition looks across 수리영's profile toward the night sea. She sits off-center with her folded posture contained by the bench, while the sea occupies the open background in the direction of her unfocused gaze.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward night sea; night sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: bus-stop bench (Occupied by 수리영) — The bench is seen from the side, aligned beneath 수리영's seated profile; used as Supports 수리영's seated profile and establishes the bus-stop location; sea (Visible at night beyond the bench); used as Open background beyond 수리영 and the destination of her distant stare.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key nighttime ambience and a dark, desaturated palette hold 수리영 against the sea without assigning an unestablished artificial source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S15sh1__bgfirst_bg.png",
  "asset_id": "d0b96b9a-6cb2-422c-aa02-1fa21f9eaaef",
  "input_asset_ids": [
   "1c0a0efd-bcaa-4629-9905-850e46ea4b65",
   "8e219d20-8f4f-477e-bd00-b0635c7afb5f"
  ]
 },
 "S15sh1": {
  "input_fingerprint": "e4497a69ce7c69a0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤바다를 배경으로 버스정류장 벤치에 앉아 멍하니 앞을 보는 수리영의 측면.\n\nLOCATION (lock): On a bench at an outdoor coastal bus stop, facing the dark night sea. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level beside the bus-stop bench, the static medium composition looks across 수리영's profile toward the night sea. She sits off-center with her folded posture contained by the bench, while the sea occupies the open background in the direction of her unfocused gaze.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward night sea; night sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: bus-stop bench (Occupied by 수리영) — The bench is seen from the side, aligned beneath 수리영's seated profile; used as Supports 수리영's seated profile and establishes the bus-stop location; sea (Visible at night beyond the bench); used as Open background beyond 수리영 and the destination of her distant stare.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key nighttime ambience and a dark, desaturated palette hold 수리영 against the sea without assigning an unestablished artificial source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains her phone and continues receiving no answer from Minsuk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤바다를 배경으로 버스정류장 벤치에 앉아 멍하니 앞을 보는 수리영의 측면.\n\nLOCATION (lock): On a bench at an outdoor coastal bus stop, facing the dark night sea. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level beside the bus-stop bench, the static medium composition looks across 수리영's profile toward the night sea. She sits off-center with her folded posture contained by the bench, while the sea occupies the open background in the direction of her unfocused gaze.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward night sea; night sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: bus-stop bench (Occupied by 수리영) — The bench is seen from the side, aligned beneath 수리영's seated profile; used as Supports 수리영's seated profile and establishes the bus-stop location; sea (Visible at night beyond the bench); used as Open background beyond 수리영 and the destination of her distant stare.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key nighttime ambience and a dark, desaturated palette hold 수리영 against the sea without assigning an unestablished artificial source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains her phone and continues receiving no answer from Minsuk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 밤바다를 배경으로 버스정류장 벤치에 앉아 멍하니 앞을 보는 수리영의 측면.\n\nLOCATION (lock): On a bench at an outdoor coastal bus stop, facing the dark night sea. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level beside the bus-stop bench, the static medium composition looks across 수리영's profile toward the night sea. She sits off-center with her folded posture contained by the bench, while the sea occupies the open background in the direction of her unfocused gaze.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward night sea; night sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: bus-stop bench (Occupied by 수리영) — The bench is seen from the side, aligned beneath 수리영's seated profile; used as Supports 수리영's seated profile and establishes the bus-stop location; sea (Visible at night beyond the bench); used as Open background beyond 수리영 and the destination of her distant stare.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key nighttime ambience and a dark, desaturated palette hold 수리영 against the sea without assigning an unestablished artificial source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains her phone and continues receiving no answer from Minsuk.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S15sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S15sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_해안_정류장_4344c8.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 앵글과 배경 구도를 정확히 구현했으며, 레퍼런스의 의상(비니 포함)과 소품(스마트폰)까지 가장 완벽하게 일치합니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도와 배경, 분위기는 매우 우수하나 캐릭터 레퍼런스에 있는 비니를 누락하여 복장 일치도가 다소 떨어집니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 앵글과 배경 구도를 정확히 구현했으며, 레퍼런스의 의상(비니 포함)과 소품(스마트폰)까지 가장 완벽하게 일치합니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도와 배경, 분위기는 매우 우수하나 캐릭터 레퍼런스에 있는 비니를 누락하여 복장 일치도가 다소 떨어집니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "미디엄 샷 프레이밍을 정확히 구사했으며, 캐릭터 레퍼런스의 의상(비니 포함)과 휴대폰을 쥐고 있는 상태를 충실하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "로케이션과 기본적인 구도는 잘 따랐으나, 캐릭터 레퍼런스의 핵심 의상인 비니가 누락되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "미디엄 샷 프레이밍을 정확히 구사했으며, 캐릭터 레퍼런스의 의상(비니 포함)과 휴대폰을 쥐고 있는 상태를 충실하게 재현했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "로케이션과 기본적인 구도는 잘 따랐으나, 캐릭터 레퍼런스의 핵심 의상인 비니가 누락되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 12,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "지정된 앵글과 배경 구도를 정확히 구현했으며, 레퍼런스의 의상(비니 포함)과 소품(스마트폰)까지 가장 완벽하게 일치합니다."
   },
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "구도와 배경, 분위기는 매우 우수하나 캐릭터 레퍼런스에 있는 비니를 누락하여 복장 일치도가 다소 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_해안_정류장_4344c8.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "휴대폰을 들고 있는 인물의 왼손 해부학 구조가 잘못되었습니다 (엄지손가락이 바깥쪽을 향하고 있어 오른손처럼 보임).",
     "fix_en": "Redraw the left hand with correct anatomy so the thumb is on the inside, towards the character's body."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the left hand with correct anatomy so the thumb is on the inside, towards the character's body.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "인물, 배경, 야간 조명, 구도 및 휴대폰 소지 상태까지 프롬프트의 모든 지시사항을 완벽하게 구현했으며 세부 묘사가 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "지시문과 레퍼런스를 훌륭하게 반영했으나, 휴대폰을 쥐고 있는 손의 형태가 A에 비해 미세하게 덜 자연스럽습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "인물, 배경, 야간 조명, 구도 및 휴대폰 소지 상태까지 프롬프트의 모든 지시사항을 완벽하게 구현했으며 세부 묘사가 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "지시문과 레퍼런스를 훌륭하게 반영했으나, 휴대폰을 쥐고 있는 손의 형태가 A에 비해 미세하게 덜 자연스럽습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 미디엄 샷 프레이밍, 인물과 장소의 레퍼런스, 그리고 핸드폰을 쥐고 있는 상태까지 정확하게 구현되었습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "A와 시각적으로 동일한 수준으로 프롬프트의 지시사항(프레이밍, 장소, 인물, 소품)을 충실히 반영했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 미디엄 샷 프레이밍, 인물과 장소의 레퍼런스, 그리고 핸드폰을 쥐고 있는 상태까지 정확하게 구현되었습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "A와 시각적으로 동일한 수준으로 프롬프트의 지시사항(프레이밍, 장소, 인물, 소품)을 충실히 반영했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 17,
     "B": 16
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S15sh1__bgfirst_bg.png",
   "bg_asset_id": "d0b96b9a-6cb2-422c-aa02-1fa21f9eaaef",
   "bg_record_key": "S15sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "해안_정류장",
   "groupbg_asset_id": "8e219d20-8f4f-477e-bd00-b0635c7afb5f"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S15sh4::variants": {
  "author_fp": "6c30c1e543c70f30",
  "author": {
   "variants": [
    {
     "approach_ko": "깊은 초점과 강한 소실점으로 수리영의 고립된 질주와 떠남의 거리를 강조한다.",
     "prompt_en": "Use a restrained wide-angle perspective with deep focus, holding the centered departure axis in clean, severe symmetry. Keep Su-riyeong’s complete airborne stride sharply resolved while the receding bench and converging coastal path create a strong sense of distance behind and ahead of her. Let the sparse nighttime ambience fall off naturally into desaturated shadow, with subtle tonal separation around her clothing and hair rather than added illumination."
    },
    {
     "approach_ko": "추적 촬영의 미세한 흔들림과 주변부 운동감으로 다급한 질주의 체감을 살린다.",
     "prompt_en": "Photograph the run with urgent tracking energy: Su-riyeong’s centered back and suspended foot remain legible, while slight directional motion softness gathers along the path and the receding bench. Favor a more compressed, layered depth feeling within the wide frame, making her movement seem forceful against the dark route opening ahead. Preserve natural low-light texture, restrained contrast, and a faintly imperfect live-action immediacy."
    }
   ]
  },
  "reused": false
 },
 "S15sh4": {
  "input_fingerprint": "6d49071504ef7851",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 해안가 인도를 따라 달려가는 수리영의 뒷모습, 한 발이 지면에서 막 떨어져 허공에 떠 있는 역동적인 달리기 자세.\n\nLOCATION (lock): Along the dark seaside sidewalk beside the bus stop, running away toward home. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Several steps behind 수리영 at waist height, the tracking camera lowers just enough to retain her complete running stride as she moves away from the bench. Her back remains centered with one foot suspended above the ground, while the path opens ahead and the bench recedes without breaking the forward axis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, moves toward path ahead; path ahead in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: coastal path (Extending away from the bus-stop area); used as Defines the centered departure axis beneath 수리영's full stride; bus-stop bench (Left behind as 수리영 runs away) — Its long side falls behind the camera-following axis; used as Receding spatial reference marking the point 수리영 has left.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark, desaturated nighttime ambience keeps the moving figure legible against the coastal path without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong keeps her phone with her as she leaves the bus stop and runs home.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 해안가 인도를 따라 달려가는 수리영의 뒷모습, 한 발이 지면에서 막 떨어져 허공에 떠 있는 역동적인 달리기 자세.\n\nLOCATION (lock): Along the dark seaside sidewalk beside the bus stop, running away toward home. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Several steps behind 수리영 at waist height, the tracking camera lowers just enough to retain her complete running stride as she moves away from the bench. Her back remains centered with one foot suspended above the ground, while the path opens ahead and the bench recedes without breaking the forward axis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, moves toward path ahead; path ahead in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: coastal path (Extending away from the bus-stop area); used as Defines the centered departure axis beneath 수리영's full stride; bus-stop bench (Left behind as 수리영 runs away) — Its long side falls behind the camera-following axis; used as Receding spatial reference marking the point 수리영 has left.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark, desaturated nighttime ambience keeps the moving figure legible against the coastal path without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong keeps her phone with her as she leaves the bus stop and runs home.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained wide-angle perspective with deep focus, holding the centered departure axis in clean, severe symmetry. Keep Su-riyeong’s complete airborne stride sharply resolved while the receding bench and converging coastal path create a strong sense of distance behind and ahead of her. Let the sparse nighttime ambience fall off naturally into desaturated shadow, with subtle tonal separation around her clothing and hair rather than added illumination.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 해안가 인도를 따라 달려가는 수리영의 뒷모습, 한 발이 지면에서 막 떨어져 허공에 떠 있는 역동적인 달리기 자세.\n\nLOCATION (lock): Along the dark seaside sidewalk beside the bus stop, running away toward home. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Several steps behind 수리영 at waist height, the tracking camera lowers just enough to retain her complete running stride as she moves away from the bench. Her back remains centered with one foot suspended above the ground, while the path opens ahead and the bench recedes without breaking the forward axis.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, moves toward path ahead; path ahead in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: coastal path (Extending away from the bus-stop area); used as Defines the centered departure axis beneath 수리영's full stride; bus-stop bench (Left behind as 수리영 runs away) — Its long side falls behind the camera-following axis; used as Receding spatial reference marking the point 수리영 has left.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark, desaturated nighttime ambience keeps the moving figure legible against the coastal path without adding an unsupported light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong keeps her phone with her as she leaves the bus stop and runs home.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph the run with urgent tracking energy: Su-riyeong’s centered back and suspended foot remain legible, while slight directional motion softness gathers along the path and the receding bench. Favor a more compressed, layered depth feeling within the wide frame, making her movement seem forceful against the dark route opening ahead. Preserve natural low-light texture, restrained contrast, and a faintly imperfect live-action immediacy.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "요청된 앵글과 역동적인 달리기 자세를 잘 구현했으며, 손에 든 휴대폰 상태도 정확히 유지했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "달리는 자세와 구도는 좋으나, 휴대폰을 들고 있다는 지시를 누락했고 기둥에 텍스트 형태의 찌그러짐이 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S15sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "캐릭터의 자세가 꼿꼿하게 선 채로 허공에 떠 있는 듯하며, 프롬프트에서 요구한 무릎을 굽히고 팔을 젓는 '역동적인 달리기 자세'가 아닙니다.",
     "fix_en": "Change the character's posture to a dynamic running stride with a forward-leaning torso, bent knees, and active arm swinging."
    },
    {
     "issue_ko": "캐릭터가 굽이 둥글고 끈이 없는 로우탑 신발을 신고 있어, 캐릭터 레퍼런스에 있는 끈이 달린 검은색 하이탑 스니커즈와 일치하지 않습니다.",
     "fix_en": "Change the character's shoes to the laced black high-top sneakers shown in the character reference."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Stage Su-riyeong in a clearly dynamic running stride, with a stronger forward lean, active arm drive, and one foot visibly airborne, while preserving her centered midground placement and the existing camera framing.",
     "issue_ko": "수리영의 상체가 곧고 팔 동작이 약해 역동적인 달리기보다 걷는 자세로 보인다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the character's posture to a dynamic running stride with a forward-leaning torso, bent knees, and active arm swinging.\n- Change the character's shoes to the laced black high-top sneakers shown in the character reference.\n- Stage Su-riyeong in a clearly dynamic running stride, with a stronger forward lean, active arm drive, and one foot visibly airborne, while preserving her centered midground placement and the existing camera framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "'역동적인 달리기 자세'라는 지시를 충족하지 못하고 공중에 뜬 다리가 뻣뻣하고 부자연스럽게 묘사되었습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 넓은 샷 스케일과 역동적인 달리기 자세, 그리고 들고 있는 스마트폰의 상태를 모두 정확하고 자연스럽게 구현했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "'역동적인 달리기 자세'라는 지시를 충족하지 못하고 공중에 뜬 다리가 뻣뻣하고 부자연스럽게 묘사되었습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 넓은 샷 스케일과 역동적인 달리기 자세, 그리고 들고 있는 스마트폰의 상태를 모두 정확하고 자연스럽게 구현했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "역동적인 달리기 자세와 공중에 뜬 발의 묘사가 프롬프트의 요구사항을 사실적으로 잘 충족했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "다리 자세가 경직되어 있어 역동적인 달리기라기보다는 부자연스럽게 떠 있는 것처럼 보입니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "역동적인 달리기 자세와 공중에 뜬 발의 묘사가 프롬프트의 요구사항을 사실적으로 잘 충족했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "다리 자세가 경직되어 있어 역동적인 달리기라기보다는 부자연스럽게 떠 있는 것처럼 보입니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh1"
  }
 },
 "S15sh6::variants": {
  "author_fp": "eadd829d25257137",
  "author": {
   "variants": [
    {
     "approach_ko": "긴 렌즈의 압축감과 날카로운 차폐물 윤곽으로 포착한 은밀한 감시 시점.",
     "prompt_en": "Use a restrained long-lens feel to compress the distant action into the narrow openings, with the dark structural edges sharply defined as rigid masks around the view. Keep the distant figures legible but small, arranging the open gaps so the eye travels laterally from the bench-side watcher to the suspended running bound. Sparse, low-key illumination and subdued contrast create the cold precision of covert surveillance rather than an expressive action image."
    },
    {
     "approach_ko": "넓은 렌즈와 거대한 암부 전경을 강조해 숨어 엿보는 공간감을 살린 시점.",
     "prompt_en": "Use a wider lens feel that makes the nearby structures loom as heavy, irregular fields of darkness while the revealed sidewalk and bench recede deeply beyond them. Let the foreground occluders fall slightly soft at their closest edges, directing attention through one restricted corridor toward the two small, separated figures; preserve clear spatial depth between concealment, bench, and departing runner. Shape the available night light into faint pools and edge glints, leaving most of the frame in textured shadow for a tense, predatory sense of hidden presence."
    }
   ]
  },
  "reused": false
 },
 "S15sh6": {
  "input_fingerprint": "32563ce2a460650d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 구조물 사이로 멀리서 달리기 자세로 도약 중인 수리영과 벤치 앞에 서 있는 혜수의 모습이 은밀하게 담긴 누군가의 시점 쇼트.\n\nLOCATION (lock): From concealment among the dark exterior structures near the coastal bus stop, watching the bench and adjacent sidewalk from a distance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: A hard cut relocates the camera far from the bench and path at standing eye level, peering through narrow gaps between dark structures. 수리영 appears small in the distant right side during the next running bound, while 혜수 remains by the bench on the left within the same concealed wide frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, background, moves toward path away from the bench; 혜수 in the middle-left of the frame, background, looks toward 수리영 running away; bus-stop bench in the lower-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: structures (Partially occluding the view) — Their near sides face the camera while narrow gaps reveal the distant path and bench; used as Large foreground occluders whose gaps create the hidden observer's restricted view; bus-stop bench (Unoccupied, with 혜수 positioned in front of it) — Its side is visible beside 혜수 in the distant left portion of the frame; used as Distant anchor fixing 혜수's position relative to the departing 수리영.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The desaturated nighttime view remains low-key, with foreground structures held dark enough to make the concealed observation feel predatory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the dark seaside, bus-stop bench, pavement, and nighttime lighting from the reference. Exclude the prior seated version of the young woman; place the officer by the bench and the young woman farther away in a running stride, viewed through dark foreground structures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is still running away with her phone, while Hye-su remains by the bench watching her go.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 구조물 사이로 멀리서 달리기 자세로 도약 중인 수리영과 벤치 앞에 서 있는 혜수의 모습이 은밀하게 담긴 누군가의 시점 쇼트.\n\nLOCATION (lock): From concealment among the dark exterior structures near the coastal bus stop, watching the bench and adjacent sidewalk from a distance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: A hard cut relocates the camera far from the bench and path at standing eye level, peering through narrow gaps between dark structures. 수리영 appears small in the distant right side during the next running bound, while 혜수 remains by the bench on the left within the same concealed wide frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, background, moves toward path away from the bench; 혜수 in the middle-left of the frame, background, looks toward 수리영 running away; bus-stop bench in the lower-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: structures (Partially occluding the view) — Their near sides face the camera while narrow gaps reveal the distant path and bench; used as Large foreground occluders whose gaps create the hidden observer's restricted view; bus-stop bench (Unoccupied, with 혜수 positioned in front of it) — Its side is visible beside 혜수 in the distant left portion of the frame; used as Distant anchor fixing 혜수's position relative to the departing 수리영.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The desaturated nighttime view remains low-key, with foreground structures held dark enough to make the concealed observation feel predatory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the dark seaside, bus-stop bench, pavement, and nighttime lighting from the reference. Exclude the prior seated version of the young woman; place the officer by the bench and the young woman farther away in a running stride, viewed through dark foreground structures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is still running away with her phone, while Hye-su remains by the bench watching her go.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained long-lens feel to compress the distant action into the narrow openings, with the dark structural edges sharply defined as rigid masks around the view. Keep the distant figures legible but small, arranging the open gaps so the eye travels laterally from the bench-side watcher to the suspended running bound. Sparse, low-key illumination and subdued contrast create the cold precision of covert surveillance rather than an expressive action image.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 구조물 사이로 멀리서 달리기 자세로 도약 중인 수리영과 벤치 앞에 서 있는 혜수의 모습이 은밀하게 담긴 누군가의 시점 쇼트.\n\nLOCATION (lock): From concealment among the dark exterior structures near the coastal bus stop, watching the bench and adjacent sidewalk from a distance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: A hard cut relocates the camera far from the bench and path at standing eye level, peering through narrow gaps between dark structures. 수리영 appears small in the distant right side during the next running bound, while 혜수 remains by the bench on the left within the same concealed wide frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, background, moves toward path away from the bench; 혜수 in the middle-left of the frame, background, looks toward 수리영 running away; bus-stop bench in the lower-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: structures (Partially occluding the view) — Their near sides face the camera while narrow gaps reveal the distant path and bench; used as Large foreground occluders whose gaps create the hidden observer's restricted view; bus-stop bench (Unoccupied, with 혜수 positioned in front of it) — Its side is visible beside 혜수 in the distant left portion of the frame; used as Distant anchor fixing 혜수's position relative to the departing 수리영.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The desaturated nighttime view remains low-key, with foreground structures held dark enough to make the concealed observation feel predatory.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the dark seaside, bus-stop bench, pavement, and nighttime lighting from the reference. Exclude the prior seated version of the young woman; place the officer by the bench and the young woman farther away in a running stride, viewed through dark foreground structures.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is still running away with her phone, while Hye-su remains by the bench watching her go.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a wider lens feel that makes the nearby structures loom as heavy, irregular fields of darkness while the revealed sidewalk and bench recede deeply beyond them. Let the foreground occluders fall slightly soft at their closest edges, directing attention through one restricted corridor toward the two small, separated figures; preserve clear spatial depth between concealment, bench, and departing runner. Shape the available night light into faint pools and edge glints, leaving most of the frame in textured shadow for a tense, predatory sense of hidden presence.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "전경의 어두운 구조물 틈새로, 벤치 앞의 혜수가 달아나는 수리영을 응시하는 시점 쇼트의 프레이밍과 시선 처리를 정확히 구현했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "도약 중인 수리영을 바라보아야 할 혜수가 카메라를 정면으로 응시하고 있어 시선 방향 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S15sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   },
   {
    "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:900092>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영이 휴대전화를 들고 있지 않습니다.",
     "fix_en": "Place a smartphone in Su-riyeong's hand."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reverse Su-riyeong’s running direction so she moves screen-right along the path, away from Hye-su and the bench, while preserving all other framing.",
     "issue_ko": "수리영이 벤치와 혜수가 있는 왼쪽을 향해 달리는 축으로 배치되어, 벤치에서 멀어져 달려가야 한다는 동선과 반대다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Place a smartphone in Su-riyeong's hand.\n- Reverse Su-riyeong’s running direction so she moves screen-right along the path, away from Hye-su and the bench, while preserving all other framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 프레이밍, 인물 배치 및 의상을 훌륭하게 구현했으며, 수리영이 휴대폰을 들고 뛰는 디테일(carried state)까지 정확히 반영했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "전반적인 구도와 배경, 인물의 배치는 매우 우수하나, 수리영의 손에 있어야 할 휴대폰이 누락되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 프레이밍, 인물 배치 및 의상을 훌륭하게 구현했으며, 수리영이 휴대폰을 들고 뛰는 디테일(carried state)까지 정확히 반영했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "전반적인 구도와 배경, 인물의 배치는 매우 우수하나, 수리영의 손에 있어야 할 휴대폰이 누락되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 은밀한 시점의 프레이밍을 정확히 구현했으며, 수리영이 휴대폰을 쥐고 달리는 이전 샷의 상태까지 충실하게 유지했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "구도와 인물의 배치 등은 요구사항에 부합하나, 지시문과 달리 수리영의 손에 들려 있어야 할 휴대폰이 누락되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 은밀한 시점의 프레이밍을 정확히 구현했으며, 수리영이 휴대폰을 쥐고 달리는 이전 샷의 상태까지 충실하게 유지했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "구도와 인물의 배치 등은 요구사항에 부합하나, 지시문과 달리 수리영의 손에 들려 있어야 할 휴대폰이 누락되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh4"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S16sh3::bgfirst_bg": {
  "input_fingerprint": "78f0240bf9c76b73",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 혜수의 손이 특정 번호가 띄워진 핸드폰 액정을 김형사의 얼굴 앞까지 불쑥 내민 찰나.\n\nLOCATION (lock): Inside the police station’s night-duty office, beside the detective’s workstation under the room’s working lights.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height just behind 혜수, the dolly-in follows her extended arm until the phone's front surface sits directly before 김형사's face. The screen occupies less than a third of the close frame, with 혜수's shoulder and forearm forming the near edge and 김형사's recoiling features providing scale behind it.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: mobile phone screen (On and displaying the selected contact number) — The illuminated front face is turned toward 김형사 and the camera, displaying 강민숙's saved number; used as Near-plane focal point positioned between 혜수 and 김형사.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Desaturated office ambience is held low-key while the active phone screen provides a restrained local emphasis justified by the visible display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 혜수의 손이 특정 번호가 띄워진 핸드폰 액정을 김형사의 얼굴 앞까지 불쑥 내민 찰나.\n\nLOCATION (lock): Inside the police station’s night-duty office, beside the detective’s workstation under the room’s working lights.\n\nTIME OF DAY (lock): dawn.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height just behind 혜수, the dolly-in follows her extended arm until the phone's front surface sits directly before 김형사's face. The screen occupies less than a third of the close frame, with 혜수's shoulder and forearm forming the near edge and 김형사's recoiling features providing scale behind it.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: mobile phone screen (On and displaying the selected contact number) — The illuminated front face is turned toward 김형사 and the camera, displaying 강민숙's saved number; used as Near-plane focal point positioned between 혜수 and 김형사.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Desaturated office ambience is held low-key while the active phone screen provides a restrained local emphasis justified by the visible display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S16sh3__bgfirst_bg.png",
  "asset_id": "59e8f761-3c1c-418b-a323-cf18a94026d6",
  "input_asset_ids": [
   "c8c2a5b8-96ea-4497-9d59-768a38bb7040",
   "73b5302b-b087-4abd-9e28-61a89334ac21"
  ]
 },
 "S16sh3": {
  "input_fingerprint": "db07bc3e3b8aa0dd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 혜수의 손이 특정 번호가 띄워진 핸드폰 액정을 김형사의 얼굴 앞까지 불쑥 내민 찰나.\n\nLOCATION (lock): Inside the police station’s night-duty office, beside the detective’s workstation under the room’s working lights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height just behind 혜수, the dolly-in follows her extended arm until the phone's front surface sits directly before 김형사's face. The screen occupies less than a third of the close frame, with 혜수's shoulder and forearm forming the near edge and 김형사's recoiling features providing scale behind it.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: mobile phone screen (On and displaying the selected contact number) — The illuminated front face is turned toward 김형사 and the camera, displaying 강민숙's saved number; used as Near-plane focal point positioned between 혜수 and 김형사.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Desaturated office ambience is held low-key while the active phone screen provides a restrained local emphasis justified by the visible display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Hye-su holds the phone displaying Minsuk’s saved number after repeated calls have gone unanswered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락); 김형사 (한국인 30대 남성 얼굴, 성인 남성의 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 혜수의 손이 특정 번호가 띄워진 핸드폰 액정을 김형사의 얼굴 앞까지 불쑥 내민 찰나.\n\nLOCATION (lock): Inside the police station’s night-duty office, beside the detective’s workstation under the room’s working lights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height just behind 혜수, the dolly-in follows her extended arm until the phone's front surface sits directly before 김형사's face. The screen occupies less than a third of the close frame, with 혜수's shoulder and forearm forming the near edge and 김형사's recoiling features providing scale behind it.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: mobile phone screen (On and displaying the selected contact number) — The illuminated front face is turned toward 김형사 and the camera, displaying 강민숙's saved number; used as Near-plane focal point positioned between 혜수 and 김형사.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Desaturated office ambience is held low-key while the active phone screen provides a restrained local emphasis justified by the visible display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Hye-su holds the phone displaying Minsuk’s saved number after repeated calls have gone unanswered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락); 김형사 (한국인 30대 남성 얼굴, 성인 남성의 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 혜수의 손이 특정 번호가 띄워진 핸드폰 액정을 김형사의 얼굴 앞까지 불쑥 내민 찰나.\n\nLOCATION (lock): Inside the police station’s night-duty office, beside the detective’s workstation under the room’s working lights. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height just behind 혜수, the dolly-in follows her extended arm until the phone's front surface sits directly before 김형사's face. The screen occupies less than a third of the close frame, with 혜수's shoulder and forearm forming the near edge and 김형사's recoiling features providing scale behind it.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: mobile phone screen (On and displaying the selected contact number) — The illuminated front face is turned toward 김형사 and the camera, displaying 강민숙's saved number; used as Near-plane focal point positioned between 혜수 and 김형사.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Desaturated office ambience is held low-key while the active phone screen provides a restrained local emphasis justified by the visible display.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Hye-su holds the phone displaying Minsuk’s saved number after repeated calls have gone unanswered.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락); 김형사 (한국인 30대 남성 얼굴, 성인 남성의 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S16sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S16sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L13B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요청된 클로즈업 프레이밍을 정확히 구현하여 핸드폰이 김형사의 얼굴 바로 앞에 위치하며, 화면의 '강민숙' 텍스트도 명확히 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 클로즈업 대신 미디엄 샷으로 너무 넓게 촬영되었으며, 핸드폰과 얼굴 사이의 거리가 멀고 화면 텍스트도 부정확합니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요청된 클로즈업 프레이밍을 정확히 구현하여 핸드폰이 김형사의 얼굴 바로 앞에 위치하며, 화면의 '강민숙' 텍스트도 명확히 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 클로즈업 대신 미디엄 샷으로 너무 넓게 촬영되었으며, 핸드폰과 얼굴 사이의 거리가 멀고 화면 텍스트도 부정확합니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요청된 클로즈업 앵글과 폰 화면을 김형사 얼굴 바로 앞에 내미는 역동적인 구도를 잘 구현했으며, 화면 텍스트도 비교적 정확합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "프레임이 너무 넓어 요청된 클로즈업 스케일과 역동적인 긴장감을 살리지 못했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요청된 클로즈업 앵글과 폰 화면을 김형사 얼굴 바로 앞에 내미는 역동적인 구도를 잘 구현했으며, 화면 텍스트도 비교적 정확합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "프레임이 너무 넓어 요청된 클로즈업 스케일과 역동적인 긴장감을 살리지 못했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 8,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "요청된 클로즈업 프레이밍을 정확히 구현하여 핸드폰이 김형사의 얼굴 바로 앞에 위치하며, 화면의 '강민숙' 텍스트도 명확히 표현되었습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "지정된 클로즈업 대신 미디엄 샷으로 너무 넓게 촬영되었으며, 핸드폰과 얼굴 사이의 거리가 멀고 화면 텍스트도 부정확합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L13B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:900092>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "핸드폰을 쥐고 있는 혜수의 오른손 손가락 구조와 관절이 인체 해부학적으로 심하게 왜곡되어 있습니다.",
     "fix_en": "Redraw Hye-su's right hand holding the phone to feature anatomically correct fingers and a natural grip."
    },
    {
     "issue_ko": "핸드폰 화면에 프롬프트가 지시한 '강민숙' 대신 '강만숙'이라고 잘못 표기되어 있습니다.",
     "fix_en": "Correct the name displayed on the mobile phone screen to read exactly \"강민숙\"."
    },
    {
     "issue_ko": "우측 CCTV 모니터 상단에 프롬프트에서 금지한 임의의 텍스트('CCTV')가 크게 삽입되어 있습니다.",
     "fix_en": "Remove the prominent \"CCTV\" text overlay from the top left corner of the computer monitor."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Tighten to a close-up around Hye-su’s near shoulder and forearm, the phone, and Kim’s recoiling face, excluding most of his torso and the workstation.",
     "issue_ko": "프레임에 김형사의 상반신과 넓은 워크스테이션이 과도하게 포함되어 지정된 클로즈업보다 훨씬 넓다."
    },
    {
     "fix_en": "Reposition the phone directly along the camera-to-Kim axis so it overlaps the central area of his face while remaining under one-third of the frame.",
     "issue_ko": "휴대폰이 김형사의 얼굴 바로 앞에 겹쳐지지 않고 얼굴 오른쪽 볼 옆으로 치우쳐 있어 요구된 전후 배치가 어긋난다."
    },
    {
     "fix_en": "Redirect Kim’s gaze to the phone display while preserving his recoiling posture.",
     "issue_ko": "김형사의 시선이 눈앞의 휴대폰 화면이 아니라 화면 왼쪽의 혜수를 향하고 있다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw Hye-su's right hand holding the phone to feature anatomically correct fingers and a natural grip.\n- Correct the name displayed on the mobile phone screen to read exactly \"강민숙\".\n- Remove the prominent \"CCTV\" text overlay from the top left corner of the computer monitor.\n- Tighten to a close-up around Hye-su’s near shoulder and forearm, the phone, and Kim’s recoiling face, excluding most of his torso and the workstation.\n- Reposition the phone directly along the camera-to-Kim axis so it overlaps the central area of his face while remaining under one-third of the frame.\n- Redirect Kim’s gaze to the phone display while preserving his recoiling posture.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 클로즈업 구도와 숄더뷰 앵글을 정확히 구현했으며, 인물의 특징과 지정된 소품(화면이 켜진 핸드폰)이 지시사항에 맞게 잘 표현되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "김형사의 머리 뒤에 불필요한 얼굴이 중복 생성되는 치명적인 신체 구조 오류(Hard Violation)가 있어 실격입니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 클로즈업 구도와 숄더뷰 앵글을 정확히 구현했으며, 인물의 특징과 지정된 소품(화면이 켜진 핸드폰)이 지시사항에 맞게 잘 표현되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "김형사의 머리 뒤에 불필요한 얼굴이 중복 생성되는 치명적인 신체 구조 오류(Hard Violation)가 있어 실격입니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "치명적인 오류(형사의 얼굴 뒤에 기형적으로 중복된 머리 형상)가 존재하여 실격입니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "제시된 클로즈업 앵글, 핸드폰 화면의 텍스트, 인물들의 위치와 표정, 그리고 지정된 배경까지 매우 충실하게 구현했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "치명적인 오류(형사의 얼굴 뒤에 기형적으로 중복된 머리 형상)가 존재하여 실격입니다."
     },
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "제시된 클로즈업 앵글, 핸드폰 화면의 텍스트, 인물들의 위치와 표정, 그리고 지정된 배경까지 매우 충실하게 구현했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 16,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S16sh3__bgfirst_bg.png",
   "bg_asset_id": "59e8f761-3c1c-418b-a323-cf18a94026d6",
   "bg_record_key": "S16sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L13B01",
    "B": "L13B02"
   },
   "assigned": "L13B02",
   "choice": "Candidate B",
   "confident": true,
   "reason_ko": "샷 텍스트에서 핸드폰 액정을 내미는 상황이 묘사될 뿐 특정 사무실 구역을 강하게 한정짓지 않으며, 두 후보 이미지가 동일한 공간을 보여주므로 현재 할당된 후보를 유지합니다.",
   "kept": "L13B02"
  }
 },
 "S16sh5::variants": {
  "author_fp": "443dc873fa219e66",
  "author": {
   "variants": [
    {
     "approach_ko": "얕은 심도와 압축된 원근으로 굳은 표정만 고립시키는 밀도 높은 심리적 클로즈업.",
     "prompt_en": "Use a compressed, intimate lens feel with extremely shallow depth of field, isolating the facial planes while the office dissolves into soft, abstract edge shapes. Keep the eyes critically sharp and let focus fall away across the nearer cheek and hair, with restrained low-key modeling that carves the tension between the brows and leaves the frame claustrophobic and still."
    },
    {
     "approach_ko": "가까운 광각감과 조금 더 살아 있는 공간 원근으로 불안을 즉각적으로 마주하게 하는 클로즈업.",
     "prompt_en": "Use a close, subtly wider-perspective lens feel so the face has immediate physical presence and the surrounding office edges retain faint spatial depth without becoming legible. Hold a firmer depth of field across both eyes, brow, and mouth; shape the desaturated low-key ambience with gentle tonal falloff across the face, emphasizing rigid stillness and unresolved unease rather than pictorial softness."
    }
   ]
  },
  "reused": false
 },
 "S16sh5": {
  "input_fingerprint": "3aee32b0b111fdc1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 미간을 깊게 구긴 채 허공을 응시하는 혜수의 불안하고 굳은 안면 클로즈업.\n\nLOCATION (lock): Inside the police station’s night-duty office, near the desks where only one detective is on shift. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: On the reverse angle at 혜수's eye level, the static camera holds a tight three-quarter facial close-up rather than following her drifting gaze. Her furrowed brow and fixed eyes occupy most of the centered frame, with the office context reduced to indistinct edge shapes.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 혜수 in the middle-center of the frame, foreground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated office ambience concentrates tonal contrast around 혜수's furrowed brow and rigid gaze without specifying an unestablished source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Hye-su retains the phone with Minsuk’s number as the continued lack of response leaves her visibly uneasy.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 미간을 깊게 구긴 채 허공을 응시하는 혜수의 불안하고 굳은 안면 클로즈업.\n\nLOCATION (lock): Inside the police station’s night-duty office, near the desks where only one detective is on shift. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: On the reverse angle at 혜수's eye level, the static camera holds a tight three-quarter facial close-up rather than following her drifting gaze. Her furrowed brow and fixed eyes occupy most of the centered frame, with the office context reduced to indistinct edge shapes.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 혜수 in the middle-center of the frame, foreground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated office ambience concentrates tonal contrast around 혜수's furrowed brow and rigid gaze without specifying an unestablished source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Hye-su retains the phone with Minsuk’s number as the continued lack of response leaves her visibly uneasy.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a compressed, intimate lens feel with extremely shallow depth of field, isolating the facial planes while the office dissolves into soft, abstract edge shapes. Keep the eyes critically sharp and let focus fall away across the nearer cheek and hair, with restrained low-key modeling that carves the tension between the brows and leaves the frame claustrophobic and still.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 미간을 깊게 구긴 채 허공을 응시하는 혜수의 불안하고 굳은 안면 클로즈업.\n\nLOCATION (lock): Inside the police station’s night-duty office, near the desks where only one detective is on shift. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: On the reverse angle at 혜수's eye level, the static camera holds a tight three-quarter facial close-up rather than following her drifting gaze. Her furrowed brow and fixed eyes occupy most of the centered frame, with the office context reduced to indistinct edge shapes.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 혜수 in the middle-center of the frame, foreground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated office ambience concentrates tonal contrast around 혜수's furrowed brow and rigid gaze without specifying an unestablished source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Hye-su retains the phone with Minsuk’s number as the continued lack of response leaves her visibly uneasy.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a close, subtly wider-perspective lens feel so the face has immediate physical presence and the surrounding office edges retain faint spatial depth without becoming legible. Hold a firmer depth of field across both eyes, brow, and mouth; shape the desaturated low-key ambience with gentle tonal falloff across the face, emphasizing rigid stillness and unresolved unease rather than pictorial softness.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지정된 3/4 측면 각도 대신 정면 클로즈업으로 연출되었으나, 인물의 표정, 휴대폰 파지 상태, 흐릿한 배경 등 주요 지시사항을 사실적으로 잘 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "화면 우측에 해부학적으로 주인공의 것일 수 없는 손이 등장하며(치명적 위반), 이전 컷에서 서 있던 형사가 배경에 앉아 있어 샷 연속성에도 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S16sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:900092>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라 앵글이 3/4 측면(three-quarter)이 아닌 정면을 향하고 있으며, 허공을 응시해야 할 시선이 렌즈를 똑바로 바라보고 있습니다.",
     "fix_en": "Change the camera angle and subject's face to a three-quarter profile, and direct her gaze off-camera into space."
    },
    {
     "issue_ko": "캐릭터가 들고 있는 스마트폰의 화면이 본인이 아닌 카메라를 향하고 있어 사물 방향 규칙을 위반했으며, 화면의 텍스트가 좌우 반전되어 노출되었습니다.",
     "fix_en": "Rotate the phone so its screen faces the character's eyes and its back faces the camera, removing the mirrored text."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Shift the camera to Hye-su’s eye-level three-quarter reverse angle while preserving the same tight centered close-up scale.",
     "issue_ko": "카메라가 지정된 역방향의 3/4 각도가 아니라 혜수를 정면으로 촬영하고 있다."
    },
    {
     "fix_en": "Redirect Hye-su’s fixed gaze slightly off the camera axis into empty space without changing her head position or framing.",
     "issue_ko": "혜수의 시선이 허공으로 빗나가 고정되기보다 카메라 렌즈를 정면으로 응시하는 축에 가깝다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the camera angle and subject's face to a three-quarter profile, and direct her gaze off-camera into space.\n- Rotate the phone so its screen faces the character's eyes and its back faces the camera, removing the mirrored text.\n- Shift the camera to Hye-su’s eye-level three-quarter reverse angle while preserving the same tight centered close-up scale.\n- Redirect Hye-su’s fixed gaze slightly off the camera axis into empty space without changing her head position or framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "미간을 찌푸린 채 허공을 응시하는 3/4 측면 안면 클로즈업을 완벽히 구현했으며, 휴대폰의 후면을 노출해 인물 시선에 맞춘 소품 방향 규칙도 정확히 준수했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "정면 앵글과 렌즈를 향한 시선으로 인해 '허공을 응시하는 3/4 측면 클로즈업' 지시를 위반했으며, 휴대폰 화면이 카메라를 향해 있어 소품 방향 규칙에도 어긋납니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "미간을 찌푸린 채 허공을 응시하는 3/4 측면 안면 클로즈업을 완벽히 구현했으며, 휴대폰의 후면을 노출해 인물 시선에 맞춘 소품 방향 규칙도 정확히 준수했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "정면 앵글과 렌즈를 향한 시선으로 인해 '허공을 응시하는 3/4 측면 클로즈업' 지시를 위반했으며, 휴대폰 화면이 카메라를 향해 있어 소품 방향 규칙에도 어긋납니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지시된 3/4 측면 안면 클로즈업과 허공을 응시하는 시선을 정확히 연출했으며, 휴대폰의 화면을 캐릭터 쪽으로 향하게 쥔 현실적인 방향도 규칙에 잘 부합합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "3/4 측면이 아닌 정면 클로즈업으로 촬영되었으며, 캐릭터가 조작 중인 휴대폰의 화면(기능면)이 카메라를 향해 있어 소품 방향 규칙을 정면으로 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "지시된 3/4 측면 안면 클로즈업과 허공을 응시하는 시선을 정확히 연출했으며, 휴대폰의 화면을 캐릭터 쪽으로 향하게 쥔 현실적인 방향도 규칙에 잘 부합합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "3/4 측면이 아닌 정면 클로즈업으로 촬영되었으며, 캐릭터가 조작 중인 휴대폰의 화면(기능면)이 카메라를 향해 있어 소품 방향 규칙을 정면으로 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 19
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S16sh3"
  }
 },
 "S17sh1::variants": {
  "author_fp": "e424cd0687be0865",
  "author": {
   "variants": [
    {
     "approach_ko": "깊은 초점과 엄격한 수평 구도로 고양이와 창문 사이의 긴장된 거리를 또렷하게 유지한다.",
     "prompt_en": "Photograph this as a restrained deep-focus architectural tableau, with the wall’s lateral geometry held rigid and level and both the perched figure and distant window clearly legible. Use a natural, moderately wide lens feel with minimal distortion, generous surrounding space, crisp weathered surface detail, and subdued dawn contrast; allow the isolated interior light to remain controlled rather than blown out. The image should feel watchful, static, and quietly ominous."
    },
    {
     "approach_ko": "압축된 원근과 선택적 초점으로 고양이의 경계심을 강조하고 먼 창문은 부드러운 빛의 표적으로 남긴다.",
     "prompt_en": "Use a compressed, selective-focus treatment that visually tightens the separation across the wide frame while preserving the assigned scale and placements. Hold the cat in precise profile focus, let the distant window remain recognizable but gently softened with a restrained luminous bloom, and allow the intervening wall texture to fall away gradually through depth. Shape the blue dawn ambience with low-key directional modeling and dense shadows, creating a suspended, predatory stillness before movement."
    }
   ]
  },
  "reused": false
 },
 "S17sh1": {
  "input_fingerprint": "4f9007da29ad61db",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 푸르스름한 새벽빛 아래, 담벼락 위에 앉은 검은 고양이가 불 켜진 옥탑방 창문을 노려보는 정지 구도.\n\nLOCATION (lock): On the rooftop boundary wall outside the aging villa’s rooftop room, facing its illuminated window in blue dawn light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static wide composition from wall-top height several meters away, placing the black cat in profile on the left third and the lit rooftop-room window deeper in the right background along its eyeline. The wall creates a horizontal visual path between animal and window, with the cat held in tense stillness before the anticipated lateral move.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 검은 고양이 in the middle-left of the frame, midground, looks toward lit rooftop-room window; lit rooftop-room window in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 담벼락 (Present with the cat seated on top) — Its top edge runs laterally across the frame, with the cat perched on the visible upper side; used as Supports the cat and forms a horizontal compositional line toward the rooftop room; 옥탑방 창문 (Lit) — The camera sees the illuminated rooftop-room interior side through the window from a distance; used as Distant visual target at the end of the cat's eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bluish dawn ambient light is held in a dark, desaturated palette against the isolated illumination of the rooftop-room window.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rooftop room’s light remains on through dawn while the apartment is still without any sign of Minsuk.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 푸르스름한 새벽빛 아래, 담벼락 위에 앉은 검은 고양이가 불 켜진 옥탑방 창문을 노려보는 정지 구도.\n\nLOCATION (lock): On the rooftop boundary wall outside the aging villa’s rooftop room, facing its illuminated window in blue dawn light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static wide composition from wall-top height several meters away, placing the black cat in profile on the left third and the lit rooftop-room window deeper in the right background along its eyeline. The wall creates a horizontal visual path between animal and window, with the cat held in tense stillness before the anticipated lateral move.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 검은 고양이 in the middle-left of the frame, midground, looks toward lit rooftop-room window; lit rooftop-room window in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 담벼락 (Present with the cat seated on top) — Its top edge runs laterally across the frame, with the cat perched on the visible upper side; used as Supports the cat and forms a horizontal compositional line toward the rooftop room; 옥탑방 창문 (Lit) — The camera sees the illuminated rooftop-room interior side through the window from a distance; used as Distant visual target at the end of the cat's eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bluish dawn ambient light is held in a dark, desaturated palette against the isolated illumination of the rooftop-room window.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rooftop room’s light remains on through dawn while the apartment is still without any sign of Minsuk.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph this as a restrained deep-focus architectural tableau, with the wall’s lateral geometry held rigid and level and both the perched figure and distant window clearly legible. Use a natural, moderately wide lens feel with minimal distortion, generous surrounding space, crisp weathered surface detail, and subdued dawn contrast; allow the isolated interior light to remain controlled rather than blown out. The image should feel watchful, static, and quietly ominous.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dawn.\n\nSHOT TEXT (authoritative, Korean): 푸르스름한 새벽빛 아래, 담벼락 위에 앉은 검은 고양이가 불 켜진 옥탑방 창문을 노려보는 정지 구도.\n\nLOCATION (lock): On the rooftop boundary wall outside the aging villa’s rooftop room, facing its illuminated window in blue dawn light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold a static wide composition from wall-top height several meters away, placing the black cat in profile on the left third and the lit rooftop-room window deeper in the right background along its eyeline. The wall creates a horizontal visual path between animal and window, with the cat held in tense stillness before the anticipated lateral move.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 검은 고양이 in the middle-left of the frame, midground, looks toward lit rooftop-room window; lit rooftop-room window in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 담벼락 (Present with the cat seated on top) — Its top edge runs laterally across the frame, with the cat perched on the visible upper side; used as Supports the cat and forms a horizontal compositional line toward the rooftop room; 옥탑방 창문 (Lit) — The camera sees the illuminated rooftop-room interior side through the window from a distance; used as Distant visual target at the end of the cat's eyeline.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bluish dawn ambient light is held in a dark, desaturated palette against the isolated illumination of the rooftop-room window.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The rooftop room’s light remains on through dawn while the apartment is still without any sign of Minsuk.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a compressed, selective-focus treatment that visually tightens the separation across the wide frame while preserving the assigned scale and placements. Hold the cat in precise profile focus, let the distant window remain recognizable but gently softened with a restrained luminous bloom, and allow the intervening wall texture to fall away gradually through depth. Shape the blue dawn ambience with low-key directional modeling and dense shadows, creating a suspended, predatory stillness before movement.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 8,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "고양이를 화면 중앙 좌측 중경에 배치하라는 구도 지시를 정확히 따랐으며, 새벽의 푸른 조명과 참조 이미지의 디테일을 충실히 재현했습니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "고양이가 지시된 중경이 아닌 전경에 지나치게 가깝고 크게 배치되어 구도 조건을 어겼으며, 참조 이미지의 일부 디테일(빨랫줄 등)이 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L04B02.png"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/background_chain/seed_bg_rooftop_villa_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "옥탑방의 문과 창문이 STRUCTURE LOOK 레퍼런스(깨끗한 회색 문, 창살 없는 미닫이창)가 아닌 위치 사진(녹슨 문, 방범창)의 형태와 재질을 따르고 있습니다.",
     "fix_en": "Change the rooftop room's door to a clean, rust-free grey surface and replace the barred window with an unbarred sliding window to match the STRUCTURE LOOK reference."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reframe the cat smaller and farther from the camera, placing it in the middle-left midground while preserving its right-facing profile and eyeline to the window.",
     "issue_ko": "고양이가 중좌측 중경이 아니라 좌하단 전경에 지나치게 크게 배치되어 있다."
    },
    {
     "fix_en": "Shift the viewpoint so the cat-supporting wall top runs level and laterally across the frame toward the distant window, forming the specified horizontal visual path.",
     "issue_ko": "고양이를 받치는 담벼락 상단이 창문 쪽으로 이어지는 수평 구도를 만들지 못하고 전경에서 꺾이고 사선으로 후퇴한다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the rooftop room's door to a clean, rust-free grey surface and replace the barred window with an unbarred sliding window to match the STRUCTURE LOOK reference.\n- Reframe the cat smaller and farther from the camera, placing it in the middle-left midground while preserving its right-facing profile and eyeline to the window.\n- Shift the viewpoint so the cat-supporting wall top runs level and laterally across the frame toward the distant window, forming the specified horizontal visual path.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지침에 따라 창문 내부에 인물을 배제하여 프롬프트를 잘 충족했으나, 구조물 기준 사진을 우선시하지 않고 위치 기준 사진의 방범창을 묘사한 점이 약간 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "창문 내부에 사람들을 생성하여 프롬프트의 '인물 배제' 및 '민석의 흔적이 없는' 지침을 정면으로 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지침에 따라 창문 내부에 인물을 배제하여 프롬프트를 잘 충족했으나, 구조물 기준 사진을 우선시하지 않고 위치 기준 사진의 방범창을 묘사한 점이 약간 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "창문 내부에 사람들을 생성하여 프롬프트의 '인물 배제' 및 '민석의 흔적이 없는' 지침을 정면으로 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "창문 안에 인물이 뚜렷하게 보여 '인물 없음(NO PEOPLE)' 지시사항을 위반했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "인물 없는 빈 방, 올바른 조명, 고양이의 위치와 구도 등 모든 지침을 충실히 반영했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "창문 안에 인물이 뚜렷하게 보여 '인물 없음(NO PEOPLE)' 지시사항을 위반했습니다."
     },
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "인물 없는 빈 방, 올바른 조명, 고양이의 위치와 구도 등 모든 지침을 충실히 반영했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "플레이트+seed만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only",
  "plate_select": {
   "candidates": {
    "A": "L04B01",
    "B": "L04B02",
    "C": "L04B03"
   },
   "assigned": "L04B01",
   "choice": "Candidate A",
   "confident": true,
   "reason_ko": "후보들은 모두 옥탑방 앞 옥상과 담벼락이라는 동일한 서브공간을 보여줍니다. 조명이나 분위기(새벽빛, 불 켜진 창문)가 아닌 공간 정체성으로 판단해야 하므로, 서브공간의 변경을 요구하는 명확한 근거가 없어 현재 할당된 후보 A를 유지합니다.",
   "kept": "L04B01"
  }
 },
 "S18sh2::bgfirst_bg": {
  "input_fingerprint": "252a0f0f94878fe9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 두 눈을 번쩍 뜬 채 고개를 빳빳하게 세운 수리영의 경직된 안면 클로즈업.\n\nLOCATION (lock): Beside the bed in the rooftop apartment’s main bedroom, in the early daylight after a night spent waiting.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At mattress height, complete the forward move into a locked close-up as 수리영 snaps awake, her rigid face occupying most of the frame in a near-frontal angle. Keep only a narrow strip of mattress and room behind her, emphasizing the change in her gaze as it searches beyond the lens rather than altering the established axis further.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 침대 매트리스 가장자리 (수리영 is seated with her head lifted from it); used as A narrow lower-frame reference establishing that she has awakened where she dozed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination keeps her widened eyes and rigid facial tension readable without specifying an ungrounded source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 두 눈을 번쩍 뜬 채 고개를 빳빳하게 세운 수리영의 경직된 안면 클로즈업.\n\nLOCATION (lock): Beside the bed in the rooftop apartment’s main bedroom, in the early daylight after a night spent waiting.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At mattress height, complete the forward move into a locked close-up as 수리영 snaps awake, her rigid face occupying most of the frame in a near-frontal angle. Keep only a narrow strip of mattress and room behind her, emphasizing the change in her gaze as it searches beyond the lens rather than altering the established axis further.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 침대 매트리스 가장자리 (수리영 is seated with her head lifted from it); used as A narrow lower-frame reference establishing that she has awakened where she dozed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination keeps her widened eyes and rigid facial tension readable without specifying an ungrounded source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S18sh2__bgfirst_bg.png",
  "asset_id": "74220265-8ba6-40a2-b5d0-6d0318745296",
  "input_asset_ids": [
   "1c9180ef-c789-4b11-8aba-69882f9d2429",
   "1099d4f5-d043-4238-a20d-1b898b0e8997"
  ]
 },
 "S18sh2": {
  "input_fingerprint": "abdef981778c9d78",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 번쩍 뜬 채 고개를 빳빳하게 세운 수리영의 경직된 안면 클로즈업.\n\nLOCATION (lock): Beside the bed in the rooftop apartment’s main bedroom, in the early daylight after a night spent waiting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At mattress height, complete the forward move into a locked close-up as 수리영 snaps awake, her rigid face occupying most of the frame in a near-frontal angle. Keep only a narrow strip of mattress and room behind her, emphasizing the change in her gaze as it searches beyond the lens rather than altering the established axis further.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 침대 매트리스 가장자리 (수리영 is seated with her head lifted from it); used as A narrow lower-frame reference establishing that she has awakened where she dozed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination keeps her widened eyes and rigid facial tension readable without specifying an ungrounded source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains empty of Minsuk, and Su-riyeong’s room remains excessively clean with no corpse or blood behind the curtain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 번쩍 뜬 채 고개를 빳빳하게 세운 수리영의 경직된 안면 클로즈업.\n\nLOCATION (lock): Beside the bed in the rooftop apartment’s main bedroom, in the early daylight after a night spent waiting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At mattress height, complete the forward move into a locked close-up as 수리영 snaps awake, her rigid face occupying most of the frame in a near-frontal angle. Keep only a narrow strip of mattress and room behind her, emphasizing the change in her gaze as it searches beyond the lens rather than altering the established axis further.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 침대 매트리스 가장자리 (수리영 is seated with her head lifted from it); used as A narrow lower-frame reference establishing that she has awakened where she dozed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination keeps her widened eyes and rigid facial tension readable without specifying an ungrounded source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains empty of Minsuk, and Su-riyeong’s room remains excessively clean with no corpse or blood behind the curtain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 번쩍 뜬 채 고개를 빳빳하게 세운 수리영의 경직된 안면 클로즈업.\n\nLOCATION (lock): Beside the bed in the rooftop apartment’s main bedroom, in the early daylight after a night spent waiting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At mattress height, complete the forward move into a locked close-up as 수리영 snaps awake, her rigid face occupying most of the frame in a near-frontal angle. Keep only a narrow strip of mattress and room behind her, emphasizing the change in her gaze as it searches beyond the lens rather than altering the established axis further.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 침대 매트리스 가장자리 (수리영 is seated with her head lifted from it); used as A narrow lower-frame reference establishing that she has awakened where she dozed.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated illumination keeps her widened eyes and rigid facial tension readable without specifying an ungrounded source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains empty of Minsuk, and Su-riyeong’s room remains excessively clean with no corpse or blood behind the curtain.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S18sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S18sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B06.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 클로즈업 구도와 안면 연기를 잘 소화했으며, 배경에 레퍼런스 이미지의 방 구조와 가구(문, 탁자, 침대)를 정확히 구현해 로케이션 조건을 훌륭히 충족합니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "요구된 표정과 조명 분위기는 잘 살렸으나, 배경이 지나치게 어둡고 흐려 필수 로케이션 레퍼런스의 세부 요소들을 확인하기 어렵습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 클로즈업 구도와 안면 연기를 잘 소화했으며, 배경에 레퍼런스 이미지의 방 구조와 가구(문, 탁자, 침대)를 정확히 구현해 로케이션 조건을 훌륭히 충족합니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "요구된 표정과 조명 분위기는 잘 살렸으나, 배경이 지나치게 어둡고 흐려 필수 로케이션 레퍼런스의 세부 요소들을 확인하기 어렵습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 안면 클로즈업 샷 크기를 완벽히 구현하여 얼굴이 프레임의 대부분을 차지하도록 연출한 점이 매우 우수합니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "배경과 인물은 잘 묘사되었으나, 프레임이 너무 넓어져 얼굴이 프레임을 가득 채우지 못하고 배경이 과도하게 노출되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 안면 클로즈업 샷 크기를 완벽히 구현하여 얼굴이 프레임의 대부분을 차지하도록 연출한 점이 매우 우수합니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "배경과 인물은 잘 묘사되었으나, 프레임이 너무 넓어져 얼굴이 프레임을 가득 채우지 못하고 배경이 과도하게 노출되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 12
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "totals": {
   "A": 12,
   "B": 12
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지정된 클로즈업 구도와 안면 연기를 잘 소화했으며, 배경에 레퍼런스 이미지의 방 구조와 가구(문, 탁자, 침대)를 정확히 구현해 로케이션 조건을 훌륭히 충족합니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "요구된 표정과 조명 분위기는 잘 살렸으나, 배경이 지나치게 어둡고 흐려 필수 로케이션 레퍼런스의 세부 요소들을 확인하기 어렵습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B06.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "시선이 렌즈 정면을 향하고 있습니다. 레이아웃 스케치의 화살표 방향대로 화면 오른쪽을 향해야 합니다.",
     "fix_en": "Redirect her gaze to look toward the right side of the frame, strictly following the arrow direction in the layout sketch."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Tighten the close-up so her rigid near-frontal face occupies most of the frame, leaving only a narrow strip of mattress along the bottom and minimal room visible behind her.",
     "issue_ko": "현재 인물이 너무 작고 하단 매트리스와 주변 방이 넓게 보여, 얼굴이 프레임 대부분을 차지하고 매트리스는 좁은 하단 띠로만 보여야 하는 지정 클로즈업보다 구도가 느슨하다."
    },
    {
     "fix_en": "Redirect her gaze toward screen right beyond the lens while keeping her head rigid and near-frontal.",
     "issue_ko": "수리영의 시선이 렌즈를 정면으로 향하지만, 레이아웃 화살표대로 렌즈 너머 화면 오른쪽을 향해야 한다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redirect her gaze to look toward the right side of the frame, strictly following the arrow direction in the layout sketch.\n- Tighten the close-up so her rigid near-frontal face occupies most of the frame, leaving only a narrow strip of mattress along the bottom and minimal room visible behind her.\n- Redirect her gaze toward screen right beyond the lens while keeping her head rigid and near-frontal.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 원본 배경을 완벽하게 유지하고 빳빳하게 굳은 표정과 구도를 잘 살렸으나, 스케치가 지시한 우측 시선 방향을 정확히 반영하지 못했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "스케치에 표시된 시선 방향은 따랐으나, 완전히 고정되어야 할 배경(열린 문과 복도, 오른쪽 침대 헤드보드 등)을 완전히 무시하고 임의로 변경하여 우선순위에서 감점되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 원본 배경을 완벽하게 유지하고 빳빳하게 굳은 표정과 구도를 잘 살렸으나, 스케치가 지시한 우측 시선 방향을 정확히 반영하지 못했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "스케치에 표시된 시선 방향은 따랐으나, 완전히 고정되어야 할 배경(열린 문과 복도, 오른쪽 침대 헤드보드 등)을 완전히 무시하고 임의로 변경하여 우선순위에서 감점되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "레이아웃 스케치가 지시한 타이트한 안면 클로즈업 구도와 오른쪽을 향하는 시선 방향을 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "스케치의 시선 방향(오른쪽)을 무시하고 렌즈를 응시했으며, 프레임이 너무 넓어 클로즈업 요구사항을 충족하지 못했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "레이아웃 스케치가 지시한 타이트한 안면 클로즈업 구도와 오른쪽을 향하는 시선 방향을 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "스케치의 시선 방향(오른쪽)을 무시하고 렌즈를 응시했으며, 프레임이 너무 넓어 클로즈업 요구사항을 충족하지 못했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 11,
     "B": 11
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S18sh2__bgfirst_bg.png",
   "bg_asset_id": "74220265-8ba6-40a2-b5d0-6d0318745296",
   "bg_record_key": "S18sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B06",
   "choice": "Candidate F",
   "confident": false,
   "reason_ko": "\"두 눈을 번쩍 뜬 채\"라는 묘사 외에 구체적인 공간을 지시하는 단서가 없어, 침대에서 깨어나는 상황으로 해석 가능한 현재 할당된 후보를 유지함.",
   "kept": "L06B06"
  }
 },
 "S18sh12::variants": {
  "author_fp": "3e1d45ff59978d3b",
  "author": {
   "variants": [
    {
     "approach_ko": "자연스러운 원근감과 섬세한 측광으로 반사된 얼굴과 붉은 원 사이의 심리적 긴장을 강조한다.",
     "prompt_en": "Use a restrained, natural-perspective lens feel, making the reflected face and the crimson ring equal points of tension across the frame. Shape the dim early-morning bathroom light into soft directional modeling on skin and grazing definition across the tiles, with gentle falloff beyond the two focal anchors. Keep the visible mirror edge as a narrow, dark physical boundary, and render the rough stroke seated in the tile texture and broken naturally by grout."
    },
    {
     "approach_ko": "완만한 압축 원근과 평평한 저조도 조명으로 거울면과 타일 격자의 기하학적 불안을 부각한다.",
     "prompt_en": "Use a gently compressed lens feel to flatten the reflected depth into an uneasy arrangement of face, mirror plane, tiled grid, and ring. Favor broad, cool, low-key illumination with minimal facial modeling, letting the desaturated geometry dominate while the restrained crimson remains the sole sharp color accent. Preserve shallow focus but tune the focal plane so both the reflected face and the rough circular stroke read cleanly, with the rest of the bathroom receding softly."
    }
   ]
  },
  "reused": false
 },
 "S18sh12": {
  "input_fingerprint": "291b61a978b21336",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 화장실 거울 반사면 안, 수리영의 등 뒤 타일 벽면에 거칠고 선명한 붉은 원이 그려져 있는 시점 컷.\n\nLOCATION (lock): Inside the rooftop apartment’s bathroom, viewed through the mirror toward the tiled wall behind the sink. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the dolly-in at eye level just beside the mirror axis, framing 수리영's left-right-reversed reflection on the left half and the sharply readable red circle on the reflected wall behind her to the right. The mirror frame remains visible at an edge so the image cannot be mistaken for a direct view, while shallow focus still holds both her reflected face and the circle legible.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영's mirror reflection in the middle-left of the frame, midground; red circle reflected on bathroom wall in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 욕실 거울 (The red circle is visible only within the reflection at this instant) — Its camera-facing surface reflects a left-right-reversed 수리영 and the bathroom wall behind her, with part of the mirror frame visible; used as Primary reflective plane containing both 수리영 and the supernatural mark; 거울에 비친 욕실 벽면 (A red circle has suddenly appeared in the reflected view) — The reflected camera-facing wall area bears a rough, vivid red circle; used as Carries the sharply readable crimson mark behind her reflected figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated bathroom illumination is punctured only by the sharply restrained crimson of the reflected circle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong has taken out her doll-adorned backpack for departure. The red circle exists only as a momentary apparition on the bathroom wall and leaves no lasting mark after she turns.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 화장실 거울 반사면 안, 수리영의 등 뒤 타일 벽면에 거칠고 선명한 붉은 원이 그려져 있는 시점 컷.\n\nLOCATION (lock): Inside the rooftop apartment’s bathroom, viewed through the mirror toward the tiled wall behind the sink. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the dolly-in at eye level just beside the mirror axis, framing 수리영's left-right-reversed reflection on the left half and the sharply readable red circle on the reflected wall behind her to the right. The mirror frame remains visible at an edge so the image cannot be mistaken for a direct view, while shallow focus still holds both her reflected face and the circle legible.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영's mirror reflection in the middle-left of the frame, midground; red circle reflected on bathroom wall in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 욕실 거울 (The red circle is visible only within the reflection at this instant) — Its camera-facing surface reflects a left-right-reversed 수리영 and the bathroom wall behind her, with part of the mirror frame visible; used as Primary reflective plane containing both 수리영 and the supernatural mark; 거울에 비친 욕실 벽면 (A red circle has suddenly appeared in the reflected view) — The reflected camera-facing wall area bears a rough, vivid red circle; used as Carries the sharply readable crimson mark behind her reflected figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated bathroom illumination is punctured only by the sharply restrained crimson of the reflected circle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong has taken out her doll-adorned backpack for departure. The red circle exists only as a momentary apparition on the bathroom wall and leaves no lasting mark after she turns.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained, natural-perspective lens feel, making the reflected face and the crimson ring equal points of tension across the frame. Shape the dim early-morning bathroom light into soft directional modeling on skin and grazing definition across the tiles, with gentle falloff beyond the two focal anchors. Keep the visible mirror edge as a narrow, dark physical boundary, and render the rough stroke seated in the tile texture and broken naturally by grout.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 화장실 거울 반사면 안, 수리영의 등 뒤 타일 벽면에 거칠고 선명한 붉은 원이 그려져 있는 시점 컷.\n\nLOCATION (lock): Inside the rooftop apartment’s bathroom, viewed through the mirror toward the tiled wall behind the sink. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Settle the dolly-in at eye level just beside the mirror axis, framing 수리영's left-right-reversed reflection on the left half and the sharply readable red circle on the reflected wall behind her to the right. The mirror frame remains visible at an edge so the image cannot be mistaken for a direct view, while shallow focus still holds both her reflected face and the circle legible.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영's mirror reflection in the middle-left of the frame, midground; red circle reflected on bathroom wall in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 욕실 거울 (The red circle is visible only within the reflection at this instant) — Its camera-facing surface reflects a left-right-reversed 수리영 and the bathroom wall behind her, with part of the mirror frame visible; used as Primary reflective plane containing both 수리영 and the supernatural mark; 거울에 비친 욕실 벽면 (A red circle has suddenly appeared in the reflected view) — The reflected camera-facing wall area bears a rough, vivid red circle; used as Carries the sharply readable crimson mark behind her reflected figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key, desaturated bathroom illumination is punctured only by the sharply restrained crimson of the reflected circle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong has taken out her doll-adorned backpack for departure. The red circle exists only as a momentary apparition on the bathroom wall and leaves no lasting mark after she turns.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a gently compressed lens feel to flatten the reflected depth into an uneasy arrangement of face, mirror plane, tiled grid, and ring. Favor broad, cool, low-key illumination with minimal facial modeling, letting the desaturated geometry dominate while the restrained crimson remains the sole sharp color accent. Preserve shallow focus but tune the focal plane so both the reflected face and the rough circular stroke read cleanly, with the rest of the bathroom receding softly.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 6,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "거울의 모서리를 포함한 구도와 반사된 피사체의 배치는 지시와 일치하나, 거울 밖 실제 타일 벽면에 붉은 자국이 유출된 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "거울이 두 개의 패널로 분리된 형태(콜라주/패널 위반)이며, 붉은 원이 반사된 벽이 아닌 거울 표면에 평면적으로 겹쳐져 있어 재질감을 상실했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B04.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "붉은 원의 하단부가 거울 프레임을 넘어 거울 밖의 실제 타일 벽면까지 침범하여 그려져 있습니다.",
     "fix_en": "Confine the entire red circle strictly to the reflected wall inside the mirror, erasing the portion that bleeds onto the physical wall outside the mirror frame."
    },
    {
     "issue_ko": "붉은 원은 거울 반사면에만 보여야 한다는 지시와 달리, 우측 거울 밖 실제 벽면(전경 인물 앞)에도 붉은 자국이 있습니다.",
     "fix_en": "Remove the red smear from the physical wall tiles on the far right side of the frame."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reframe so only Su-riyeong’s mirror reflection occupies the middle-left, removing her large direct-view head and shoulder from the right foreground while preserving the visible mirror edge.",
     "issue_ko": "화면 오른쪽 전경에 수리영의 실제 머리와 어깨가 크게 들어와, 거울 속 중경 인물과 반사된 배경 표식만으로 구성해야 하는 구도를 침범한다."
    },
    {
     "fix_en": "Shift the framing so the reflected red circle is centered in the middle-right of the full frame, keeping Su-riyeong’s reflection in the middle-left.",
     "issue_ko": "붉은 원의 중심이 전체 화면의 중앙 부근에 놓여 있어, 계약에서 지정한 중간 오른쪽 배치보다 지나치게 왼쪽이다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Confine the entire red circle strictly to the reflected wall inside the mirror, erasing the portion that bleeds onto the physical wall outside the mirror frame.\n- Remove the red smear from the physical wall tiles on the far right side of the frame.\n- Reframe so only Su-riyeong’s mirror reflection occupies the middle-left, removing her large direct-view head and shoulder from the right foreground while preserving the visible mirror edge.\n- Shift the framing so the reflected red circle is centered in the middle-right of the full frame, keeping Su-riyeong’s reflection in the middle-left.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "거울 속 반사된 인물과 붉은 원의 위치, 배낭 디테일 등 프롬프트의 요구사항을 정확히 충족한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "거울 밖 실제 벽면에 붉은 자국이 나타나 '반사면에만 보인다'는 설정을 위반했으며, 원의 형태도 온전하지 않고 앞쪽에 실제 인물의 어깨가 노출되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "거울 속 반사된 인물과 붉은 원의 위치, 배낭 디테일 등 프롬프트의 요구사항을 정확히 충족한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "거울 밖 실제 벽면에 붉은 자국이 나타나 '반사면에만 보인다'는 설정을 위반했으며, 원의 형태도 온전하지 않고 앞쪽에 실제 인물의 어깨가 노출되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "거울 안의 인물과 붉은 원의 배치가 지시문과 정확히 일치하며, 불필요한 전경 인물이나 외부 벽면의 얼룩 없이 깔끔하게 연출되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "거울 밖 실제 벽면에 붉은 얼룩이 나타나는 오류가 있으며, 원의 형태가 왜곡되고 우측 전경에 지시되지 않은 인물의 어깨가 포함되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "거울 안의 인물과 붉은 원의 배치가 지시문과 정확히 일치하며, 불필요한 전경 인물이나 외부 벽면의 얼룩 없이 깔끔하게 연출되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "거울 밖 실제 벽면에 붉은 얼룩이 나타나는 오류가 있으며, 원의 형태가 왜곡되고 우측 전경에 지시되지 않은 인물의 어깨가 포함되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B04",
   "choice": "Candidate D",
   "confident": true,
   "reason_ko": "화장실 거울과 타일 벽면, 그리고 붉은 원이 명시된 텍스트와 일치하는 유일한 화장실 거울 시점 이미지입니다.",
   "kept": "L06B04"
  }
 },
 "S18sh16::variants": {
  "author_fp": "5d75284b4a4adf77",
  "author": {
   "variants": [
    {
     "approach_ko": "정면의 평면성과 넓은 초점으로 사진 인화면의 변형을 냉정하고 법의학적인 디테일처럼 포착한다.",
     "prompt_en": "Use a rigid, dead-on macro treatment with the image plane held square and nearly all of the framed detail in crisp focus. Let dim, soft early-morning ambient light reveal the print grain, intact protective surface, frame edge, and tabletop texture with restrained tonal separation; keep the palette cold, dark, and desaturated. The clinical stillness of the composition should make the arrested downward deformation feel precise and deeply unnatural."
    },
    {
     "approach_ko": "극도로 얕은 초점을 정상적인 미소와 흘러내리는 형상의 경계에 고정해 주관적이고 숨 막히는 공포를 만든다.",
     "prompt_en": "Photograph the insert with an extremely shallow, intimate focus plane fixed on the transition between the intact smile and the descending distortion, allowing the retained frame edge and tabletop boundary to fall gently soft. Use low, grazing early-morning ambient light to draw subtle relief across the photographic surface while the surrounding tones sink into muted shadow. Emphasize the downward visual pull of the warped features so the frozen instant feels unstable and subjectively horrifying."
    }
   ]
  },
  "reused": false
 },
 "S18sh16": {
  "input_fingerprint": "c11bb3bf48f7a959",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 식탁 위 액자 속, 환하게 웃는 강민숙의 얼굴 형상이 촛농처럼 기괴하게 뭉개져 반쯤 흘러내린 mid-action 찰나의 클로즈업.\n\nLOCATION (lock): At the dining table in the rooftop apartment’s living area, focused on a framed family photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At photograph height, hold the midpoint of the direct push-in so the framed image and the partially collapsing smile occupy a tight isolated detail, with a slim portion of the frame and tabletop retained as physical context. The downward flow of 강민숙's facial features is frozen mid-action rather than shown as a completed transformation.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 식탁 위 액자 사진 (강민숙's photographed face is partially distorted at the captured instant) — The photograph's image-bearing front faces the camera directly, showing 강민숙 smiling as half of her face distorts and flows downward; used as Primary content surface showing the impossible facial distortion; 식탁 (The framed photograph rests on it); used as Provides a narrow environmental boundary beneath the isolated photograph.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: A dark, desaturated treatment preserves the photograph's tactile detail while the warped facial features carry the uncanny emphasis without an invented light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains the doll-adorned backpack. The framed family photograph remains physically intact despite the momentary vision of Minsuk’s face melting.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 식탁 위 액자 속, 환하게 웃는 강민숙의 얼굴 형상이 촛농처럼 기괴하게 뭉개져 반쯤 흘러내린 mid-action 찰나의 클로즈업.\n\nLOCATION (lock): At the dining table in the rooftop apartment’s living area, focused on a framed family photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At photograph height, hold the midpoint of the direct push-in so the framed image and the partially collapsing smile occupy a tight isolated detail, with a slim portion of the frame and tabletop retained as physical context. The downward flow of 강민숙's facial features is frozen mid-action rather than shown as a completed transformation.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 식탁 위 액자 사진 (강민숙's photographed face is partially distorted at the captured instant) — The photograph's image-bearing front faces the camera directly, showing 강민숙 smiling as half of her face distorts and flows downward; used as Primary content surface showing the impossible facial distortion; 식탁 (The framed photograph rests on it); used as Provides a narrow environmental boundary beneath the isolated photograph.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: A dark, desaturated treatment preserves the photograph's tactile detail while the warped facial features carry the uncanny emphasis without an invented light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains the doll-adorned backpack. The framed family photograph remains physically intact despite the momentary vision of Minsuk’s face melting.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a rigid, dead-on macro treatment with the image plane held square and nearly all of the framed detail in crisp focus. Let dim, soft early-morning ambient light reveal the print grain, intact protective surface, frame edge, and tabletop texture with restrained tonal separation; keep the palette cold, dark, and desaturated. The clinical stillness of the composition should make the arrested downward deformation feel precise and deeply unnatural.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 식탁 위 액자 속, 환하게 웃는 강민숙의 얼굴 형상이 촛농처럼 기괴하게 뭉개져 반쯤 흘러내린 mid-action 찰나의 클로즈업.\n\nLOCATION (lock): At the dining table in the rooftop apartment’s living area, focused on a framed family photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At photograph height, hold the midpoint of the direct push-in so the framed image and the partially collapsing smile occupy a tight isolated detail, with a slim portion of the frame and tabletop retained as physical context. The downward flow of 강민숙's facial features is frozen mid-action rather than shown as a completed transformation.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 식탁 위 액자 사진 (강민숙's photographed face is partially distorted at the captured instant) — The photograph's image-bearing front faces the camera directly, showing 강민숙 smiling as half of her face distorts and flows downward; used as Primary content surface showing the impossible facial distortion; 식탁 (The framed photograph rests on it); used as Provides a narrow environmental boundary beneath the isolated photograph.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: A dark, desaturated treatment preserves the photograph's tactile detail while the warped facial features carry the uncanny emphasis without an invented light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains the doll-adorned backpack. The framed family photograph remains physically intact despite the momentary vision of Minsuk’s face melting.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph the insert with an extremely shallow, intimate focus plane fixed on the transition between the intact smile and the descending distortion, allowing the retained frame edge and tabletop boundary to fall gently soft. Use low, grazing early-morning ambient light to draw subtle relief across the photographic surface while the surrounding tones sink into muted shadow. Emphasize the downward visual pull of the warped features so the frozen instant feels unstable and subjectively horrifying.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 대로 카메라를 정면으로 바라보는 구도를 잘 유지했으며, 얼굴이 촛농처럼 녹아내리는 찰나의 질감과 디테일이 훌륭하게 표현되었습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "액자가 카메라를 정면으로 바라보지 않고 비스듬하게 누워 있어 지정된 카메라 구도(정면)를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B04.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라 프레이밍이 피사체인 액자를 정면에서 수직으로 바라보지 않고 비스듬한 각도로 촬영되었습니다.",
     "fix_en": "Adjust the camera angle to face the framed photograph completely dead-on, keeping the image plane perfectly square to the lens."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reposition the camera at the photograph’s height and align it rigidly perpendicular to the image plane for a square, dead-on view.",
     "issue_ko": "카메라가 액자 정면축에서 비껴나 있고 약간 위에서 내려다봐 사진 면이 사선 원근으로 보인다."
    },
    {
     "fix_en": "Push in tighter so the photographed face dominates the insert, retaining only a slim frame edge and a narrow strip of tabletop.",
     "issue_ko": "액자 주변, 특히 왼쪽 식탁 면이 지나치게 넓게 보여 요구된 타이트한 디테일 클로즈업보다 프레이밍이 느슨하다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Adjust the camera angle to face the framed photograph completely dead-on, keeping the image plane perfectly square to the lens.\n- Reposition the camera at the photograph’s height and align it rigidly perpendicular to the image plane for a square, dead-on view.\n- Push in tighter so the photographed face dominates the insert, retaining only a slim frame edge and a narrow strip of tabletop.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "정면에서 프레임의 수직과 수평을 맞춘 'dead-on' 및 'square' 구도 지시를 정확히 구현하여 몰입감을 높였습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "카메라 앵글이 비스듬하게 설정되어 액자가 틀어져 보이므로, 정면(dead-on) 및 평면 유지(square) 지시를 어겼습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "정면에서 프레임의 수직과 수평을 맞춘 'dead-on' 및 'square' 구도 지시를 정확히 구현하여 몰입감을 높였습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "카메라 앵글이 비스듬하게 설정되어 액자가 틀어져 보이므로, 정면(dead-on) 및 평면 유지(square) 지시를 어겼습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 정면(dead-on) 앵글을 정확히 구현했으며, 촛농처럼 흘러내리는 얼굴의 디테일과 어둡고 차가운 조명 톤을 충실히 반영했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "얼굴이 녹아내리는 묘사는 훌륭하나, 카메라와 평행하게 정면을 향해야 한다는 구도 지침을 어기고 액자가 약간 틀어져 있습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 정면(dead-on) 앵글을 정확히 구현했으며, 촛농처럼 흘러내리는 얼굴의 디테일과 어둡고 차가운 조명 톤을 충실히 반영했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "얼굴이 녹아내리는 묘사는 훌륭하나, 카메라와 평행하게 정면을 향해야 한다는 구도 지침을 어기고 액자가 약간 틀어져 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 16
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B04",
   "choice": "Candidate D",
   "confident": true,
   "reason_ko": "원문에 '식탁 위'라는 명확한 지시가 있으며, 여러 후보(C, D, E) 중 식탁이 등장하는 현재 할당된 앵글(D)을 변경할 만한 다른 특정 구절이 없으므로 현재 할당본을 유지합니다.",
   "kept": "L06B04"
  }
 },
 "S19sh3::bgfirst_bg": {
  "input_fingerprint": "e3596a1049aea393",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 형광등 아래 마트 진열대 앞, 수리영과 유니폼을 입은 마트 남자 직원이 마주 선 채 정지된 측면 구도.\n\nLOCATION (lock): Inside a quiet supermarket aisle in the early morning, directly in front of the merchandise shelves under fluorescent lighting.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height beside the aisle, hold a static profile two-shot with 수리영 on the left and 마트 남자 직원 on the right, separated by a narrow conversational gap. Keep both upper bodies and part of the adjacent shelving visible, using the aisle lines to contain their exchange without shifting the established profile axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마트 진열대 (Present in the quiet market interior) — The shelf fronts angle away from the camera along the aisle beside the two figures; used as Frames the conversation laterally and establishes the market aisle; 형광등 (On but dim) — The light fixtures are positioned above the aisle, with their underside facing the floor; used as Provides the grounded overhead illumination described for the shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim fluorescent illumination holds the aisle in a dark, desaturated register with subdued facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어두운 형광등 아래 마트 진열대 앞, 수리영과 유니폼을 입은 마트 남자 직원이 마주 선 채 정지된 측면 구도.\n\nLOCATION (lock): Inside a quiet supermarket aisle in the early morning, directly in front of the merchandise shelves under fluorescent lighting.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height beside the aisle, hold a static profile two-shot with 수리영 on the left and 마트 남자 직원 on the right, separated by a narrow conversational gap. Keep both upper bodies and part of the adjacent shelving visible, using the aisle lines to contain their exchange without shifting the established profile axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마트 진열대 (Present in the quiet market interior) — The shelf fronts angle away from the camera along the aisle beside the two figures; used as Frames the conversation laterally and establishes the market aisle; 형광등 (On but dim) — The light fixtures are positioned above the aisle, with their underside facing the floor; used as Provides the grounded overhead illumination described for the shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim fluorescent illumination holds the aisle in a dark, desaturated register with subdued facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S19sh3__bgfirst_bg.png",
  "asset_id": "13026940-6875-47ef-b528-6426212934f3",
  "input_asset_ids": [
   "4ccedf97-5c76-456f-9d39-c0374f94e549",
   "90f1d3de-fe9f-4886-8599-a3e716868d96"
  ]
 },
 "S19sh3": {
  "input_fingerprint": "3364afda16eaaa1c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 어두운 형광등 아래 마트 진열대 앞, 수리영과 유니폼을 입은 마트 남자 직원이 마주 선 채 정지된 측면 구도.\n\nLOCATION (lock): Inside a quiet supermarket aisle in the early morning, directly in front of the merchandise shelves under fluorescent lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height beside the aisle, hold a static profile two-shot with 수리영 on the left and 마트 남자 직원 on the right, separated by a narrow conversational gap. Keep both upper bodies and part of the adjacent shelving visible, using the aisle lines to contain their exchange without shifting the established profile axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마트 진열대 (Present in the quiet market interior) — The shelf fronts angle away from the camera along the aisle beside the two figures; used as Frames the conversation laterally and establishes the market aisle; 형광등 (On but dim) — The light fixtures are positioned above the aisle, with their underside facing the floor; used as Provides the grounded overhead illumination described for the shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim fluorescent illumination holds the aisle in a dark, desaturated register with subdued facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s bicycle remains parked outside the mart, and she continues to carry her doll-adorned backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 마트 남자 직원 (한국인 성인 남성 얼굴, 일반적인 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 어두운 형광등 아래 마트 진열대 앞, 수리영과 유니폼을 입은 마트 남자 직원이 마주 선 채 정지된 측면 구도.\n\nLOCATION (lock): Inside a quiet supermarket aisle in the early morning, directly in front of the merchandise shelves under fluorescent lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height beside the aisle, hold a static profile two-shot with 수리영 on the left and 마트 남자 직원 on the right, separated by a narrow conversational gap. Keep both upper bodies and part of the adjacent shelving visible, using the aisle lines to contain their exchange without shifting the established profile axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마트 진열대 (Present in the quiet market interior) — The shelf fronts angle away from the camera along the aisle beside the two figures; used as Frames the conversation laterally and establishes the market aisle; 형광등 (On but dim) — The light fixtures are positioned above the aisle, with their underside facing the floor; used as Provides the grounded overhead illumination described for the shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim fluorescent illumination holds the aisle in a dark, desaturated register with subdued facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s bicycle remains parked outside the mart, and she continues to carry her doll-adorned backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 마트 남자 직원 (한국인 성인 남성 얼굴, 일반적인 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 어두운 형광등 아래 마트 진열대 앞, 수리영과 유니폼을 입은 마트 남자 직원이 마주 선 채 정지된 측면 구도.\n\nLOCATION (lock): Inside a quiet supermarket aisle in the early morning, directly in front of the merchandise shelves under fluorescent lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height beside the aisle, hold a static profile two-shot with 수리영 on the left and 마트 남자 직원 on the right, separated by a narrow conversational gap. Keep both upper bodies and part of the adjacent shelving visible, using the aisle lines to contain their exchange without shifting the established profile axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마트 진열대 (Present in the quiet market interior) — The shelf fronts angle away from the camera along the aisle beside the two figures; used as Frames the conversation laterally and establishes the market aisle; 형광등 (On but dim) — The light fixtures are positioned above the aisle, with their underside facing the floor; used as Provides the grounded overhead illumination described for the shot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dim fluorescent illumination holds the aisle in a dark, desaturated register with subdued facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong’s bicycle remains parked outside the mart, and she continues to carry her doll-adorned backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 마트 남자 직원 (한국인 성인 남성 얼굴, 일반적인 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S19sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S19sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L11B04.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 측면 미디엄 샷과 어두운 조명을 잘 구현했으며, 수리영의 레퍼런스 복장(비니, 재킷)과 배낭을 완벽하게 유지했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "구도와 마트 직원의 유니폼 설정은 좋으나, 수리영의 복장이 레퍼런스와 전혀 다른 회색 후드티로 변경되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 측면 미디엄 샷과 어두운 조명을 잘 구현했으며, 수리영의 레퍼런스 복장(비니, 재킷)과 배낭을 완벽하게 유지했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "구도와 마트 직원의 유니폼 설정은 좋으나, 수리영의 복장이 레퍼런스와 전혀 다른 회색 후드티로 변경되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "수리영의 인물 레퍼런스(비니, 짙은 재킷)와 인형이 달린 백팩, 마트 직원의 유니폼을 명시된 측면 미디엄 투샷으로 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "요구된 구도와 조명은 잘 따랐으나, 수리영의 의상(비니 누락, 회색 후드티 착용)이 지정된 레퍼런스와 전혀 일치하지 않습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "수리영의 인물 레퍼런스(비니, 짙은 재킷)와 인형이 달린 백팩, 마트 직원의 유니폼을 명시된 측면 미디엄 투샷으로 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "요구된 구도와 조명은 잘 따랐으나, 수리영의 의상(비니 누락, 회색 후드티 착용)이 지정된 레퍼런스와 전혀 일치하지 않습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 16
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 10,
   "B": 16
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "지시된 측면 미디엄 샷과 어두운 조명을 잘 구현했으며, 수리영의 레퍼런스 복장(비니, 재킷)과 배낭을 완벽하게 유지했습니다."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "구도와 마트 직원의 유니폼 설정은 좋으나, 수리영의 복장이 레퍼런스와 전혀 다른 회색 후드티로 변경되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L11B04.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": []
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Tighten the framing to a medium two-shot showing both figures primarily from the waist up while preserving Su-riyeong on the left, the employee on the right, their profile axis, and the narrow gap.",
     "issue_ko": "두 인물이 허벅지까지 보이고 주변 통로가 넓게 포함되어, 지정된 미디엄 숏보다 프레이밍이 넓고 인물의 화면 크기가 작다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Tighten the framing to a medium two-shot showing both figures primarily from the waist up while preserving Su-riyeong on the left, the employee on the right, their profile axis, and the narrow gap.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 앵글과 인물 배치를 정확히 구현했으며, 배경 레퍼런스의 미세한 디테일(진열대 라벨 등)까지 완벽하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "전반적인 구도와 인물 지시사항을 잘 따랐으나, 배경의 세부적인 텍스트와 디테일 묘사에서 A에 비해 약간 떨어집니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 앵글과 인물 배치를 정확히 구현했으며, 배경 레퍼런스의 미세한 디테일(진열대 라벨 등)까지 완벽하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "전반적인 구도와 인물 지시사항을 잘 따랐으나, 배경의 세부적인 텍스트와 디테일 묘사에서 A에 비해 약간 떨어집니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 미디엄 샷과 측면 투샷 구도를 정확히 구현했으며, 인상착의와 어두운 마트 조명 분위기 및 소품 디테일이 우수합니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "구도와 배경, 인물의 위치 등 전반적인 프롬프트 요구사항을 잘 따랐으나, 손과 소품 등 미세한 디테일의 선명도가 A에 비해 약간 떨어집니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 미디엄 샷과 측면 투샷 구도를 정확히 구현했으며, 인상착의와 어두운 마트 조명 분위기 및 소품 디테일이 우수합니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "구도와 배경, 인물의 위치 등 전반적인 프롬프트 요구사항을 잘 따랐으나, 손과 소품 등 미세한 디테일의 선명도가 A에 비해 약간 떨어집니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 13
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S19sh3__bgfirst_bg.png",
   "bg_asset_id": "13026940-6875-47ef-b528-6426212934f3",
   "bg_record_key": "S19sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L11B01",
    "B": "L11B02",
    "C": "L11B03",
    "D": "L11B04"
   },
   "assigned": "L11B04",
   "choice": "Candidate D",
   "confident": true,
   "reason_ko": "\"어두운 형광등 아래 마트 진열대 앞\"이라는 묘사에 부합하는 마트 내부 진열대 공간입니다.",
   "kept": "L11B04"
  }
 },
 "S19sh7::bgfirst_bg": {
  "input_fingerprint": "55c37a6fa062c727",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고개를 뒤로 젖혀 마트 외부 천장의 CCTV를 뚫어지게 올려다보는 수리영의 상반신.\n\nLOCATION (lock): Under the covered exterior entrance of the supermarket, looking up toward a surveillance camera mounted above the doorway.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the store, tighten at upper-torso height in front of 수리영 and preserve the upward tilt along her raised eyeline, placing her chest and craned neck low in frame while leaving the exterior CCTV above her. The key change is her gaze lifting decisively toward the camera unit as the tilt reaches its endpoint.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the lower-center of the frame, foreground, looks toward exterior CCTV camera; exterior CCTV camera in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 마트 외부 CCTV (Fixed beneath the exterior ceiling) — Its lens-bearing front is angled toward the area outside the store and remains visible above 수리영; used as The elevated object anchoring 수리영's upward sightline; 마트 외부 천장 (The CCTV is mounted beneath it) — Its underside faces the camera above 수리영; used as Provides the overhead boundary that holds the CCTV above her.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural early-morning exterior ambient light is rendered with restrained contrast and a dark, desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고개를 뒤로 젖혀 마트 외부 천장의 CCTV를 뚫어지게 올려다보는 수리영의 상반신.\n\nLOCATION (lock): Under the covered exterior entrance of the supermarket, looking up toward a surveillance camera mounted above the doorway.\n\nTIME OF DAY (lock): early morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the store, tighten at upper-torso height in front of 수리영 and preserve the upward tilt along her raised eyeline, placing her chest and craned neck low in frame while leaving the exterior CCTV above her. The key change is her gaze lifting decisively toward the camera unit as the tilt reaches its endpoint.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the lower-center of the frame, foreground, looks toward exterior CCTV camera; exterior CCTV camera in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 마트 외부 CCTV (Fixed beneath the exterior ceiling) — Its lens-bearing front is angled toward the area outside the store and remains visible above 수리영; used as The elevated object anchoring 수리영's upward sightline; 마트 외부 천장 (The CCTV is mounted beneath it) — Its underside faces the camera above 수리영; used as Provides the overhead boundary that holds the CCTV above her.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural early-morning exterior ambient light is rendered with restrained contrast and a dark, desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S19sh7__bgfirst_bg.png",
  "asset_id": "32b71ae0-8060-42ea-a606-54a70d09a8b8",
  "input_asset_ids": [
   "b81f99ab-3ae8-4425-a213-25682362a146",
   "c2e803ac-08f2-4adb-99eb-cb3af6964a7e",
   "e10f4013-8646-462c-8bb2-fdcce13eef2b"
  ]
 },
 "S19sh7": {
  "input_fingerprint": "7eac1ad113e4d8c5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 고개를 뒤로 젖혀 마트 외부 천장의 CCTV를 뚫어지게 올려다보는 수리영의 상반신.\n\nLOCATION (lock): Under the covered exterior entrance of the supermarket, looking up toward a surveillance camera mounted above the doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the store, tighten at upper-torso height in front of 수리영 and preserve the upward tilt along her raised eyeline, placing her chest and craned neck low in frame while leaving the exterior CCTV above her. The key change is her gaze lifting decisively toward the camera unit as the tilt reaches its endpoint.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the lower-center of the frame, foreground, looks toward exterior CCTV camera; exterior CCTV camera in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 마트 외부 CCTV (Fixed beneath the exterior ceiling) — Its lens-bearing front is angled toward the area outside the store and remains visible above 수리영; used as The elevated object anchoring 수리영's upward sightline; 마트 외부 천장 (The CCTV is mounted beneath it) — Its underside faces the camera above 수리영; used as Provides the overhead boundary that holds the CCTV above her.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural early-morning exterior ambient light is rendered with restrained contrast and a dark, desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her bicycle remains parked at the mart and the doll-adorned backpack remains with her as she checks the exterior CCTV position.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 고개를 뒤로 젖혀 마트 외부 천장의 CCTV를 뚫어지게 올려다보는 수리영의 상반신.\n\nLOCATION (lock): Under the covered exterior entrance of the supermarket, looking up toward a surveillance camera mounted above the doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the store, tighten at upper-torso height in front of 수리영 and preserve the upward tilt along her raised eyeline, placing her chest and craned neck low in frame while leaving the exterior CCTV above her. The key change is her gaze lifting decisively toward the camera unit as the tilt reaches its endpoint.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the lower-center of the frame, foreground, looks toward exterior CCTV camera; exterior CCTV camera in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 마트 외부 CCTV (Fixed beneath the exterior ceiling) — Its lens-bearing front is angled toward the area outside the store and remains visible above 수리영; used as The elevated object anchoring 수리영's upward sightline; 마트 외부 천장 (The CCTV is mounted beneath it) — Its underside faces the camera above 수리영; used as Provides the overhead boundary that holds the CCTV above her.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural early-morning exterior ambient light is rendered with restrained contrast and a dark, desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her bicycle remains parked at the mart and the doll-adorned backpack remains with her as she checks the exterior CCTV position.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 고개를 뒤로 젖혀 마트 외부 천장의 CCTV를 뚫어지게 올려다보는 수리영의 상반신.\n\nLOCATION (lock): Under the covered exterior entrance of the supermarket, looking up toward a surveillance camera mounted above the doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Outside the store, tighten at upper-torso height in front of 수리영 and preserve the upward tilt along her raised eyeline, placing her chest and craned neck low in frame while leaving the exterior CCTV above her. The key change is her gaze lifting decisively toward the camera unit as the tilt reaches its endpoint.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the lower-center of the frame, foreground, looks toward exterior CCTV camera; exterior CCTV camera in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 마트 외부 CCTV (Fixed beneath the exterior ceiling) — Its lens-bearing front is angled toward the area outside the store and remains visible above 수리영; used as The elevated object anchoring 수리영's upward sightline; 마트 외부 천장 (The CCTV is mounted beneath it) — Its underside faces the camera above 수리영; used as Provides the overhead boundary that holds the CCTV above her.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural early-morning exterior ambient light is rendered with restrained contrast and a dark, desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her bicycle remains parked at the mart and the doll-adorned backpack remains with her as she checks the exterior CCTV position.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S19sh7__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S19sh7.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L11B02.png"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/background_chain/seed_bg_large_market_sel.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요청된 정면 로우앵글 카메라 축과 프레이밍을 정확히 구현하였으며, 인물의 시선과 목선이 지시대로 잘 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "건물 외관의 디테일은 참조와 일치하지만, 정면에서 올려다보는 카메라 축을 무시하고 측면 앵글로 촬영되어 프레이밍 지시를 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요청된 정면 로우앵글 카메라 축과 프레이밍을 정확히 구현하였으며, 인물의 시선과 목선이 지시대로 잘 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "건물 외관의 디테일은 참조와 일치하지만, 정면에서 올려다보는 카메라 축을 무시하고 측면 앵글로 촬영되어 프레이밍 지시를 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "피사체 정면에서 올려다보는 카메라 축과 지정된 미디엄 샷 프레이밍(우선순위 2)을 정확히 구현하여 배경 구조의 변형에도 불구하고 승리했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "장소는 잘 재현했으나, 지시된 정면 앵글 대신 측면 구도와 너무 넓은 화각을 사용하여 프레이밍(우선순위 2)에서 감점되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "피사체 정면에서 올려다보는 카메라 축과 지정된 미디엄 샷 프레이밍(우선순위 2)을 정확히 구현하여 배경 구조의 변형에도 불구하고 승리했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "장소는 잘 재현했으나, 지시된 정면 앵글 대신 측면 구도와 너무 넓은 화각을 사용하여 프레이밍(우선순위 2)에서 감점되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 8,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "요청된 정면 로우앵글 카메라 축과 프레이밍을 정확히 구현하였으며, 인물의 시선과 목선이 지시대로 잘 표현되었습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "건물 외관의 디테일은 참조와 일치하지만, 정면에서 올려다보는 카메라 축을 무시하고 측면 앵글로 촬영되어 프레이밍 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L11B02.png"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/background_chain/seed_bg_large_market_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "CCTV가 상단 중앙의 천장이 아닌 좌측 상단의 벽면에 위치해 있어 프레이밍 지시를 위반했으며, 인물의 시선과도 일치하지 않습니다.",
     "fix_en": "Move the CCTV to the upper-center of the frame, mounting it on the ceiling directly above her so it aligns with her upward gaze."
    },
    {
     "issue_ko": "구조물 레퍼런스에 있는 평평한 처마 지붕 대신, 공간에 존재하지 않는 거대한 경사진 터널형 지붕 구조가 임의로 생성되었습니다.",
     "fix_en": "Replace the invented sloped tunnel ceiling with the flat, horizontal overhanging roof shown in the structure reference."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reframe so the exterior CCTV occupies the upper-center of the frame while keeping Suriyoung lower-center.",
     "issue_ko": "외부 CCTV가 계약된 화면 상단 중앙이 아니라 상단 왼쪽에 배치되어 있다."
    },
    {
     "fix_en": "Turn Suriyoung's raised eyeline toward the CCTV so her gaze axis terminates directly on the camera unit.",
     "issue_ko": "수리영의 시선이 상단 중앙을 향해 있어 왼쪽 위 CCTV와 연결되지 않는다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Move the CCTV to the upper-center of the frame, mounting it on the ceiling directly above her so it aligns with her upward gaze.\n- Replace the invented sloped tunnel ceiling with the flat, horizontal overhanging roof shown in the structure reference.\n- Reframe so the exterior CCTV occupies the upper-center of the frame while keeping Suriyoung lower-center.\n- Turn Suriyoung's raised eyeline toward the CCTV so her gaze axis terminates directly on the camera unit.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "수리영의 모습과 장소는 잘 구현되었으나, 프레이밍 지시 및 텍스트와 달리 CCTV가 천장 중앙이 아닌 좌측 벽면에 위치해 있습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "수리영의 상반신 구도와 시선, 그리고 프레임 상단 중앙 천장에 달린 CCTV의 위치까지 프롬프트의 지시를 매우 정확하게 따랐습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "수리영의 모습과 장소는 잘 구현되었으나, 프레이밍 지시 및 텍스트와 달리 CCTV가 천장 중앙이 아닌 좌측 벽면에 위치해 있습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "수리영의 상반신 구도와 시선, 그리고 프레임 상단 중앙 천장에 달린 CCTV의 위치까지 프롬프트의 지시를 매우 정확하게 따랐습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프레임 레이아웃 지침에 맞춰 천장 CCTV를 화면 상단 중앙에 정확히 배치하였으며, 피사체의 시선 방향도 대상과 자연스럽게 일치합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "CCTV가 화면 좌측 벽면에 치우쳐 있어 상단 중앙에 배치하라는 레이아웃 지시를 어겼으며, 인물의 시선이 CCTV를 향하지 않습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프레임 레이아웃 지침에 맞춰 천장 CCTV를 화면 상단 중앙에 정확히 배치하였으며, 피사체의 시선 방향도 대상과 자연스럽게 일치합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "CCTV가 화면 좌측 벽면에 치우쳐 있어 상단 중앙에 배치하라는 레이아웃 지시를 어겼으며, 인물의 시선이 CCTV를 향하지 않습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S19sh7__bgfirst_bg.png",
   "bg_asset_id": "32b71ae0-8060-42ea-a606-54a70d09a8b8",
   "bg_record_key": "S19sh7::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready",
  "plate_select": {
   "candidates": {
    "A": "L11B01",
    "B": "L11B02",
    "C": "L11B03",
    "D": "L11B04"
   },
   "assigned": "L11B03",
   "choice": "Candidate C",
   "confident": true,
   "reason_ko": "원문에 '마트 외부 천장의 CCTV'가 명시되어 있으며, 현재 할당된 후보 C가 마트 외부와 CCTV가 설치된 처마(천장) 아래 공간을 보여주므로 이를 유지합니다.",
   "kept": "L11B03"
  }
 },
 "S19sh8::variants": {
  "author_fp": "bc9ef3eab5775402",
  "author": {
   "variants": [
    {
     "approach_ko": "압축된 망원 시선과 얕은 심도로 얼굴의 결연한 정적을 고립시키는 정면 초상.",
     "prompt_en": "Use a compressed long-lens feel and very shallow depth of field, holding the facial plane, phone, and supporting hand in precise focus while the entrance dissolves into muted shapes. Keep the composition nearly symmetrical and visually sealed around her expression, with soft, cool early-morning ambience and restrained contrast emphasizing stillness rather than movement."
    },
    {
     "approach_ko": "조금 더 자연스러운 원근과 출입구의 구조선을 활용해 인물을 압박하는 공간감 있는 정면 클로즈업.",
     "prompt_en": "Use a natural perspective with moderately retained depth, allowing the softly rendered entrance geometry and canopy lines to remain legible as a rigid frame around the centered portrait. Let the early-morning ambient light model one side of her face slightly more than the other, preserving realistic skin and fabric texture while the subdued exterior recedes behind her."
    }
   ]
  },
  "reused": false
 },
 "S19sh8": {
  "input_fingerprint": "ae82e703b9f25415",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 수리영이 핸드폰을 귀에 바짝 댄 채 입술을 굳게 다물고 정면을 응시하는 정지 순간.\n\nLOCATION (lock): Outside the supermarket entrance beneath the exterior CCTV coverage, where the daughter calls the police officer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold close at eye level directly in front of 수리영, her phone pressed to one ear and her upper body centered with minimal exterior context at the edges. The emphasized change from the upward-looking shot is her gaze settling straight ahead as her lips clamp shut, creating a deliberate confrontational portrait.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 마트 외부 출입부 (Present behind her) — The store-facing side remains behind 수리영 and outside primary focus; used as Soft, limited context behind the fixed frontal portrait.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural early-morning ambient light keeps her fixed expression stark within a muted, desaturated exterior palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the market exterior canopy, fluorescent lighting, daytime look, and the young woman's clothing from the reference. Exclude the overhead CCTV emphasis and her upward-looking pose; show her holding the phone to her ear and facing forward.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains her backpack and holds her phone to call Hye-su about the CCTV.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 수리영이 핸드폰을 귀에 바짝 댄 채 입술을 굳게 다물고 정면을 응시하는 정지 순간.\n\nLOCATION (lock): Outside the supermarket entrance beneath the exterior CCTV coverage, where the daughter calls the police officer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold close at eye level directly in front of 수리영, her phone pressed to one ear and her upper body centered with minimal exterior context at the edges. The emphasized change from the upward-looking shot is her gaze settling straight ahead as her lips clamp shut, creating a deliberate confrontational portrait.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 마트 외부 출입부 (Present behind her) — The store-facing side remains behind 수리영 and outside primary focus; used as Soft, limited context behind the fixed frontal portrait.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural early-morning ambient light keeps her fixed expression stark within a muted, desaturated exterior palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the market exterior canopy, fluorescent lighting, daytime look, and the young woman's clothing from the reference. Exclude the overhead CCTV emphasis and her upward-looking pose; show her holding the phone to her ear and facing forward.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains her backpack and holds her phone to call Hye-su about the CCTV.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a compressed long-lens feel and very shallow depth of field, holding the facial plane, phone, and supporting hand in precise focus while the entrance dissolves into muted shapes. Keep the composition nearly symmetrical and visually sealed around her expression, with soft, cool early-morning ambience and restrained contrast emphasizing stillness rather than movement.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): early morning.\n\nSHOT TEXT (authoritative, Korean): 수리영이 핸드폰을 귀에 바짝 댄 채 입술을 굳게 다물고 정면을 응시하는 정지 순간.\n\nLOCATION (lock): Outside the supermarket entrance beneath the exterior CCTV coverage, where the daughter calls the police officer. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold close at eye level directly in front of 수리영, her phone pressed to one ear and her upper body centered with minimal exterior context at the edges. The emphasized change from the upward-looking shot is her gaze settling straight ahead as her lips clamp shut, creating a deliberate confrontational portrait.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 마트 외부 출입부 (Present behind her) — The store-facing side remains behind 수리영 and outside primary focus; used as Soft, limited context behind the fixed frontal portrait.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural early-morning ambient light keeps her fixed expression stark within a muted, desaturated exterior palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the market exterior canopy, fluorescent lighting, daytime look, and the young woman's clothing from the reference. Exclude the overhead CCTV emphasis and her upward-looking pose; show her holding the phone to her ear and facing forward.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains her backpack and holds her phone to call Hye-su about the CCTV.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a natural perspective with moderately retained depth, allowing the softly rendered entrance geometry and canopy lines to remain legible as a rigid frame around the centered portrait. Let the early-morning ambient light model one side of her face slightly more than the other, preserving realistic skin and fabric texture while the subdued exterior recedes behind her.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 3,
   "B": 8
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "배경 구조물이 레퍼런스의 콘크리트 캐노피 및 입구와 전혀 일치하지 않으며, 캐릭터의 헤어스타일과 인형 키링 등 이전 샷과의 디테일 연결성이 떨어집니다."
   },
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "지정된 장소의 배경(콘크리트 구조물, 유리문)을 완벽하게 유지하면서, 지시된 클로즈업 프레이밍과 캐릭터의 디테일(묶은 머리, 인형 키링) 및 정면을 응시하는 연기를 정확히 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S19sh7_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "핸드폰을 귀에 대고 있으나 화면이 바깥쪽을 향하고 있어 방향이 잘못되었습니다.",
     "fix_en": "Flip the phone so the screen faces her ear and the back of the phone faces the camera."
    },
    {
     "issue_ko": "오른쪽 가방 끈에 달린 인형 열쇠고리의 형태가 뭉개져 재킷과 융합되어 있습니다.",
     "fix_en": "Render the doll keychain on the right shoulder strap as a clearly defined, separate physical object."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Flip the phone so the screen faces her ear and the back of the phone faces the camera.\n- Render the doll keychain on the right shoulder strap as a clearly defined, separate physical object.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 클로즈업 프레이밍과 굳은 표정을 훌륭하게 구현했으며, 핸드폰의 후면이 카메라를 향하도록 소품의 방향 규칙을 정확히 준수했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "프레이밍과 인물의 표정은 좋으나, 핸드폰의 화면(기능면)이 카메라 쪽을 향하고 있어 소품 방향에 대한 명시적인 지침을 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 클로즈업 프레이밍과 굳은 표정을 훌륭하게 구현했으며, 핸드폰의 후면이 카메라를 향하도록 소품의 방향 규칙을 정확히 준수했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "프레이밍과 인물의 표정은 좋으나, 핸드폰의 화면(기능면)이 카메라 쪽을 향하고 있어 소품 방향에 대한 명시적인 지침을 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "이전 샷의 모자가 누락된 점은 아쉬우나, 요구된 클로즈업 구도와 입술을 굳게 다물고 정면을 응시하는 자세를 해부학적 오류 없이 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "모자 누락뿐만 아니라, 핸드폰을 쥔 손가락의 구조와 옷에 매달린 인형의 형태가 심하게 뭉개지는 물리적/해부학적 오류가 발생했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "이전 샷의 모자가 누락된 점은 아쉬우나, 요구된 클로즈업 구도와 입술을 굳게 다물고 정면을 응시하는 자세를 해부학적 오류 없이 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "모자 누락뿐만 아니라, 핸드폰을 쥔 손가락의 구조와 옷에 매달린 인형의 형태가 심하게 뭉개지는 물리적/해부학적 오류가 발생했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S19sh7"
  }
 },
 "S20sh2::variants": {
  "author_fp": "1b981d3a8bd9d601",
  "author": {
   "variants": [
    {
     "approach_ko": "정면 대칭의 임상적 증거 화면처럼, 화면 전체의 노이즈와 압축 손상을 선명하고 냉정하게 포착한다.",
     "prompt_en": "Photograph the monitor with rigid, forensic symmetry, using a neutral close-focus lens feel and even edge-to-edge sharpness across the display plane. Hold the crushed blacks, weak electronic luminance, blocky compression, and granular video noise with clinical clarity; the two separated human figures remain physically plausible but equally unreadable through the recording damage. Keep the slim bezel crisp and unobtrusive, making the paused image feel like cold, unyielding evidence."
    },
    {
     "approach_ko": "화면에 바짝 밀착한 압박감 있는 시선으로, 픽셀 질감과 어둠 속 뭉개짐이 인물을 집어삼키는 듯 강조한다.",
     "prompt_en": "Use an intensely intimate optical treatment that feels pressed against the monitor, with the central screen texture and interference sharply resolved while focus falls off subtly toward the thin bezel. Let dense noise, smeared motion residue, and compressed shadow detail dominate the visual rhythm, swallowing identifying features without turning either separated person into an abstract silhouette. Render the faint electronic glow as murky and oppressive, emphasizing the disturbing stillness of the paused frame."
    }
   ]
  },
  "reused": false
 },
 "S20sh2": {
  "input_fingerprint": "270a58b83e471948",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둡고 노이즈가 낀 CCTV 화면 속, 두 사람의 형상이 기괴하게 뭉개진 채 멈춘 모니터 클로즈업.\n\nLOCATION (lock): At the CCTV monitoring desk inside the supermarket’s security office, lit primarily by the surveillance monitors. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the push-in directly on the monitor axis at screen-center height, holding the paused CCTV image close enough that noise and the two distorted human forms occupy most of the composition while a thin display edge remains visible. Keep the forms separated within the screen but equally unreadable, making the display rather than the surrounding office the sole visual focus.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: paused CCTV monitor image in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: CCTV 모니터 화면 (On and paused at the difficult-to-identify section) — The display face is square to the camera and shows a paused, unusually dark recording with two distorted human forms and visible image noise; used as Primary device surface presenting the obscured evidence; 모니터 외곽 (Present around the active screen) — Only a slim front bezel around the display faces the camera; used as A narrow edge confirming that the image is being viewed on a physical monitor.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark monitor image supplies a subdued electronic luminance, with heavy tonal compression and visible video noise rather than invented environmental lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The relevant section of the CCTV recording remains uniquely dark and distorted, preventing identification of Minsuk’s male companion.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둡고 노이즈가 낀 CCTV 화면 속, 두 사람의 형상이 기괴하게 뭉개진 채 멈춘 모니터 클로즈업.\n\nLOCATION (lock): At the CCTV monitoring desk inside the supermarket’s security office, lit primarily by the surveillance monitors. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the push-in directly on the monitor axis at screen-center height, holding the paused CCTV image close enough that noise and the two distorted human forms occupy most of the composition while a thin display edge remains visible. Keep the forms separated within the screen but equally unreadable, making the display rather than the surrounding office the sole visual focus.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: paused CCTV monitor image in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: CCTV 모니터 화면 (On and paused at the difficult-to-identify section) — The display face is square to the camera and shows a paused, unusually dark recording with two distorted human forms and visible image noise; used as Primary device surface presenting the obscured evidence; 모니터 외곽 (Present around the active screen) — Only a slim front bezel around the display faces the camera; used as A narrow edge confirming that the image is being viewed on a physical monitor.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark monitor image supplies a subdued electronic luminance, with heavy tonal compression and visible video noise rather than invented environmental lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The relevant section of the CCTV recording remains uniquely dark and distorted, preventing identification of Minsuk’s male companion.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph the monitor with rigid, forensic symmetry, using a neutral close-focus lens feel and even edge-to-edge sharpness across the display plane. Hold the crushed blacks, weak electronic luminance, blocky compression, and granular video noise with clinical clarity; the two separated human figures remain physically plausible but equally unreadable through the recording damage. Keep the slim bezel crisp and unobtrusive, making the paused image feel like cold, unyielding evidence.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둡고 노이즈가 낀 CCTV 화면 속, 두 사람의 형상이 기괴하게 뭉개진 채 멈춘 모니터 클로즈업.\n\nLOCATION (lock): At the CCTV monitoring desk inside the supermarket’s security office, lit primarily by the surveillance monitors. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the push-in directly on the monitor axis at screen-center height, holding the paused CCTV image close enough that noise and the two distorted human forms occupy most of the composition while a thin display edge remains visible. Keep the forms separated within the screen but equally unreadable, making the display rather than the surrounding office the sole visual focus.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: paused CCTV monitor image in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: CCTV 모니터 화면 (On and paused at the difficult-to-identify section) — The display face is square to the camera and shows a paused, unusually dark recording with two distorted human forms and visible image noise; used as Primary device surface presenting the obscured evidence; 모니터 외곽 (Present around the active screen) — Only a slim front bezel around the display faces the camera; used as A narrow edge confirming that the image is being viewed on a physical monitor.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark monitor image supplies a subdued electronic luminance, with heavy tonal compression and visible video noise rather than invented environmental lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The relevant section of the CCTV recording remains uniquely dark and distorted, preventing identification of Minsuk’s male companion.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse an intensely intimate optical treatment that feels pressed against the monitor, with the central screen texture and interference sharply resolved while focus falls off subtly toward the thin bezel. Let dense noise, smeared motion residue, and compressed shadow detail dominate the visual rhythm, swallowing identifying features without turning either separated person into an abstract silhouette. Render the faint electronic glow as murky and oppressive, emphasizing the disturbing stillness of the paused frame.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지정된 꽉 찬 클로즈업 구도를 정확히 따르며, 두 사람의 뭉개진 형상과 화면 노이즈를 사실적인 CCTV 시각 효과로 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "모니터 속 화면이 4분할된 패널 위에 거대한 인물들이 물리적으로 불가능하게 겹쳐진 콜라주 형태로 렌더링되어 심각한 규칙 위반입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L14B01.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "좌우에 보이는 주변 모니터 화면과 베젤에 날짜, 숫자, 브랜드 로고 등의 텍스트가 포함되어 있습니다 (텍스트 금지 지시 위반).",
     "fix_en": "Remove all text, numbers, timestamps, and logos from the screens and bezels of all visible monitors."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Push in and crop symmetrically so the paused central CCTV image fills the frame with only its slim bezel visible, excluding both adjacent monitors and the monitor stand.",
     "issue_ko": "프레이밍이 너무 넓어 중앙 화면 외에 양옆 모니터와 하단 스탠드까지 보여, 디스플레이와 얇은 베젤만 프레임을 지배해야 한다는 클로즈업 구성을 위반한다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove all text, numbers, timestamps, and logos from the screens and bezels of all visible monitors.\n- Push in and crop symmetrically so the paused central CCTV image fills the frame with only its slim bezel visible, excluding both adjacent monitors and the monitor stand.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 얇은 베젤 외에 양옆의 다른 모니터와 텍스트가 노출되어 프레이밍 및 텍스트 금지 지침을 위반했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요청된 화면 비율과 단일 모니터 클로즈업 구도를 정확히 따랐으며, 텍스트 없이 기괴하게 왜곡된 두 사람의 형상을 잘 구현했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 얇은 베젤 외에 양옆의 다른 모니터와 텍스트가 노출되어 프레이밍 및 텍스트 금지 지침을 위반했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요청된 화면 비율과 단일 모니터 클로즈업 구도를 정확히 따랐으며, 텍스트 없이 기괴하게 왜곡된 두 사람의 형상을 잘 구현했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "텍스트 금지 지침을 엄격히 준수했으며, 단일 모니터 화면에 온전히 집중하는 클로즈업 구도를 잘 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "주변 모니터 화면에 날짜와 식별 기호 등 텍스트가 노출되어 '텍스트 금지' 규칙을 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "텍스트 금지 지침을 엄격히 준수했으며, 단일 모니터 화면에 온전히 집중하는 클로즈업 구도를 잘 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "주변 모니터 화면에 날짜와 식별 기호 등 텍스트가 노출되어 '텍스트 금지' 규칙을 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S20sh3::variants": {
  "author_fp": "7cd9a58362ff6ae7",
  "author": {
   "variants": [
    {
     "approach_ko": "눈높이의 압축된 얕은 심도로 얼굴과 주먹의 긴장을 밀착해 포착한다.",
     "prompt_en": "Use an intimate eye-level three-quarter medium framing with a gently compressed perspective. Let the soft monitor edge form a narrow, defocused foreground veil, while very shallow focus binds the deeply furrowed brow and clenched hands into the same tense visual plane; allow the office behind her to fall into subdued, desaturated softness. Shape the monitor glow delicately across her brow, cheek contours, knuckles, and fabric folds, preserving restrained shadow detail and natural skin texture."
    },
    {
     "approach_ko": "책상 높이의 비스듬한 시점과 층진 깊이로 아래의 꽉 쥔 손에서 위의 굳은 얼굴까지 불안을 끌어올린다.",
     "prompt_en": "Photograph from a slightly lower, desk-height oblique angle, using layered depth to lead upward from the clenched hands in the lower frame to the furrowed face above. Keep the monitor edge as a soft diagonal foreground anchor and render the office with modestly greater spatial definition, yet hold primary sharpness on her upper body so the surroundings remain secondary. Use the cool screen light as a low directional wash that emphasizes the tension in her fingers and brow against the dim, restrained office tones."
    }
   ]
  },
  "reused": false
 },
 "S20sh3": {
  "input_fingerprint": "385961a3c7335f6a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 불빛에 반사된 수리영의 미간이 깊게 패고 양손을 꽉 쥔 불안한 상반신.\n\nLOCATION (lock): Behind the monitoring desk inside the supermarket security office, facing the glowing CCTV screens in the dim room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral track away from the monitor axis into a close three-quarter upper-body view of 수리영, placing her furrowed face in the upper portion and her clenched hands within the lower portion of the same frame. A soft edge of the monitor remains in the near foreground, but the focus transfer isolates her tightening anxiety rather than changing her established position behind the seated viewers.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: CCTV 모니터 (On and displaying the CCTV footage) — Only part of the display face is visible near the frame edge, turned away enough that its footage is no longer readable; used as Soft foreground edge linking her reaction to the footage she is watching; 마트 총무과 작업 공간 (Occupied during the CCTV review) — The room extends behind 수리영 from the monitor area; used as Provides the administrative-office context behind her without competing for focus.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Monitor light catches 수리영's brow and clenched hands against a dark, desaturated office treatment with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same abnormally dark, unreadable CCTV segment remains on the monitor through repeated playback; Su-riyeong still has her backpack with her.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 불빛에 반사된 수리영의 미간이 깊게 패고 양손을 꽉 쥔 불안한 상반신.\n\nLOCATION (lock): Behind the monitoring desk inside the supermarket security office, facing the glowing CCTV screens in the dim room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral track away from the monitor axis into a close three-quarter upper-body view of 수리영, placing her furrowed face in the upper portion and her clenched hands within the lower portion of the same frame. A soft edge of the monitor remains in the near foreground, but the focus transfer isolates her tightening anxiety rather than changing her established position behind the seated viewers.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: CCTV 모니터 (On and displaying the CCTV footage) — Only part of the display face is visible near the frame edge, turned away enough that its footage is no longer readable; used as Soft foreground edge linking her reaction to the footage she is watching; 마트 총무과 작업 공간 (Occupied during the CCTV review) — The room extends behind 수리영 from the monitor area; used as Provides the administrative-office context behind her without competing for focus.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Monitor light catches 수리영's brow and clenched hands against a dark, desaturated office treatment with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same abnormally dark, unreadable CCTV segment remains on the monitor through repeated playback; Su-riyeong still has her backpack with her.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse an intimate eye-level three-quarter medium framing with a gently compressed perspective. Let the soft monitor edge form a narrow, defocused foreground veil, while very shallow focus binds the deeply furrowed brow and clenched hands into the same tense visual plane; allow the office behind her to fall into subdued, desaturated softness. Shape the monitor glow delicately across her brow, cheek contours, knuckles, and fabric folds, preserving restrained shadow detail and natural skin texture.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 불빛에 반사된 수리영의 미간이 깊게 패고 양손을 꽉 쥔 불안한 상반신.\n\nLOCATION (lock): Behind the monitoring desk inside the supermarket security office, facing the glowing CCTV screens in the dim room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the lateral track away from the monitor axis into a close three-quarter upper-body view of 수리영, placing her furrowed face in the upper portion and her clenched hands within the lower portion of the same frame. A soft edge of the monitor remains in the near foreground, but the focus transfer isolates her tightening anxiety rather than changing her established position behind the seated viewers.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: CCTV 모니터 (On and displaying the CCTV footage) — Only part of the display face is visible near the frame edge, turned away enough that its footage is no longer readable; used as Soft foreground edge linking her reaction to the footage she is watching; 마트 총무과 작업 공간 (Occupied during the CCTV review) — The room extends behind 수리영 from the monitor area; used as Provides the administrative-office context behind her without competing for focus.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Monitor light catches 수리영's brow and clenched hands against a dark, desaturated office treatment with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same abnormally dark, unreadable CCTV segment remains on the monitor through repeated playback; Su-riyeong still has her backpack with her.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph from a slightly lower, desk-height oblique angle, using layered depth to lead upward from the clenched hands in the lower frame to the furrowed face above. Keep the monitor edge as a soft diagonal foreground anchor and render the office with modestly greater spatial definition, yet hold primary sharpness on her upper body so the surroundings remain secondary. Use the cool screen light as a low directional wash that emphasizes the tension in her fingers and brow against the dim, restrained office tones.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 9,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 9,
    "verdict_ko": "캐릭터 레퍼런스의 의상(비니 포함)과 외모를 정확히 재현했으며, 지정된 앵글과 프레이밍, 모니터 배치 및 꽉 쥔 양손의 연출을 훌륭하게 수행했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "지시문에 명시되지 않은 인물이 배경에 등장하는 심각한 위반(발명된 인물)을 범했으며, 캐릭터 레퍼런스의 비니를 누락했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S20sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "모니터 화면이 프레임 가장자리에 일부만 보이고 영상이 읽히지 않을 정도로 돌아가 있어야 한다는 프롬프트 지시와 달리, 모니터가 화면 좌측을 크게 차지하며 안의 CCTV 영상 형태가 명확하게 보입니다.",
     "fix_en": "Angle the monitor away from the camera and shift it further to the left edge so that only a narrow, soft sliver of the bezel is visible and the screen content is completely hidden."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reframe laterally so the monitor becomes only a narrow, defocused diagonal edge at the near frame boundary, with its display face turned farther away and Su-riyeong remaining dominant.",
     "issue_ko": "모니터가 화면 왼쪽의 큰 면적을 차지하고 표시 화면이 정면에 가깝게 보여, 계약의 ‘프레임 가장자리에 남는 부드러운 전경 모서리’보다 지나치게 크고 지배적이다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Angle the monitor away from the camera and shift it further to the left edge so that only a narrow, soft sliver of the bezel is visible and the screen content is completely hidden.\n- Reframe laterally so the monitor becomes only a narrow, defocused diagonal edge at the near frame boundary, with its display face turned farther away and Su-riyeong remaining dominant.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "카메라 시점에서 모니터 화면이 보이지 않게 배치하여 '화면 내용을 읽을 수 없어야 한다'는 지시를 정확히 따랐으며, 인물의 표정과 제스처를 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "모니터 화면이 카메라를 향해 있어 내용이 너무 잘 보이며, 이는 프롬프트의 지시사항 및 사물 방향 규칙에 위배됩니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "카메라 시점에서 모니터 화면이 보이지 않게 배치하여 '화면 내용을 읽을 수 없어야 한다'는 지시를 정확히 따랐으며, 인물의 표정과 제스처를 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "모니터 화면이 카메라를 향해 있어 내용이 너무 잘 보이며, 이는 프롬프트의 지시사항 및 사물 방향 규칙에 위배됩니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "지시문은 모니터 화면의 일부가 보이며 어두운 CCTV 영상이 재생되고 있어야 한다고 명시했으나, 모니터의 뒷면만 노출하여 이를 위반했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "모니터 화면의 일부가 프레임 가장자리에 걸치도록 앵글을 잘 설정했으며, 꽉 쥔 두 손과 찌푸린 미간, 메고 있는 백팩 등 모든 세부 사항을 충실히 반영했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "지시문은 모니터 화면의 일부가 보이며 어두운 CCTV 영상이 재생되고 있어야 한다고 명시했으나, 모니터의 뒷면만 노출하여 이를 위반했습니다."
     },
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "모니터 화면의 일부가 프레임 가장자리에 걸치도록 앵글을 잘 설정했으며, 꽉 쥔 두 손과 찌푸린 미간, 메고 있는 백팩 등 모든 세부 사항을 충실히 반영했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 13
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S20sh2"
  }
 },
 "S21sh6::bgfirst_bg": {
  "input_fingerprint": "d51c08c93a988148",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 사진을 내려다보는 수리영의 극도로 커진 눈동자와 백지장처럼 질린 안색을 담은 고정된 안면 클로즈업.\n\nLOCATION (lock): At the officer’s desk inside the police station office, where enlarged crime-scene photographs are spread across the desktop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Across the near edge of 혜수's desk at seated eye level, finish the dolly-in as a locked extreme facial close-up of 수리영, with her enlarged eyes and drained expression occupying nearly the entire frame. Her gaze remains angled down toward the photographs below frame, while only a blurred lower edge of desk material anchors the source of her shock.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 혜수의 책상 가장자리 (The crime-scene photographs are spread across the desk beyond the visible edge) — Its near edge crosses the bottom of frame between camera and 수리영; used as Blurred lower-frame anchor indicating the photographs she is examining without diluting the facial close-up; 살인사건 확대 사진들 (Spread untidily across the desk) — Their image-bearing faces lie upward on the desk toward 수리영, with only indistinct portions entering the lower edge of the close-up; used as Out-of-focus content below her eyeline that motivates the reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the daytime police office is treated in a low-key, desaturated manner that emphasizes her pallor and widened eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 사진을 내려다보는 수리영의 극도로 커진 눈동자와 백지장처럼 질린 안색을 담은 고정된 안면 클로즈업.\n\nLOCATION (lock): At the officer’s desk inside the police station office, where enlarged crime-scene photographs are spread across the desktop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Across the near edge of 혜수's desk at seated eye level, finish the dolly-in as a locked extreme facial close-up of 수리영, with her enlarged eyes and drained expression occupying nearly the entire frame. Her gaze remains angled down toward the photographs below frame, while only a blurred lower edge of desk material anchors the source of her shock.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 혜수의 책상 가장자리 (The crime-scene photographs are spread across the desk beyond the visible edge) — Its near edge crosses the bottom of frame between camera and 수리영; used as Blurred lower-frame anchor indicating the photographs she is examining without diluting the facial close-up; 살인사건 확대 사진들 (Spread untidily across the desk) — Their image-bearing faces lie upward on the desk toward 수리영, with only indistinct portions entering the lower edge of the close-up; used as Out-of-focus content below her eyeline that motivates the reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the daytime police office is treated in a low-key, desaturated manner that emphasizes her pallor and widened eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S21sh6__bgfirst_bg.png",
  "asset_id": "a6820f11-8576-4d93-a947-715dac51aaf6",
  "input_asset_ids": [
   "577a5a08-7beb-47ce-a62f-89e08d052fb3",
   "73b5302b-b087-4abd-9e28-61a89334ac21"
  ]
 },
 "S21sh6": {
  "input_fingerprint": "78f3c952430193df",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사진을 내려다보는 수리영의 극도로 커진 눈동자와 백지장처럼 질린 안색을 담은 고정된 안면 클로즈업.\n\nLOCATION (lock): At the officer’s desk inside the police station office, where enlarged crime-scene photographs are spread across the desktop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Across the near edge of 혜수's desk at seated eye level, finish the dolly-in as a locked extreme facial close-up of 수리영, with her enlarged eyes and drained expression occupying nearly the entire frame. Her gaze remains angled down toward the photographs below frame, while only a blurred lower edge of desk material anchors the source of her shock.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 혜수의 책상 가장자리 (The crime-scene photographs are spread across the desk beyond the visible edge) — Its near edge crosses the bottom of frame between camera and 수리영; used as Blurred lower-frame anchor indicating the photographs she is examining without diluting the facial close-up; 살인사건 확대 사진들 (Spread untidily across the desk) — Their image-bearing faces lie upward on the desk toward 수리영, with only indistinct portions entering the lower edge of the close-up; used as Out-of-focus content below her eyeline that motivates the reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the daytime police office is treated in a low-key, desaturated manner that emphasizes her pallor and widened eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gruesome enlarged crime-scene photographs remain spread across Hye-su’s desk, while Su-riyeong retains her doll-adorned backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사진을 내려다보는 수리영의 극도로 커진 눈동자와 백지장처럼 질린 안색을 담은 고정된 안면 클로즈업.\n\nLOCATION (lock): At the officer’s desk inside the police station office, where enlarged crime-scene photographs are spread across the desktop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Across the near edge of 혜수's desk at seated eye level, finish the dolly-in as a locked extreme facial close-up of 수리영, with her enlarged eyes and drained expression occupying nearly the entire frame. Her gaze remains angled down toward the photographs below frame, while only a blurred lower edge of desk material anchors the source of her shock.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 혜수의 책상 가장자리 (The crime-scene photographs are spread across the desk beyond the visible edge) — Its near edge crosses the bottom of frame between camera and 수리영; used as Blurred lower-frame anchor indicating the photographs she is examining without diluting the facial close-up; 살인사건 확대 사진들 (Spread untidily across the desk) — Their image-bearing faces lie upward on the desk toward 수리영, with only indistinct portions entering the lower edge of the close-up; used as Out-of-focus content below her eyeline that motivates the reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the daytime police office is treated in a low-key, desaturated manner that emphasizes her pallor and widened eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gruesome enlarged crime-scene photographs remain spread across Hye-su’s desk, while Su-riyeong retains her doll-adorned backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사진을 내려다보는 수리영의 극도로 커진 눈동자와 백지장처럼 질린 안색을 담은 고정된 안면 클로즈업.\n\nLOCATION (lock): At the officer’s desk inside the police station office, where enlarged crime-scene photographs are spread across the desktop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Across the near edge of 혜수's desk at seated eye level, finish the dolly-in as a locked extreme facial close-up of 수리영, with her enlarged eyes and drained expression occupying nearly the entire frame. Her gaze remains angled down toward the photographs below frame, while only a blurred lower edge of desk material anchors the source of her shock.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 혜수의 책상 가장자리 (The crime-scene photographs are spread across the desk beyond the visible edge) — Its near edge crosses the bottom of frame between camera and 수리영; used as Blurred lower-frame anchor indicating the photographs she is examining without diluting the facial close-up; 살인사건 확대 사진들 (Spread untidily across the desk) — Their image-bearing faces lie upward on the desk toward 수리영, with only indistinct portions entering the lower edge of the close-up; used as Out-of-focus content below her eyeline that motivates the reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the daytime police office is treated in a low-key, desaturated manner that emphasizes her pallor and widened eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The gruesome enlarged crime-scene photographs remain spread across Hye-su’s desk, while Su-riyeong retains her doll-adorned backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S21sh6__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S21sh6.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L13B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 극단적인 안면 클로즈업 앵글과 흐릿하게 처리된 책상 가장자리 및 사진 구도를 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "얼굴이 화면을 가득 채워야 하는 안면 클로즈업 지시를 어기고 상반신과 배경을 너무 넓게 포착했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 극단적인 안면 클로즈업 앵글과 흐릿하게 처리된 책상 가장자리 및 사진 구도를 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "얼굴이 화면을 가득 채워야 하는 안면 클로즈업 지시를 어기고 상반신과 배경을 너무 넓게 포착했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 '극단적인 안면 클로즈업' 앵글과 화면 비율을 완벽히 구현했으며, 흐릿한 사진과 책상 가장자리 배치도 정확합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 안면 클로즈업을 무시하고 앵글을 넓혀 상반신과 초점이 맞는 사진들을 노출함으로써 프레임 지침을 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 '극단적인 안면 클로즈업' 앵글과 화면 비율을 완벽히 구현했으며, 흐릿한 사진과 책상 가장자리 배치도 정확합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "지정된 안면 클로즈업을 무시하고 앵글을 넓혀 상반신과 초점이 맞는 사진들을 노출함으로써 프레임 지침을 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 8
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 15,
   "B": 8
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 극단적인 안면 클로즈업 앵글과 흐릿하게 처리된 책상 가장자리 및 사진 구도를 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "얼굴이 화면을 가득 채워야 하는 안면 클로즈업 지시를 어기고 상반신과 배경을 너무 넓게 포착했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L13B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영의 양쪽 눈동자(홍채와 동공) 형태가 일그러져 있어 해부학적으로 정상적인 사람의 눈이 아닙니다.",
     "fix_en": "Restore anatomically normal, perfectly round irises and solid pupils to both of Su-riyeong's eyes."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Dolly in tighter so Su-riyeong’s face occupies nearly the entire frame, leaving only a thin, blurred sliver of the desk edge and photographs along the bottom.",
     "issue_ko": "얼굴보다 어깨와 사무실 배경, 책상 위 사진들이 너무 넓게 보여 계약의 극단적 안면 클로즈업보다 프레이밍이 느슨하다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Restore anatomically normal, perfectly round irises and solid pupils to both of Su-riyeong's eyes.\n- Dolly in tighter so Su-riyeong’s face occupies nearly the entire frame, leaving only a thin, blurred sliver of the desk edge and photographs along the bottom.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "레이아웃 스케치와 프레이밍 지시사항(얼굴이 화면을 거의 채우는 익스트림 클로즈업)을 완벽하게 구현했으며 창백한 안색과 커진 눈동자의 디테일도 우수합니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "표정과 소품의 방향은 좋으나, 카메라 샷이 너무 넓게 잡혀 레이아웃 스케치와 '극단적인 안면 클로즈업'이라는 프레이밍 지시를 어겼습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "레이아웃 스케치와 프레이밍 지시사항(얼굴이 화면을 거의 채우는 익스트림 클로즈업)을 완벽하게 구현했으며 창백한 안색과 커진 눈동자의 디테일도 우수합니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "표정과 소품의 방향은 좋으나, 카메라 샷이 너무 넓게 잡혀 레이아웃 스케치와 '극단적인 안면 클로즈업'이라는 프레이밍 지시를 어겼습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "스케치에 제시된 극단적인 안면 클로즈업 구도를 정확하게 구현하였으며, 피사체의 시선과 창백한 안색 등 지시사항을 훌륭히 따랐습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 극단적인 안면 클로즈업에 비해 프레임이 너무 넓게 설정되었으며, 눈동자 방향이 부자연스럽습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "스케치에 제시된 극단적인 안면 클로즈업 구도를 정확하게 구현하였으며, 피사체의 시선과 창백한 안색 등 지시사항을 훌륭히 따랐습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "요구된 극단적인 안면 클로즈업에 비해 프레임이 너무 넓게 설정되었으며, 눈동자 방향이 부자연스럽습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 16
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S21sh6__bgfirst_bg.png",
   "bg_asset_id": "a6820f11-8576-4d93-a947-715dac51aaf6",
   "bg_record_key": "S21sh6::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L13B01",
    "B": "L13B02"
   },
   "assigned": "L13B01",
   "choice": "Candidate A",
   "confident": true,
   "reason_ko": "샷은 사진을 내려다보는 인물의 안면 클로즈업(고정된 안면 클로즈업)을 묘사하고 있으므로, 특정 서브공간을 명확히 구분할 만한 단서가 없어 현재 할당된 후보를 유지합니다.",
   "kept": "L13B01"
  }
 },
 "S21sh10::variants": {
  "author_fp": "ae3fc8280b92514a",
  "author": {
   "variants": [
    {
     "approach_ko": "왜곡 없는 정면 관찰과 균일한 초점으로 추적 결과를 차갑고 증거물처럼 선명하게 포착한다.",
     "prompt_en": "Use a neutral, rectilinear lens feel with edge-to-edge focus and disciplined, forensic clarity. Expose the display so the maritime detail and tracking information remain cleanly readable without blooming, while the narrow physical margins retain subdued texture and cool office shadow. Keep the composition rigid, centered, and clinically still."
    },
    {
     "approach_ko": "주변부를 어둠에 가라앉히고 화면의 절제된 붉은 강조를 살려 좌표 발견의 불길한 충격을 부각한다.",
     "prompt_en": "Photograph the screen as an ominous pool of controlled light within the low-key office, allowing the bezel and workstation margin to fall toward deep desaturated shadow. Preserve full readability across the display, but emphasize the restrained crimson point through local contrast and subtle screen luminance rather than saturation or glow. Hold the image tense and motionless, with the physical edges gently receding from the luminous map."
    }
   ]
  },
  "reused": false
 },
 "S21sh10": {
  "input_fingerprint": "ee1b42d7a07d9b3d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 해상 지도 이미지 위로 붉은색 GPS 좌표 점이 찍혀 있는 모니터 화면 클로즈업.\n\nLOCATION (lock): At the detective’s computer workstation inside the police station office, focused on the location-tracking result displayed on the monitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold directly perpendicular to the monitor at close range, with the maritime map centered and the GPS point, coordinates, timestamp, and offline status all legible inside the screen. Keep a narrow strip of bezel and workstation edge around the display so the result remains physically grounded while the map occupies most of the visible screen area.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Kim Detective's monitor (On, displaying Kang Min-sook's last phone location) — The active front face is square to camera, showing the maritime map, red GPS point, coordinates, timestamp, source information, and offline status; used as Primary focal surface carrying the location-tracking result; workstation edge (Visible only at the frame margin) — Its front edge runs beneath the screen, nearly parallel to the lower frame; used as Thin peripheral context beneath the monitor.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The active display provides restrained crimson emphasis within a dark, desaturated office palette and low-key tonal contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The computer displays Minsuk’s last phone location as an offshore GPS coordinate; Su-riyeong still has her backpack nearby.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 해상 지도 이미지 위로 붉은색 GPS 좌표 점이 찍혀 있는 모니터 화면 클로즈업.\n\nLOCATION (lock): At the detective’s computer workstation inside the police station office, focused on the location-tracking result displayed on the monitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold directly perpendicular to the monitor at close range, with the maritime map centered and the GPS point, coordinates, timestamp, and offline status all legible inside the screen. Keep a narrow strip of bezel and workstation edge around the display so the result remains physically grounded while the map occupies most of the visible screen area.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Kim Detective's monitor (On, displaying Kang Min-sook's last phone location) — The active front face is square to camera, showing the maritime map, red GPS point, coordinates, timestamp, source information, and offline status; used as Primary focal surface carrying the location-tracking result; workstation edge (Visible only at the frame margin) — Its front edge runs beneath the screen, nearly parallel to the lower frame; used as Thin peripheral context beneath the monitor.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The active display provides restrained crimson emphasis within a dark, desaturated office palette and low-key tonal contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The computer displays Minsuk’s last phone location as an offshore GPS coordinate; Su-riyeong still has her backpack nearby.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a neutral, rectilinear lens feel with edge-to-edge focus and disciplined, forensic clarity. Expose the display so the maritime detail and tracking information remain cleanly readable without blooming, while the narrow physical margins retain subdued texture and cool office shadow. Keep the composition rigid, centered, and clinically still.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거친 해상 지도 이미지 위로 붉은색 GPS 좌표 점이 찍혀 있는 모니터 화면 클로즈업.\n\nLOCATION (lock): At the detective’s computer workstation inside the police station office, focused on the location-tracking result displayed on the monitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold directly perpendicular to the monitor at close range, with the maritime map centered and the GPS point, coordinates, timestamp, and offline status all legible inside the screen. Keep a narrow strip of bezel and workstation edge around the display so the result remains physically grounded while the map occupies most of the visible screen area.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Kim Detective's monitor (On, displaying Kang Min-sook's last phone location) — The active front face is square to camera, showing the maritime map, red GPS point, coordinates, timestamp, source information, and offline status; used as Primary focal surface carrying the location-tracking result; workstation edge (Visible only at the frame margin) — Its front edge runs beneath the screen, nearly parallel to the lower frame; used as Thin peripheral context beneath the monitor.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The active display provides restrained crimson emphasis within a dark, desaturated office palette and low-key tonal contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The computer displays Minsuk’s last phone location as an offshore GPS coordinate; Su-riyeong still has her backpack nearby.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph the screen as an ominous pool of controlled light within the low-key office, allowing the bezel and workstation margin to fall toward deep desaturated shadow. Preserve full readability across the display, but emphasize the restrained crimson point through local contrast and subtle screen luminance rather than saturation or glow. Hold the image tense and motionless, with the physical edges gently receding from the luminous map.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "요구된 클로즈업 구도와 얇은 베젤, 화면 마진의 책상 모서리를 정확히 구현했으며 한국어 기반 UI를 잘 반영했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "모니터 주변의 책상 표면과 마우스, 펜 등 불필요한 사물이 너무 많이 보이며 프레임이 넓고 UI가 영문으로 표기되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S21sh6_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 패널에 프롬프트가 명시적으로 요구한 '출처 정보(source information)' 텍스트가 누락되어 있습니다.",
     "fix_en": "Add the missing source information text to the side panel, grouping it with the coordinates and timestamp."
    },
    {
     "issue_ko": "프롬프트는 '붉은색 GPS 좌표 점(point)'을 요구했으나, 단순한 점이 아닌 입체적인 3D 맵 핀 아이콘이 그려져 있습니다.",
     "fix_en": "Replace the 3D red map pin icon with a simple, flat red dot to strictly match the 'point' instruction."
    },
    {
     "issue_ko": "레퍼런스 이미지에서 확립된 책상은 밝은 갈색 목재이나, 현재 샷에서 모니터 아래로 보이는 책상 모서리는 어두운 회색이어서 일관성이 어긋납니다.",
     "fix_en": "Change the visible desk edge beneath the monitor to a light brown wood material to match the reference image exactly."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Add the missing source information text to the side panel, grouping it with the coordinates and timestamp.\n- Replace the 3D red map pin icon with a simple, flat red dot to strictly match the 'point' instruction.\n- Change the visible desk edge beneath the monitor to a light brown wood material to match the reference image exactly.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 프레이밍을 정확히 구현했으며, 프롬프트에서 요구한 출처 정보(source information)와 책상 가장자리 디테일이 잘 표현되었습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 화면 구성은 잘 따랐으나, 모니터 화면 내에 표시되어야 할 '출처 정보(source information)'가 누락되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 프레이밍을 정확히 구현했으며, 프롬프트에서 요구한 출처 정보(source information)와 책상 가장자리 디테일이 잘 표현되었습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 화면 구성은 잘 따랐으나, 모니터 화면 내에 표시되어야 할 '출처 정보(source information)'가 누락되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 프레이밍을 정확히 따랐으며, 레퍼런스의 나무 책상 질감을 유지하고 '붉은색 점'을 명확히 묘사하여 요구사항을 충실히 반영했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "전반적인 화면 구성은 우수하나 붉은색 점 대신 핀 아이콘이 생성되었고, 레퍼런스의 책상 재질(나무)이 프레임 가장자리에 일관되게 나타나지 않았습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 프레이밍을 정확히 따랐으며, 레퍼런스의 나무 책상 질감을 유지하고 '붉은색 점'을 명확히 묘사하여 요구사항을 충실히 반영했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "전반적인 화면 구성은 우수하나 붉은색 점 대신 핀 아이콘이 생성되었고, 레퍼런스의 책상 재질(나무)이 프레임 가장자리에 일관되게 나타나지 않았습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S21sh6"
  },
  "lane_policy": "share_plan_prev_bgonly",
  "plate_select": {
   "candidates": {
    "A": "L13B01",
    "B": "L13B02"
   },
   "assigned": "L13B01",
   "choice": "A",
   "confident": true,
   "reason_ko": "모니터 화면",
   "chosen": "L13B01"
  }
 },
 "S21sh12::variants": {
  "author_fp": "81286756190faf9a",
  "author": {
   "variants": [
    {
     "approach_ko": "깊은 초점과 열린 동선의 기하학으로 멈칫한 출발 순간을 냉정하게 포착한다.",
     "prompt_en": "Use a restrained, observational medium-shot lens feel with deep, legible spatial focus. Build a strong diagonal from the near-edge workstation through Su-riyeong to the office door, preserving the open exit route as deliberate negative space; keep the exposed photographs readable as physical desk clutter without letting them dominate. Render the daytime office ambience as broad, subdued illumination with gentle falloff, cool desaturated surfaces, and crisp detail in her weight shift, turning torso, and hand just clearing the pocket."
    },
    {
     "approach_ko": "얕은 심도와 압축된 공간감으로 떠나려는 몸의 찰나와 심리적 망설임에 집중한다.",
     "prompt_en": "Give the static medium view a subtly compressed, intimate lens feel and isolate Su-riyeong with selective focus. Let the workstation edge and scattered photographs form a soft, uneasy foreground layer while her mid-stride body, backpack, and departing hand remain sharply resolved; hold the office door as a recognizable but gently softened destination behind her half-turn. Keep the ambient office light low-key and desaturated, shaping restrained highlights along her hair and clothing while the surrounding room recedes into muted tonal density."
    }
   ]
  },
  "reused": false
 },
 "S21sh12": {
  "input_fingerprint": "d5a73d03e54828ec",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 종이 쪽지를 주머니에 넣은 채, 등 뒤 사무실 문을 향해 발을 내딛고 상체를 반쯤 돌린 수리영의 mid-action 정지 구도.\n\nLOCATION (lock): Between the detective’s workstation and the office exit inside the police station, as the daughter turns toward the door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From waist-to-chest height beside the workstation, hold a static medium view as 수리영 pauses mid-stride in the center-right, one hand just leaving her pocket and her upper body half-turned toward the office door behind her. Keep the workstation along the near edge and the exit route open across the opposite side of frame, emphasizing only her changed position toward departure.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, midground, moves toward office door; office door in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: office door (Visible on her exit path) — The doorway is seen beyond and to the side of her half-turned body; used as Exit anchor behind 수리영 and destination of her arrested turn; workstation (The location-result monitor remains on) — Its side faces camera while the monitor is angled away from the lens; used as Foreground spatial anchor beside the camera; enlarged murder photographs (Scattered across Hye-su's desk) — Their image-bearing faces lie exposed across the desk, partially overlapping one another; used as Uneasy context around the workstation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient office light is rendered in a dark, desaturated grade with restrained contrast and no added color source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The handwritten coordinates remain folded in Su-riyeong’s pocket, and her doll-adorned backpack remains with her as she leaves the office.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 종이 쪽지를 주머니에 넣은 채, 등 뒤 사무실 문을 향해 발을 내딛고 상체를 반쯤 돌린 수리영의 mid-action 정지 구도.\n\nLOCATION (lock): Between the detective’s workstation and the office exit inside the police station, as the daughter turns toward the door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From waist-to-chest height beside the workstation, hold a static medium view as 수리영 pauses mid-stride in the center-right, one hand just leaving her pocket and her upper body half-turned toward the office door behind her. Keep the workstation along the near edge and the exit route open across the opposite side of frame, emphasizing only her changed position toward departure.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, midground, moves toward office door; office door in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: office door (Visible on her exit path) — The doorway is seen beyond and to the side of her half-turned body; used as Exit anchor behind 수리영 and destination of her arrested turn; workstation (The location-result monitor remains on) — Its side faces camera while the monitor is angled away from the lens; used as Foreground spatial anchor beside the camera; enlarged murder photographs (Scattered across Hye-su's desk) — Their image-bearing faces lie exposed across the desk, partially overlapping one another; used as Uneasy context around the workstation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient office light is rendered in a dark, desaturated grade with restrained contrast and no added color source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The handwritten coordinates remain folded in Su-riyeong’s pocket, and her doll-adorned backpack remains with her as she leaves the office.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained, observational medium-shot lens feel with deep, legible spatial focus. Build a strong diagonal from the near-edge workstation through Su-riyeong to the office door, preserving the open exit route as deliberate negative space; keep the exposed photographs readable as physical desk clutter without letting them dominate. Render the daytime office ambience as broad, subdued illumination with gentle falloff, cool desaturated surfaces, and crisp detail in her weight shift, turning torso, and hand just clearing the pocket.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 종이 쪽지를 주머니에 넣은 채, 등 뒤 사무실 문을 향해 발을 내딛고 상체를 반쯤 돌린 수리영의 mid-action 정지 구도.\n\nLOCATION (lock): Between the detective’s workstation and the office exit inside the police station, as the daughter turns toward the door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From waist-to-chest height beside the workstation, hold a static medium view as 수리영 pauses mid-stride in the center-right, one hand just leaving her pocket and her upper body half-turned toward the office door behind her. Keep the workstation along the near edge and the exit route open across the opposite side of frame, emphasizing only her changed position toward departure.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-right of the frame, midground, moves toward office door; office door in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: office door (Visible on her exit path) — The doorway is seen beyond and to the side of her half-turned body; used as Exit anchor behind 수리영 and destination of her arrested turn; workstation (The location-result monitor remains on) — Its side faces camera while the monitor is angled away from the lens; used as Foreground spatial anchor beside the camera; enlarged murder photographs (Scattered across Hye-su's desk) — Their image-bearing faces lie exposed across the desk, partially overlapping one another; used as Uneasy context around the workstation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient office light is rendered in a dark, desaturated grade with restrained contrast and no added color source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The handwritten coordinates remain folded in Su-riyeong’s pocket, and her doll-adorned backpack remains with her as she leaves the office.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nGive the static medium view a subtly compressed, intimate lens feel and isolate Su-riyeong with selective focus. Let the workstation edge and scattered photographs form a soft, uneasy foreground layer while her mid-stride body, backpack, and departing hand remain sharply resolved; hold the office door as a recognizable but gently softened destination behind her half-turn. Keep the ambient office light low-key and desaturated, shaping restrained highlights along her hair and clothing while the surrounding room recedes into muted tonal density.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 6,
   "A": 5
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "요청된 미디엄 샷 프레이밍에 더 근접하여 높은 우선순위(2)를 충족했으나, 쪽지를 주머니가 아닌 손에 들고 있어 소지품 상태(5)에서 감점됨."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "동선과 소지품 상태는 비교적 잘 반영되었으나, 지시된 미디엄 샷이 아닌 전신 풀 샷으로 프레이밍(우선순위 2)이 과도하게 넓어 감점됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S21sh10_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영이 종이 쪽지를 왼손에 쥐고 있으며, 프롬프트의 지시대로 주머니에 넣은 상태가 아닙니다.",
     "fix_en": "Remove the paper slip from her left hand so it is empty."
    },
    {
     "issue_ko": "배경 벽면에 부착된 포스터와 서류들에 텍스트가 포함되어 있어 텍스트 금지 규칙을 위반했습니다.",
     "fix_en": "Remove all text from the posters and documents on the background wall."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reverse Su-riyeong’s stepping direction so her lower body moves screen-left toward the office door while preserving her half-turned torso and screen position.",
     "issue_ko": "수리영의 발걸음이 중간 왼쪽의 사무실 문이 아니라 화면 오른쪽으로 향해, 계약된 퇴장 동선과 반대입니다."
    },
    {
     "fix_en": "Place her pocket-side hand just outside the pocket in a clear withdrawing motion, preserving the rest of her pose.",
     "issue_ko": "한 손이 주머니에서 막 빠져나오는 순간이어야 하지만, 현재는 손이 여전히 주머니 안에 깊이 들어가 있습니다."
    },
    {
     "fix_en": "Rotate the foreground monitor farther away from the camera so its side is emphasized and the screen is only obliquely visible.",
     "issue_ko": "전경 모니터 화면이 카메라를 향해 크게 노출되어 있으며, 모니터가 렌즈에서 비껴 돌아가 있어야 한다는 배치와 다릅니다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the paper slip from her left hand so it is empty.\n- Remove all text from the posters and documents on the background wall.\n- Reverse Su-riyeong’s stepping direction so her lower body moves screen-left toward the office door while preserving her half-turned torso and screen position.\n- Place her pocket-side hand just outside the pocket in a clear withdrawing motion, preserving the rest of her pose.\n- Rotate the foreground monitor farther away from the camera so its side is emphasized and the screen is only obliquely visible.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "종이 쪽지가 주머니에 있지 않고 손에 들려 있는 점은 아쉽지만, 모니터 화면과 지정된 구도를 매우 충실하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "모니터 화면이 어둡게 처리되었으며, 인물의 왼쪽 손가락 형태가 심하게 왜곡된 해부학적 오류(하드 위반)가 있습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "종이 쪽지가 주머니에 있지 않고 손에 들려 있는 점은 아쉽지만, 모니터 화면과 지정된 구도를 매우 충실하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "모니터 화면이 어둡게 처리되었으며, 인물의 왼쪽 손가락 형태가 심하게 왜곡된 해부학적 오류(하드 위반)가 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "왼쪽 손가락의 묘사가 기형적으로 무너져 물리적으로 불가능한 신체(우선순위 1위 위반)가 발생했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "쪽지를 주머니에 넣지 않고 손에 들고 있어 세부 지시와 다르나, 심각한 해부학적 오류 없이 샷 텍스트의 구도를 잘 따랐습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "왼쪽 손가락의 묘사가 기형적으로 무너져 물리적으로 불가능한 신체(우선순위 1위 위반)가 발생했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "쪽지를 주머니에 넣지 않고 손에 들고 있어 세부 지시와 다르나, 심각한 해부학적 오류 없이 샷 텍스트의 구도를 잘 따랐습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S21sh10"
  }
 },
 "groupbg::피시방_좌석": {
  "input_fingerprint": "a8770e6489930ff7",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "피시방_좌석",
    "tags": [
     "S22sh2",
     "S22sh3"
    ]
   },
   "context_sig": "ba37f59896c7b9e3"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\nPC방 내부: 어두운 실내에 화려한 조명과 대형 모니터들이 줄지어 있는 현대 한국의 일반적인 PC방 좌석 구역. (특징: 열을 맞춰 놓인 컴퓨터 데스크와 게이밍 의자; 디지털 지도가 띄워진 대형 모니터 화면; 어두운 공간 속 모니터 백라이트와 실내 간접 조명)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#23. 어느 PC방 / 실내\n- 모니터 화면에 앉아있는 수리영, 좌표를 찍고 위치를 확인하고 인쇄까지\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\nPC방 내부: 어두운 실내에 화려한 조명과 대형 모니터들이 줄지어 있는 현대 한국의 일반적인 PC방 좌석 구역. (특징: 열을 맞춰 놓인 컴퓨터 데스크와 게이밍 의자; 디지털 지도가 띄워진 대형 모니터 화면; 어두운 공간 속 모니터 백라이트와 실내 간접 조명)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#23. 어느 PC방 / 실내\n- 모니터 화면에 앉아있는 수리영, 좌표를 찍고 위치를 확인하고 인쇄까지\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_피시방_좌석_4632d2.png",
  "asset_id": "53b4b393-27b7-4db3-93a6-03d24d4b989c",
  "input_asset_ids": [
   "3d45218c-6ed0-4b68-9ce0-10a5f98133d1"
  ],
  "origin_tag": "S22sh2",
  "place_text": "At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow.",
  "origin_inputs": {
   "place_text": "At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow.",
   "time_of_day_en": "day",
   "conti_asset_id": "3d45218c-6ed0-4b68-9ce0-10a5f98133d1"
  }
 },
 "S22sh2::bgfirst_bg": {
  "input_fingerprint": "43a154aa528dcd6c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어둡고 푸르스름한 조명이 깔린 PC방 내부, 빛나는 모니터 앞에 앉은 수리영의 측면 상반신.\n\nLOCATION (lock): At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at seated shoulder height beside 수리영 at medium distance, perpendicular to the monitor, and dolly inward on her profile as she leans toward the workstation. Place her upper body in the left half and the glowing monitor near the right edge, allowing the screen axis to pull the camera forward toward the following reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: PC monitor (On, displaying the entered location coordinates) — The active screen face is seen obliquely from the side, with the map interface only partly readable; used as Bright visual destination at the right edge and motivation for the dolly-in; workstation (In use) — Its side plane recedes from 수리영 toward the monitor; used as Supports 수리영 and establishes the PC workstation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark bluish PC-room illumination and the glowing monitor model 수리영 in a desaturated, low-key profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어둡고 푸르스름한 조명이 깔린 PC방 내부, 빛나는 모니터 앞에 앉은 수리영의 측면 상반신.\n\nLOCATION (lock): At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at seated shoulder height beside 수리영 at medium distance, perpendicular to the monitor, and dolly inward on her profile as she leans toward the workstation. Place her upper body in the left half and the glowing monitor near the right edge, allowing the screen axis to pull the camera forward toward the following reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: PC monitor (On, displaying the entered location coordinates) — The active screen face is seen obliquely from the side, with the map interface only partly readable; used as Bright visual destination at the right edge and motivation for the dolly-in; workstation (In use) — Its side plane recedes from 수리영 toward the monitor; used as Supports 수리영 and establishes the PC workstation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark bluish PC-room illumination and the glowing monitor model 수리영 in a desaturated, low-key profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S22sh2__bgfirst_bg.png",
  "asset_id": "1cafaf94-5900-45e3-bc69-8cd7762ecc30",
  "input_asset_ids": [
   "3d45218c-6ed0-4b68-9ce0-10a5f98133d1",
   "53b4b393-27b7-4db3-93a6-03d24d4b989c"
  ]
 },
 "S22sh2": {
  "input_fingerprint": "0beb813a9caa63a4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둡고 푸르스름한 조명이 깔린 PC방 내부, 빛나는 모니터 앞에 앉은 수리영의 측면 상반신.\n\nLOCATION (lock): At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at seated shoulder height beside 수리영 at medium distance, perpendicular to the monitor, and dolly inward on her profile as she leans toward the workstation. Place her upper body in the left half and the glowing monitor near the right edge, allowing the screen axis to pull the camera forward toward the following reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: PC monitor (On, displaying the entered location coordinates) — The active screen face is seen obliquely from the side, with the map interface only partly readable; used as Bright visual destination at the right edge and motivation for the dolly-in; workstation (In use) — Its side plane recedes from 수리영 toward the monitor; used as Supports 수리영 and establishes the PC workstation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark bluish PC-room illumination and the glowing monitor model 수리영 in a desaturated, low-key profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong still carries the folded coordinate note in her pocket and keeps her backpack ready beside her for the printed maps.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둡고 푸르스름한 조명이 깔린 PC방 내부, 빛나는 모니터 앞에 앉은 수리영의 측면 상반신.\n\nLOCATION (lock): At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at seated shoulder height beside 수리영 at medium distance, perpendicular to the monitor, and dolly inward on her profile as she leans toward the workstation. Place her upper body in the left half and the glowing monitor near the right edge, allowing the screen axis to pull the camera forward toward the following reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: PC monitor (On, displaying the entered location coordinates) — The active screen face is seen obliquely from the side, with the map interface only partly readable; used as Bright visual destination at the right edge and motivation for the dolly-in; workstation (In use) — Its side plane recedes from 수리영 toward the monitor; used as Supports 수리영 and establishes the PC workstation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark bluish PC-room illumination and the glowing monitor model 수리영 in a desaturated, low-key profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong still carries the folded coordinate note in her pocket and keeps her backpack ready beside her for the printed maps.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어둡고 푸르스름한 조명이 깔린 PC방 내부, 빛나는 모니터 앞에 앉은 수리영의 측면 상반신.\n\nLOCATION (lock): At a computer booth inside a dim PC café, illuminated by cool blue ambient light and the monitor glow. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin at seated shoulder height beside 수리영 at medium distance, perpendicular to the monitor, and dolly inward on her profile as she leans toward the workstation. Place her upper body in the left half and the glowing monitor near the right edge, allowing the screen axis to pull the camera forward toward the following reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: PC monitor (On, displaying the entered location coordinates) — The active screen face is seen obliquely from the side, with the map interface only partly readable; used as Bright visual destination at the right edge and motivation for the dolly-in; workstation (In use) — Its side plane recedes from 수리영 toward the monitor; used as Supports 수리영 and establishes the PC workstation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dark bluish PC-room illumination and the glowing monitor model 수리영 in a desaturated, low-key profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong still carries the folded coordinate note in her pocket and keeps her backpack ready beside her for the printed maps.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S22sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S22sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_피시방_좌석_4632d2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "프레이밍과 PC방의 세부 묘사, 가방 배치를 잘 따랐으나 명시되지 않은 세로형 스크린이 발현된 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "허공에 떠 있는 그래픽 텍스트 오버레이가 생성되어 물질적 사실성 지침을 크게 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "프레이밍과 PC방의 세부 묘사, 가방 배치를 잘 따랐으나 명시되지 않은 세로형 스크린이 발현된 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "허공에 떠 있는 그래픽 텍스트 오버레이가 생성되어 물질적 사실성 지침을 크게 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 측면 구도와 푸른 조명 분위기를 잘 구현했으나, 유리 칸막이에 프롬프트에 없는 허공의 텍스트가 생성된 점이 감점 요소입니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "지시되지 않은 배경 인물이 등장해 치명적인 위반이며, 메인 모니터 앞에 물리적으로 어색한 패널이 겹쳐 있어 프레임 구성이 훼손되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지정된 측면 구도와 푸른 조명 분위기를 잘 구현했으나, 유리 칸막이에 프롬프트에 없는 허공의 텍스트가 생성된 점이 감점 요소입니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시되지 않은 배경 인물이 등장해 치명적인 위반이며, 메인 모니터 앞에 물리적으로 어색한 패널이 겹쳐 있어 프레임 구성이 훼손되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 10
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "totals": {
   "A": 10,
   "B": 10
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "프레이밍과 PC방의 세부 묘사, 가방 배치를 잘 따랐으나 명시되지 않은 세로형 스크린이 발현된 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "허공에 떠 있는 그래픽 텍스트 오버레이가 생성되어 물질적 사실성 지침을 크게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_피시방_좌석_4632d2.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "샷 텍스트에 언급되지 않은 인물이 배경에 앉아 있습니다.",
     "fix_en": "Remove the blurred person sitting in the background aisle."
    },
    {
     "issue_ko": "오른쪽 원본 모니터 앞에 평면적인 하얀 화면이 임의로 추가되었습니다.",
     "fix_en": "Remove the flat white screen added in front of the right-hand monitor to restore the original background."
    },
    {
     "issue_ko": "캐릭터가 레퍼런스 이미지에 있는 짙은 회색 비니를 쓰지 않았습니다.",
     "fix_en": "Add the dark grey beanie to the character's head to match the reference."
    },
    {
     "issue_ko": "백팩 주머니에 있는 종이에 금지된 텍스트가 포함되어 있습니다.",
     "fix_en": "Remove all text from the folded paper inside the backpack pocket."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Remove the distant seated person in the background while preserving Su-riyeong and the entire frame composition unchanged.",
     "issue_ko": "화면 중앙 깊은 배경에 계약에 없는 인물이 추가되어 있어, 수리영만 보여야 하는 인물 배치를 위반한다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the blurred person sitting in the background aisle.\n- Remove the flat white screen added in front of the right-hand monitor to restore the original background.\n- Add the dark grey beanie to the character's head to match the reference.\n- Remove all text from the folded paper inside the backpack pocket.\n- Remove the distant seated person in the background while preserving Su-riyeong and the entire frame composition unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정되지 않은 배경 인물을 추가하고 원본 배경에 없는 두 번째 모니터를 창조해 냈으며, 레퍼런스의 비니를 누락하여 심각한 위반에 해당합니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "지정된 배경을 정확히 유지하면서, 요구된 카메라 구도, 푸른 조명, 비니를 포함한 인물의 외모와 소품 위치까지 완벽하게 재현했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정되지 않은 배경 인물을 추가하고 원본 배경에 없는 두 번째 모니터를 창조해 냈으며, 레퍼런스의 비니를 누락하여 심각한 위반에 해당합니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "지정된 배경을 정확히 유지하면서, 요구된 카메라 구도, 푸른 조명, 비니를 포함한 인물의 외모와 소품 위치까지 완벽하게 재현했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "레퍼런스 배경과 레이아웃을 정확히 유지했으며 캐릭터의 외모와 의상(비니 포함)을 충실히 반영했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "화면 우측에 허공에 떠 있는 모니터가 생성되는 치명적인 오류가 발생했으며, 캐릭터의 비니가 누락되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "레퍼런스 배경과 레이아웃을 정확히 유지했으며 캐릭터의 외모와 의상(비니 포함)을 충실히 반영했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "화면 우측에 허공에 떠 있는 모니터가 생성되는 치명적인 오류가 발생했으며, 캐릭터의 비니가 누락되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 17
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S22sh2__bgfirst_bg.png",
   "bg_asset_id": "1cafaf94-5900-45e3-bc69-8cd7762ecc30",
   "bg_record_key": "S22sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "피시방_좌석",
   "groupbg_asset_id": "53b4b393-27b7-4db3-93a6-03d24d4b989c"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S22sh3::variants": {
  "author_fp": "e630898f32e04174",
  "author": {
   "variants": [
    {
     "approach_ko": "정면 대칭과 또렷한 화면 정보로 좌표 발견의 명료함을 강조하는 절제된 접근.",
     "prompt_en": "Use rigid frontal symmetry and clean, controlled focus, with the coastline geometry, offshore position, and muted red coordinate marker rendered crisply across the display. Keep the narrow bezel evenly visible on all sides, letting the cool desaturated room ambience recede while the monitor becomes a precise, restrained reveal."
    },
    {
     "approach_ko": "화면의 발광감과 물리적 모니터 질감을 살려 불길한 좌표의 긴장감을 강조하는 접근.",
     "prompt_en": "Treat the monitor as a tactile luminous object in the dark bluish PC-room ambience: preserve subtle pixel texture, restrained screen glare, and gentle edge falloff across the physical display without obscuring the map. Hold the offshore red marker as the sharpest and most saturated point at frame center, with the bezel slipping into cooler shadow to create a quiet, ominous pull."
    }
   ]
  },
  "reused": false
 },
 "S22sh3": {
  "input_fingerprint": "522d94a3d50fa5f4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면에 바다 위 특정 좌표가 붉게 표시된 해안가 지도가 띄워진 클로즈업.\n\nLOCATION (lock): At the same PC café workstation, focused tightly on the monitor displaying the searched maritime coordinates. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the dolly-in directly in front of the monitor at screen-center height, ending square to the display with the coastal map and the coordinate marker held at the center. Retain a narrow border of bezel around the map so the image reads as displayed information rather than an omniscient map view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: PC monitor (On, displaying the plotted coordinates) — The front screen face is perpendicular to camera and displays a coastal map with a red coordinate marker positioned over the sea; used as Primary focal surface revealing the plotted destination.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The monitor glow carries a muted crimson coordinate accent against the PC room's dark bluish, desaturated ambience.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The offshore coordinate remains marked on the digital map; the handwritten coordinates stay in her pocket, and the resulting printed maps are folded into her backpack.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면에 바다 위 특정 좌표가 붉게 표시된 해안가 지도가 띄워진 클로즈업.\n\nLOCATION (lock): At the same PC café workstation, focused tightly on the monitor displaying the searched maritime coordinates. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the dolly-in directly in front of the monitor at screen-center height, ending square to the display with the coastal map and the coordinate marker held at the center. Retain a narrow border of bezel around the map so the image reads as displayed information rather than an omniscient map view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: PC monitor (On, displaying the plotted coordinates) — The front screen face is perpendicular to camera and displays a coastal map with a red coordinate marker positioned over the sea; used as Primary focal surface revealing the plotted destination.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The monitor glow carries a muted crimson coordinate accent against the PC room's dark bluish, desaturated ambience.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The offshore coordinate remains marked on the digital map; the handwritten coordinates stay in her pocket, and the resulting printed maps are folded into her backpack.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse rigid frontal symmetry and clean, controlled focus, with the coastline geometry, offshore position, and muted red coordinate marker rendered crisply across the display. Keep the narrow bezel evenly visible on all sides, letting the cool desaturated room ambience recede while the monitor becomes a precise, restrained reveal.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면에 바다 위 특정 좌표가 붉게 표시된 해안가 지도가 띄워진 클로즈업.\n\nLOCATION (lock): At the same PC café workstation, focused tightly on the monitor displaying the searched maritime coordinates. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Complete the dolly-in directly in front of the monitor at screen-center height, ending square to the display with the coastal map and the coordinate marker held at the center. Retain a narrow border of bezel around the map so the image reads as displayed information rather than an omniscient map view.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: PC monitor (On, displaying the plotted coordinates) — The front screen face is perpendicular to camera and displays a coastal map with a red coordinate marker positioned over the sea; used as Primary focal surface revealing the plotted destination.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The monitor glow carries a muted crimson coordinate accent against the PC room's dark bluish, desaturated ambience.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The offshore coordinate remains marked on the digital map; the handwritten coordinates stay in her pocket, and the resulting printed maps are folded into her backpack.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nTreat the monitor as a tactile luminous object in the dark bluish PC-room ambience: preserve subtle pixel texture, restrained screen glare, and gentle edge falloff across the physical display without obscuring the map. Hold the offshore red marker as the sharpest and most saturated point at frame center, with the bezel slipping into cooler shadow to create a quiet, ominous pull.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 8,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "클로즈업 프레이밍 지시를 정확히 따랐으며, 레퍼런스의 푸른 지도 화면 색감을 매우 잘 유지했습니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "프레이밍과 마커 표시는 훌륭하나, 지도에 나타난 육지의 색감이 레퍼런스 화면과 다르게 붉고 따뜻하게 표현되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S22sh2_sel.png"
   }
  ],
  "critique": {
   "issues": []
  },
  "composition_critique": {
   "issues": []
  },
  "fix_skipped": true,
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh2"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "groupbg::빗속_해안길": {
  "input_fingerprint": "60852e5e5e03198c",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "빗속_해안길",
    "tags": [
     "S23sh3"
    ]
   },
   "context_sig": "017940e3c2aa2d02"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n폭우가 쏟아지는 해안가 도로: 천둥 번개를 동반한 세찬 소나기가 내리는 아스팔트 해안도로. 짙은 먹구름과 빗줄기로 인해 시야가 제한되고 어둡다. (특징: 거세게 쏟아지는 굵은 빗줄기; 번개 불빛에 순간적으로 드러나는 젖은 아스팔트 반사광; 시야를 가리는 짙고 어두운 먹구름)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#24. 해안가 / 소나기\n- 자전거를 타고 빗속을 달리는 수리영.\n\nTIME OF DAY (lock): day, stormy.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n폭우가 쏟아지는 해안가 도로: 천둥 번개를 동반한 세찬 소나기가 내리는 아스팔트 해안도로. 짙은 먹구름과 빗줄기로 인해 시야가 제한되고 어둡다. (특징: 거세게 쏟아지는 굵은 빗줄기; 번개 불빛에 순간적으로 드러나는 젖은 아스팔트 반사광; 시야를 가리는 짙고 어두운 먹구름)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#24. 해안가 / 소나기\n- 자전거를 타고 빗속을 달리는 수리영.\n\nTIME OF DAY (lock): day, stormy.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_빗속_해안길_84857e.png",
  "asset_id": "0f9806d3-c5da-48dc-89a1-6a68e1f10632",
  "input_asset_ids": [
   "9837c1bc-cea4-4ad8-9113-094f56b82b72"
  ],
  "origin_tag": "S23sh3",
  "place_text": "On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face.",
  "origin_inputs": {
   "place_text": "On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face.",
   "time_of_day_en": "day, stormy",
   "conti_asset_id": "9837c1bc-cea4-4ad8-9113-094f56b82b72"
  }
 },
 "S23sh3::bgfirst_bg": {
  "input_fingerprint": "24d8e15f51119fc6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 얼굴에 부딪혀 흩어지는 빗방울 사이로 두 눈을 부릅뜬 수리영의 비장한 안면 정면 클로즈업.\n\nLOCATION (lock): On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face.\n\nTIME OF DAY (lock): day, stormy.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Travel backward directly ahead of the bicycle at eye level while closing into the tightest frontal facial view at the end of the move. 수리영's rain-struck face occupies most of the frame, with her widened eyes held near the upper center and the bicycle path reduced to unstable peripheral traces behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground, moves toward camera position along the bicycle path.\n- KEY BACKGROUND ELEMENTS: bicycle path (Obscured by heavy rain) — The route recedes directly behind 수리영 along the camera's backward travel axis; used as Soft, unstable background reference behind the moving face; heavy rain (Falling forcefully as droplets strike and scatter); used as Visible impact and motion texture across her face and frame edges; storm clouds (Densely covering the coastal sky); used as Background weather mass that suppresses distant detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Storm-darkened daylight is punctuated by the scene's lightning, with desaturated skin tones and hard rain highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 얼굴에 부딪혀 흩어지는 빗방울 사이로 두 눈을 부릅뜬 수리영의 비장한 안면 정면 클로즈업.\n\nLOCATION (lock): On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face.\n\nTIME OF DAY (lock): day, stormy.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Travel backward directly ahead of the bicycle at eye level while closing into the tightest frontal facial view at the end of the move. 수리영's rain-struck face occupies most of the frame, with her widened eyes held near the upper center and the bicycle path reduced to unstable peripheral traces behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground, moves toward camera position along the bicycle path.\n- KEY BACKGROUND ELEMENTS: bicycle path (Obscured by heavy rain) — The route recedes directly behind 수리영 along the camera's backward travel axis; used as Soft, unstable background reference behind the moving face; heavy rain (Falling forcefully as droplets strike and scatter); used as Visible impact and motion texture across her face and frame edges; storm clouds (Densely covering the coastal sky); used as Background weather mass that suppresses distant detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Storm-darkened daylight is punctuated by the scene's lightning, with desaturated skin tones and hard rain highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S23sh3__bgfirst_bg.png",
  "asset_id": "c6a0f313-257d-4354-b2b3-b63a02ae6d12",
  "input_asset_ids": [
   "9837c1bc-cea4-4ad8-9113-094f56b82b72",
   "0f9806d3-c5da-48dc-89a1-6a68e1f10632"
  ]
 },
 "S23sh3": {
  "input_fingerprint": "e6ba32e9bc8876c0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day, stormy.\n\nSHOT TEXT (authoritative, Korean): 얼굴에 부딪혀 흩어지는 빗방울 사이로 두 눈을 부릅뜬 수리영의 비장한 안면 정면 클로즈업.\n\nLOCATION (lock): On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Travel backward directly ahead of the bicycle at eye level while closing into the tightest frontal facial view at the end of the move. 수리영's rain-struck face occupies most of the frame, with her widened eyes held near the upper center and the bicycle path reduced to unstable peripheral traces behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground, moves toward camera position along the bicycle path.\n- KEY BACKGROUND ELEMENTS: bicycle path (Obscured by heavy rain) — The route recedes directly behind 수리영 along the camera's backward travel axis; used as Soft, unstable background reference behind the moving face; heavy rain (Falling forcefully as droplets strike and scatter); used as Visible impact and motion texture across her face and frame edges; storm clouds (Densely covering the coastal sky); used as Background weather mass that suppresses distant detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Storm-darkened daylight is punctuated by the scene's lightning, with desaturated skin tones and hard rain highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong rides her bicycle through the rain with the coordinate note still in her pocket and several printed maps stored in her backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day, stormy.\n\nSHOT TEXT (authoritative, Korean): 얼굴에 부딪혀 흩어지는 빗방울 사이로 두 눈을 부릅뜬 수리영의 비장한 안면 정면 클로즈업.\n\nLOCATION (lock): On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Travel backward directly ahead of the bicycle at eye level while closing into the tightest frontal facial view at the end of the move. 수리영's rain-struck face occupies most of the frame, with her widened eyes held near the upper center and the bicycle path reduced to unstable peripheral traces behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground, moves toward camera position along the bicycle path.\n- KEY BACKGROUND ELEMENTS: bicycle path (Obscured by heavy rain) — The route recedes directly behind 수리영 along the camera's backward travel axis; used as Soft, unstable background reference behind the moving face; heavy rain (Falling forcefully as droplets strike and scatter); used as Visible impact and motion texture across her face and frame edges; storm clouds (Densely covering the coastal sky); used as Background weather mass that suppresses distant detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Storm-darkened daylight is punctuated by the scene's lightning, with desaturated skin tones and hard rain highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong rides her bicycle through the rain with the coordinate note still in her pocket and several printed maps stored in her backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): day, stormy.\n\nSHOT TEXT (authoritative, Korean): 얼굴에 부딪혀 흩어지는 빗방울 사이로 두 눈을 부릅뜬 수리영의 비장한 안면 정면 클로즈업.\n\nLOCATION (lock): On an exposed coastal road during a violent rainstorm, with wind-driven rain striking the cyclist’s face. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Travel backward directly ahead of the bicycle at eye level while closing into the tightest frontal facial view at the end of the move. 수리영's rain-struck face occupies most of the frame, with her widened eyes held near the upper center and the bicycle path reduced to unstable peripheral traces behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, foreground, moves toward camera position along the bicycle path.\n- KEY BACKGROUND ELEMENTS: bicycle path (Obscured by heavy rain) — The route recedes directly behind 수리영 along the camera's backward travel axis; used as Soft, unstable background reference behind the moving face; heavy rain (Falling forcefully as droplets strike and scatter); used as Visible impact and motion texture across her face and frame edges; storm clouds (Densely covering the coastal sky); used as Background weather mass that suppresses distant detail.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Storm-darkened daylight is punctuated by the scene's lightning, with desaturated skin tones and hard rain highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong rides her bicycle through the rain with the coordinate note still in her pocket and several printed maps stored in her backpack.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S23sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S23sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_빗속_해안길_84857e.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지정된 '안면 정면 클로즈업' 샷 크기를 정확히 구현하여 얼굴이 프레임의 대부분을 차지하도록 연출했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지정된 타이트한 클로즈업 샷보다 넓게 프레임을 잡아 불필요하게 상체와 자전거 손잡이까지 노출하여 감점되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지정된 '안면 정면 클로즈업' 샷 크기를 정확히 구현하여 얼굴이 프레임의 대부분을 차지하도록 연출했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지정된 타이트한 클로즈업 샷보다 넓게 프레임을 잡아 불필요하게 상체와 자전거 손잡이까지 노출하여 감점되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "지정된 '안면 정면 클로즈업' 프레이밍을 정확히 구현하여 프레임 대부분을 얼굴로 채웠으며, 캐릭터 레퍼런스의 비니와 복장 디테일을 더 충실히 반영했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "비장한 표정과 날씨 연출은 좋으나, 카메라 프레임이 너무 넓어져 지시된 타이트한 안면 클로즈업 대신 상반신과 자전거 핸들까지 노출되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지정된 '안면 정면 클로즈업' 프레이밍을 정확히 구현하여 프레임 대부분을 얼굴로 채웠으며, 캐릭터 레퍼런스의 비니와 복장 디테일을 더 충실히 반영했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "비장한 표정과 날씨 연출은 좋으나, 카메라 프레임이 너무 넓어져 지시된 타이트한 안면 클로즈업 대신 상반신과 자전거 핸들까지 노출되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 18,
     "B": 13
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 18,
   "B": 13
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "지정된 '안면 정면 클로즈업' 샷 크기를 정확히 구현하여 얼굴이 프레임의 대부분을 차지하도록 연출했습니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "지정된 타이트한 클로즈업 샷보다 넓게 프레임을 잡아 불필요하게 상체와 자전거 손잡이까지 노출하여 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_빗속_해안길_84857e.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "인물이 레이아웃 스케치와 캐릭터 레퍼런스에는 없는 후드를 머리에 뒤집어쓰고 있습니다.",
     "fix_en": "Remove the jacket hood covering the character's head, leaving only her beanie and freely blowing hair visible to match the layout sketch."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the jacket hood covering the character's head, leaving only her beanie and freely blowing hair visible to match the layout sketch.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "스케치의 머리카락 흩날림과 캐릭터 레퍼런스의 복장(비니)을 정확히 반영하며, 비장한 표정의 정면 클로즈업을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "요구된 구도와 배경은 잘 따랐으나, 레퍼런스와 스케치에 없는 후드를 뒤집어써 머리카락의 형태가 스케치와 다릅니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "스케치의 머리카락 흩날림과 캐릭터 레퍼런스의 복장(비니)을 정확히 반영하며, 비장한 표정의 정면 클로즈업을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "요구된 구도와 배경은 잘 따랐으나, 레퍼런스와 스케치에 없는 후드를 뒤집어써 머리카락의 형태가 스케치와 다릅니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "레이아웃 스케치에 묘사된 흩날리는 머리카락과 캐릭터 레퍼런스의 복장(후드를 쓰지 않은 상태)을 정확히 구현하여 높은 일치도를 보입니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "배경과 인물의 표정은 훌륭하게 구현되었으나, 스케치나 레퍼런스에 없는 후드를 착용하여 레이아웃의 머리카락 묘사와 어긋납니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "레이아웃 스케치에 묘사된 흩날리는 머리카락과 캐릭터 레퍼런스의 복장(후드를 쓰지 않은 상태)을 정확히 구현하여 높은 일치도를 보입니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "배경과 인물의 표정은 훌륭하게 구현되었으나, 스케치나 레퍼런스에 없는 후드를 착용하여 레이아웃의 머리카락 묘사와 어긋납니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S23sh3__bgfirst_bg.png",
   "bg_asset_id": "c6a0f313-257d-4354-b2b3-b63a02ae6d12",
   "bg_record_key": "S23sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "빗속_해안길",
   "groupbg_asset_id": "0f9806d3-c5da-48dc-89a1-6a68e1f10632"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S24sh4::variants": {
  "author_fp": "f097c5cb4f351d1f",
  "author": {
   "variants": [
    {
     "approach_ko": "정면 대칭과 깊은 초점으로 단서를 냉정하고 법의학적인 정물처럼 포착한다.",
     "prompt_en": "Photograph the locked frontal composition with severe symmetry and a neutral, undistorted lens feel. Use deep, even focus across the wall plane, restrained dusk exposure, and broad soft ambient illumination so brush texture, pinning, stains, faded emulsion, and wall wear register with clinical clarity. Let the surrounding darkness fall away gently without obscuring the ritual geometry."
    },
    {
     "approach_ko": "비스듬히 스미는 황혼빛과 얕은 초점으로 표면의 상처와 불길한 촉감을 강조한다.",
     "prompt_en": "Keep the required square-on close view, but give it a compressed, intimate lens feel with focus held decisively at the central clue and a subtle softness toward the outer wall. Shape the dusk ambience as a low, raking lateral wash that catches raised brush ridges, dried irregularities, curled photographic edges, and punctured paper while the opposite side recedes into desaturated shadow. Preserve the centered geometry as an oppressive, tactile presence rather than a graphic design."
    }
   ]
  },
  "reused": false
 },
 "S24sh4": {
  "input_fingerprint": "9c77bea48ec8562c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 수리영의 방 벽면에 피처럼 붉은 도료로 거칠게 칠해진 원과 그 중앙에 꽂힌 낡은 사진 정물.\n\nLOCATION (lock): On the wall of the daughter’s bedroom inside the ransacked rooftop apartment, centered on the red circle and pinned photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the doorway-to-wall dolly at eye level in a square, frontal close view, then hold briefly with the rough red circle centered around the photograph. Keep enough of the surrounding wall visible to preserve the ritual-like geometry while the photograph remains clearly smaller than the painted marking.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: red circle on the bedroom wall (Freshly discovered on the wall) — The marked wall face is square to camera, bearing a roughly painted blood-red circle around the photograph; used as Geometric frame surrounding the photograph and primary occult visual anchor; old photograph (Bloodied and fixed at the circle's center) — Its image-bearing front faces camera, showing the old rural restaurant scene and two young girls beneath visible blood marks; used as Central narrative clue fixed inside the red marking.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural ambient light appropriate to the room at dusk preserves a dark, desaturated field around the restrained red marking.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains ransacked, with the living-room window open and chairs overturned. A bloodstained old photograph is pinned at the center of a rough red circle on Su-riyeong’s wall, while her map-filled backpack and coordinate note remain with her.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 수리영의 방 벽면에 피처럼 붉은 도료로 거칠게 칠해진 원과 그 중앙에 꽂힌 낡은 사진 정물.\n\nLOCATION (lock): On the wall of the daughter’s bedroom inside the ransacked rooftop apartment, centered on the red circle and pinned photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the doorway-to-wall dolly at eye level in a square, frontal close view, then hold briefly with the rough red circle centered around the photograph. Keep enough of the surrounding wall visible to preserve the ritual-like geometry while the photograph remains clearly smaller than the painted marking.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: red circle on the bedroom wall (Freshly discovered on the wall) — The marked wall face is square to camera, bearing a roughly painted blood-red circle around the photograph; used as Geometric frame surrounding the photograph and primary occult visual anchor; old photograph (Bloodied and fixed at the circle's center) — Its image-bearing front faces camera, showing the old rural restaurant scene and two young girls beneath visible blood marks; used as Central narrative clue fixed inside the red marking.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural ambient light appropriate to the room at dusk preserves a dark, desaturated field around the restrained red marking.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains ransacked, with the living-room window open and chairs overturned. A bloodstained old photograph is pinned at the center of a rough red circle on Su-riyeong’s wall, while her map-filled backpack and coordinate note remain with her.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph the locked frontal composition with severe symmetry and a neutral, undistorted lens feel. Use deep, even focus across the wall plane, restrained dusk exposure, and broad soft ambient illumination so brush texture, pinning, stains, faded emulsion, and wall wear register with clinical clarity. Let the surrounding darkness fall away gently without obscuring the ritual geometry.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 수리영의 방 벽면에 피처럼 붉은 도료로 거칠게 칠해진 원과 그 중앙에 꽂힌 낡은 사진 정물.\n\nLOCATION (lock): On the wall of the daughter’s bedroom inside the ransacked rooftop apartment, centered on the red circle and pinned photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the doorway-to-wall dolly at eye level in a square, frontal close view, then hold briefly with the rough red circle centered around the photograph. Keep enough of the surrounding wall visible to preserve the ritual-like geometry while the photograph remains clearly smaller than the painted marking.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: red circle on the bedroom wall (Freshly discovered on the wall) — The marked wall face is square to camera, bearing a roughly painted blood-red circle around the photograph; used as Geometric frame surrounding the photograph and primary occult visual anchor; old photograph (Bloodied and fixed at the circle's center) — Its image-bearing front faces camera, showing the old rural restaurant scene and two young girls beneath visible blood marks; used as Central narrative clue fixed inside the red marking.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued natural ambient light appropriate to the room at dusk preserves a dark, desaturated field around the restrained red marking.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains ransacked, with the living-room window open and chairs overturned. A bloodstained old photograph is pinned at the center of a rough red circle on Su-riyeong’s wall, while her map-filled backpack and coordinate note remain with her.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nKeep the required square-on close view, but give it a compressed, intimate lens feel with focus held decisively at the central clue and a subtle softness toward the outer wall. Shape the dusk ambience as a low, raking lateral wash that catches raised brush ridges, dried irregularities, curled photographic edges, and punctured paper while the opposite side recedes into desaturated shadow. Preserve the centered geometry as an oppressive, tactile presence rather than a graphic design.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 3,
   "B": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "클로즈업 앵글 지시를 무시하고 너무 넓게 촬영했으며, 벽이 방 한가운데 기둥처럼 변형되는 심각한 공간적 오류가 발생했습니다."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지시된 클로즈업 프레임과 정면 구도를 잘 구현했으며, 레퍼런스의 벽지 질감과 난장판이 된 배경의 디테일을 충실히 반영했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B08.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프레임 오른쪽에 다른 방이 보이도록 벽이 끊어져 있어, 정면에서 벽면이 화면을 꽉 채워야 하는 기하학적 구도가 깨졌습니다.",
     "fix_en": "Extend the floral wallpaper seamlessly across the entire right side of the frame so the wall forms a continuous, flat background completely filling the space around the circle."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reframe directly toward the wall so the red circle and pinned photograph sit at the exact center of the image, removing the right-side doorway from the composition while preserving the close-up scale and surrounding wall.",
     "issue_ko": "붉은 원과 사진이 화면 중앙이 아니라 왼쪽으로 치우쳐 있고, 오른쪽 출입구가 불필요하게 큰 화면 영역을 차지한다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Extend the floral wallpaper seamlessly across the entire right side of the frame so the wall forms a continuous, flat background completely filling the space around the circle.\n- Reframe directly toward the wall so the red circle and pinned photograph sit at the exact center of the image, removing the right-side doorway from the composition while preserving the close-up scale and surrounding wall.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 정면 클로즈업 구도와 측면 조명 분위기를 완벽하게 구현하며, 요구된 질감과 디테일도 우수합니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "벽의 단면을 임의로 만들어 배경 방을 노출함으로써 '완전한 정면 뷰'라는 핵심 프레이밍 지시를 어겼습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 정면 클로즈업 구도와 측면 조명 분위기를 완벽하게 구현하며, 요구된 질감과 디테일도 우수합니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "벽의 단면을 임의로 만들어 배경 방을 노출함으로써 '완전한 정면 뷰'라는 핵심 프레이밍 지시를 어겼습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 정면 클로즈업 구도를 정확히 준수하였으며, 벽면의 질감과 조명을 사실적으로 표현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 정면 구도를 어기고 벽면 모서리와 배경을 노출하여 프레이밍 최우선 지침을 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 정면 클로즈업 구도를 정확히 준수하였으며, 벽면의 질감과 조명을 사실적으로 표현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "요구된 정면 구도를 어기고 벽면 모서리와 배경을 노출하여 프레이밍 최우선 지침을 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B01",
   "choice": "Candidate C",
   "confident": true,
   "reason_ko": "지문에서 언급된 '벽면에 피처럼 붉은 도료로 거칠게 칠해진 원'이 우측 벽면에 명확하게 그려져 있는 유일한 이미지입니다.",
   "kept": "L06B01"
  }
 },
 "S24sh8::variants": {
  "author_fp": "568741d6455a1bcb",
  "author": {
   "variants": [
    {
     "approach_ko": "깊은 초점과 사선형 공간 구도로 어질러진 방을 가로질러 열린 문으로 향하는 긴장된 이탈 동선을 강조한다.",
     "prompt_en": "Use a restrained, moderately wide lens feel with deep, legible focus, preserving the shoulder-height medium over-shoulder framing. Build strong diagonal depth from the clenched photograph at the lower edge, through the centrally weighted backpack and broken angles of the overturned chairs, toward the open doorway at middle-right. Let the cool dusk ambience fall gently across the room so the doorway reads as the clearest spatial destination, while the photograph’s blood remains the sole muted red accent."
    },
    {
     "approach_ko": "압축된 얕은 초점으로 백팩과 사진을 쥔 손의 촉감을 전면화하고, 열린 문은 부드럽지만 분명한 탈출 지점으로 남긴다.",
     "prompt_en": "Use a tighter, gently compressed lens feel within the locked shoulder-height medium over-shoulder composition. Hold crisp tactile focus on the backpack’s fabric, her turning shoulder, and the tension in the hand crushing the photograph; allow the disordered room to fall into a shallow, layered softness while keeping the open doorway at middle-right unmistakably readable. Shape the low dusk light with soft side falloff across her back and hand, creating an intimate suspended instant rather than broad spatial coverage."
    }
   ]
  },
  "reused": false
 },
 "S24sh8": {
  "input_fingerprint": "6a6d977895cc018f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 검은 백팩을 멘 수리영이 손에 구겨진 사진을 꽉 쥔 채 열린 현관문을 향해 몸을 튼 정지 찰나.\n\nLOCATION (lock): In the ransacked rooftop apartment’s living area, turning toward the open entrance door with the photograph in hand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track at shoulder height behind and slightly beside 수리영, then pause in a medium over-shoulder composition as she turns mid-step toward the open front door. Her backpack occupies the central midground, while her clenched hand and crumpled photograph sit clearly in the lower side of frame and the doorway remains ahead of her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, looks toward open front door; open front door in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: front door (Open) — The open doorway is seen obliquely ahead of her turned body; used as Departure anchor directly ahead of 수리영; dining chairs (Overturned) — Their seat and back planes lie at broken angles to the camera; used as Disordered spatial obstacle along the route through the living area; rummaged household items (Pulled apart and left in disorder) — The displaced objects appear irregularly across the room behind and beside her path; used as Peripheral evidence of the intrusion and disturbed home.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to dusk is kept low-key and desaturated, with the bloodied photograph providing the only restrained red emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ransacked apartment, open window, overturned chairs, and red wall circle remain behind, now with the photograph removed from its center. Su-riyeong leaves wearing the backpack containing her maps, the coordinate note still in her pocket, and the old photograph clenched in her hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 검은 백팩을 멘 수리영이 손에 구겨진 사진을 꽉 쥔 채 열린 현관문을 향해 몸을 튼 정지 찰나.\n\nLOCATION (lock): In the ransacked rooftop apartment’s living area, turning toward the open entrance door with the photograph in hand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track at shoulder height behind and slightly beside 수리영, then pause in a medium over-shoulder composition as she turns mid-step toward the open front door. Her backpack occupies the central midground, while her clenched hand and crumpled photograph sit clearly in the lower side of frame and the doorway remains ahead of her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, looks toward open front door; open front door in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: front door (Open) — The open doorway is seen obliquely ahead of her turned body; used as Departure anchor directly ahead of 수리영; dining chairs (Overturned) — Their seat and back planes lie at broken angles to the camera; used as Disordered spatial obstacle along the route through the living area; rummaged household items (Pulled apart and left in disorder) — The displaced objects appear irregularly across the room behind and beside her path; used as Peripheral evidence of the intrusion and disturbed home.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to dusk is kept low-key and desaturated, with the bloodied photograph providing the only restrained red emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ransacked apartment, open window, overturned chairs, and red wall circle remain behind, now with the photograph removed from its center. Su-riyeong leaves wearing the backpack containing her maps, the coordinate note still in her pocket, and the old photograph clenched in her hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained, moderately wide lens feel with deep, legible focus, preserving the shoulder-height medium over-shoulder framing. Build strong diagonal depth from the clenched photograph at the lower edge, through the centrally weighted backpack and broken angles of the overturned chairs, toward the open doorway at middle-right. Let the cool dusk ambience fall gently across the room so the doorway reads as the clearest spatial destination, while the photograph’s blood remains the sole muted red accent.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 검은 백팩을 멘 수리영이 손에 구겨진 사진을 꽉 쥔 채 열린 현관문을 향해 몸을 튼 정지 찰나.\n\nLOCATION (lock): In the ransacked rooftop apartment’s living area, turning toward the open entrance door with the photograph in hand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track at shoulder height behind and slightly beside 수리영, then pause in a medium over-shoulder composition as she turns mid-step toward the open front door. Her backpack occupies the central midground, while her clenched hand and crumpled photograph sit clearly in the lower side of frame and the doorway remains ahead of her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 수리영 in the middle-center of the frame, midground, looks toward open front door; open front door in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: front door (Open) — The open doorway is seen obliquely ahead of her turned body; used as Departure anchor directly ahead of 수리영; dining chairs (Overturned) — Their seat and back planes lie at broken angles to the camera; used as Disordered spatial obstacle along the route through the living area; rummaged household items (Pulled apart and left in disorder) — The displaced objects appear irregularly across the room behind and beside her path; used as Peripheral evidence of the intrusion and disturbed home.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to dusk is kept low-key and desaturated, with the bloodied photograph providing the only restrained red emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ransacked apartment, open window, overturned chairs, and red wall circle remain behind, now with the photograph removed from its center. Su-riyeong leaves wearing the backpack containing her maps, the coordinate note still in her pocket, and the old photograph clenched in her hand.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a tighter, gently compressed lens feel within the locked shoulder-height medium over-shoulder composition. Hold crisp tactile focus on the backpack’s fabric, her turning shoulder, and the tension in the hand crushing the photograph; allow the disordered room to fall into a shallow, layered softness while keeping the open doorway at middle-right unmistakably readable. Shape the low dusk light with soft side falloff across her back and hand, creating an intimate suspended instant rather than broad spatial coverage.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 4,
   "B": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "현관문이 아닌 발코니 문이 배경에 있고, 벽의 붉은 원 안에 지시되지 않은 무늬가 추가되었으며, 구겨진 사진을 꽉 쥔 형태가 아닙니다."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "사진을 움켜쥔 손의 형태와 복도로 이어지는 현관문의 묘사가 지침에 부합하나, 시선이 문을 향하지 않은 점은 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S24sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   },
   {
    "label": "PROP REFERENCE — 두 한국인 어린 소녀가 찍힌 낡은 사진: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1242961>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영이 오른쪽 배경에 있는 열린 현관문을 향해 시선을 두지 않고, 화면 왼쪽을 바라보고 있습니다.",
     "fix_en": "Turn Su-riyeong's head and gaze to the right so she is looking directly at the open front door in the background."
    },
    {
     "issue_ko": "지시된 '구겨진 사진'과 '꽉 쥔 손'이 반영되지 않아, 사진이 빳빳한 판지처럼 평평하고 손은 가장자리를 느슨하게 잡고 있습니다.",
     "fix_en": "Crumple the photograph heavily and change the hand pose to tightly clench it into a fist."
    },
    {
     "issue_ko": "피 묻은 사진에만 붉은색 포인트를 제한하라는 지시와 달리, 화면 우측 배경에 강렬한 붉은색 물건이 존재합니다.",
     "fix_en": "Desaturate the bright red object on the right side so the only red elements in the image are the blood spots on the photograph."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Turn Su-riyeong’s head and gaze toward the open doorway at middle-right while preserving her body position, backpack, hand, and all other framing.",
     "issue_ko": "수리영의 몸은 열린 현관문 쪽으로 향하지만 머리와 시선은 화면 왼쪽 뒤를 보고 있어, 현관문을 바라보도록 지정된 동작 축과 어긋난다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Turn Su-riyeong's head and gaze to the right so she is looking directly at the open front door in the background.\n- Crumple the photograph heavily and change the hand pose to tightly clench it into a fist.\n- Desaturate the bright red object on the right side so the only red elements in the image are the blood spots on the photograph.\n- Turn Su-riyeong’s head and gaze toward the open doorway at middle-right while preserving her body position, backpack, hand, and all other framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "심하게 구겨진 사진을 꽉 쥔 손의 디테일이 프롬프트의 핵심 소품 및 포즈 요구사항을 완벽히 충족합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "사진이 전혀 구겨지지 않고 손에 가볍게 들려 있어 '구겨진 사진을 꽉 쥔 채'라는 명시적인 지시를 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "심하게 구겨진 사진을 꽉 쥔 손의 디테일이 프롬프트의 핵심 소품 및 포즈 요구사항을 완벽히 충족합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "사진이 전혀 구겨지지 않고 손에 가볍게 들려 있어 '구겨진 사진을 꽉 쥔 채'라는 명시적인 지시를 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 미디엄 오버숄더 샷과 인물의 포즈를 잘 구현했으며, 손에 쥔 구겨진 사진의 디테일(두 소녀, 핏자국)이 레퍼런스와 가장 정확하게 일치합니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "전반적인 구도와 배경은 좋으나, 손에 쥔 구겨진 사진 표면에 인물의 얼굴이 기형적으로 왜곡되고 중복 출력되는 오류가 있습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 미디엄 오버숄더 샷과 인물의 포즈를 잘 구현했으며, 손에 쥔 구겨진 사진의 디테일(두 소녀, 핏자국)이 레퍼런스와 가장 정확하게 일치합니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "전반적인 구도와 배경은 좋으나, 손에 쥔 구겨진 사진 표면에 인물의 얼굴이 기형적으로 왜곡되고 중복 출력되는 오류가 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S24sh4"
  }
 },
 "groupbg::포구_접안부": {
  "input_fingerprint": "e815d2c3697b4489",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "포구_접안부",
    "tags": [
     "S25sh2",
     "S25sh3",
     "S25sh6"
    ]
   },
   "context_sig": "fb75e53ff01802de"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n안개 낀 포구: 안개가 짙게 깔려 배들이 제대로 보이지 않는 스산하고 텅 빈 소형 항구. 바닷물이 닿는 갯바위와 배를 대는 정박 시설이 있다. (특징: 시야를 가리는 짙은 회색 안개; 바닷물이 찰랑이는 거친 갯바위; 안개를 뚫고 다가오는 낡은 소형 어선의 실루엣; 인기척 없이 비어있는 부두)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#26. 포구 / 해질 무렵\n- 텅 빈 포구. 안개로 인해 배는 보이지 않고.\n- 갯바위에 앉아있는 수리영.\n\nTIME OF DAY (lock): dusk, foggy after rain.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated: On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n안개 낀 포구: 안개가 짙게 깔려 배들이 제대로 보이지 않는 스산하고 텅 빈 소형 항구. 바닷물이 닿는 갯바위와 배를 대는 정박 시설이 있다. (특징: 시야를 가리는 짙은 회색 안개; 바닷물이 찰랑이는 거친 갯바위; 안개를 뚫고 다가오는 낡은 소형 어선의 실루엣; 인기척 없이 비어있는 부두)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- S#26. 포구 / 해질 무렵\n- 텅 빈 포구. 안개로 인해 배는 보이지 않고.\n- 갯바위에 앉아있는 수리영.\n\nTIME OF DAY (lock): dusk, foggy after rain.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_포구_접안부_ea2e3b.png",
  "asset_id": "e334ab2c-4f6a-406a-ac03-afdb38df6db6",
  "input_asset_ids": [
   "c0b3788b-e48a-46b0-ab5f-65c160ba5533"
  ],
  "origin_tag": "S25sh2",
  "place_text": "On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog.",
  "origin_inputs": {
   "place_text": "On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog.",
   "time_of_day_en": "dusk, foggy after rain",
   "conti_asset_id": "c0b3788b-e48a-46b0-ab5f-65c160ba5533"
  }
 },
 "S25sh2::bgfirst_bg": {
  "input_fingerprint": "a6228199cbddc390",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거친 갯바위 위에 무릎을 모으고 웅크려 앉아 회색빛 바다를 멍하니 응시하는 수리영의 측면.\n\nLOCATION (lock): On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog.\n\nTIME OF DAY (lock): dusk, foggy after rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at seated shoulder height from the side of the rocks, framing 수리영 in profile with her knees drawn in while leaving a broad field of sea and fog before her gaze. Her compact figure rests to one side of the wide frame, making the empty harbor and obscured water carry most of the spatial weight.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward fog-covered sea; fog-covered sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rough coastal rocks (Exposed along the harbor edge); used as Uneven support beneath 수리영 and textured foreground boundary; sea (Gray and partly obscured); used as Primary negative space in the direction of her stare; fog (Dense across the empty harbor); used as Suppresses visibility and conceals boats offshore; harbor (Empty); used as Environmental frame around the isolated figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk ambient light is muted into gray, desaturated tones, with fog flattening the distant contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 거친 갯바위 위에 무릎을 모으고 웅크려 앉아 회색빛 바다를 멍하니 응시하는 수리영의 측면.\n\nLOCATION (lock): On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog.\n\nTIME OF DAY (lock): dusk, foggy after rain.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at seated shoulder height from the side of the rocks, framing 수리영 in profile with her knees drawn in while leaving a broad field of sea and fog before her gaze. Her compact figure rests to one side of the wide frame, making the empty harbor and obscured water carry most of the spatial weight.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward fog-covered sea; fog-covered sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rough coastal rocks (Exposed along the harbor edge); used as Uneven support beneath 수리영 and textured foreground boundary; sea (Gray and partly obscured); used as Primary negative space in the direction of her stare; fog (Dense across the empty harbor); used as Suppresses visibility and conceals boats offshore; harbor (Empty); used as Environmental frame around the isolated figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk ambient light is muted into gray, desaturated tones, with fog flattening the distant contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S25sh2__bgfirst_bg.png",
  "asset_id": "020c40ea-33de-4507-ab90-3eba86f7e749",
  "input_asset_ids": [
   "c0b3788b-e48a-46b0-ab5f-65c160ba5533",
   "e334ab2c-4f6a-406a-ac03-afdb38df6db6"
  ]
 },
 "S25sh2": {
  "input_fingerprint": "832b0979c2f0b13a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 거친 갯바위 위에 무릎을 모으고 웅크려 앉아 회색빛 바다를 멍하니 응시하는 수리영의 측면.\n\nLOCATION (lock): On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at seated shoulder height from the side of the rocks, framing 수리영 in profile with her knees drawn in while leaving a broad field of sea and fog before her gaze. Her compact figure rests to one side of the wide frame, making the empty harbor and obscured water carry most of the spatial weight.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward fog-covered sea; fog-covered sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rough coastal rocks (Exposed along the harbor edge); used as Uneven support beneath 수리영 and textured foreground boundary; sea (Gray and partly obscured); used as Primary negative space in the direction of her stare; fog (Dense across the empty harbor); used as Suppresses visibility and conceals boats offshore; harbor (Empty); used as Environmental frame around the isolated figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk ambient light is muted into gray, desaturated tones, with fog flattening the distant contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains the old photograph, the map-filled backpack, and the coordinate note during her search of the harbor; her phone remains with her and has accumulated missed calls from Hye-su.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 거친 갯바위 위에 무릎을 모으고 웅크려 앉아 회색빛 바다를 멍하니 응시하는 수리영의 측면.\n\nLOCATION (lock): On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at seated shoulder height from the side of the rocks, framing 수리영 in profile with her knees drawn in while leaving a broad field of sea and fog before her gaze. Her compact figure rests to one side of the wide frame, making the empty harbor and obscured water carry most of the spatial weight.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward fog-covered sea; fog-covered sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rough coastal rocks (Exposed along the harbor edge); used as Uneven support beneath 수리영 and textured foreground boundary; sea (Gray and partly obscured); used as Primary negative space in the direction of her stare; fog (Dense across the empty harbor); used as Suppresses visibility and conceals boats offshore; harbor (Empty); used as Environmental frame around the isolated figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk ambient light is muted into gray, desaturated tones, with fog flattening the distant contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains the old photograph, the map-filled backpack, and the coordinate note during her search of the harbor; her phone remains with her and has accumulated missed calls from Hye-su.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 거친 갯바위 위에 무릎을 모으고 웅크려 앉아 회색빛 바다를 멍하니 응시하는 수리영의 측면.\n\nLOCATION (lock): On rough shoreline rocks at the edge of an empty harbor, facing a gray sea concealed by dense evening fog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold at seated shoulder height from the side of the rocks, framing 수리영 in profile with her knees drawn in while leaving a broad field of sea and fog before her gaze. Her compact figure rests to one side of the wide frame, making the empty harbor and obscured water carry most of the spatial weight.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground, looks toward fog-covered sea; fog-covered sea in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rough coastal rocks (Exposed along the harbor edge); used as Uneven support beneath 수리영 and textured foreground boundary; sea (Gray and partly obscured); used as Primary negative space in the direction of her stare; fog (Dense across the empty harbor); used as Suppresses visibility and conceals boats offshore; harbor (Empty); used as Environmental frame around the isolated figure.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk ambient light is muted into gray, desaturated tones, with fog flattening the distant contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong retains the old photograph, the map-filled backpack, and the coordinate note during her search of the harbor; her phone remains with her and has accumulated missed calls from Hye-su.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S25sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S25sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_포구_접안부_ea2e3b.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 앵글과 '무릎을 모으고 앉은' 포즈를 정확히 구현했으며, 배낭과 사진, 메모 등 유지해야 할 소지품 디테일이 잘 표현되었습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "배경과 구도는 훌륭하나, 쪼그려 앉은 자세가 지문과 다소 차이가 있고 사진과 메모 등 일부 소지품이 보이지 않습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 앵글과 '무릎을 모으고 앉은' 포즈를 정확히 구현했으며, 배낭과 사진, 메모 등 유지해야 할 소지품 디테일이 잘 표현되었습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "배경과 구도는 훌륭하나, 쪼그려 앉은 자세가 지문과 다소 차이가 있고 사진과 메모 등 일부 소지품이 보이지 않습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 와이드 숏 구도와 안개 낀 바다 배경을 잘 구현했으며, 지시된 소지품(사진, 배낭 속 지도와 메모)을 모두 정확하게 묘사하여 프롬프트 충실도가 가장 높습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "요구된 프레이밍과 쓸쓸한 배경 분위기는 훌륭하게 표현했으나, 지시문에 명시된 소지품 중 낡은 사진과 좌표 메모가 누락되어 아쉽습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 와이드 숏 구도와 안개 낀 바다 배경을 잘 구현했으며, 지시된 소지품(사진, 배낭 속 지도와 메모)을 모두 정확하게 묘사하여 프롬프트 충실도가 가장 높습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "요구된 프레이밍과 쓸쓸한 배경 분위기는 훌륭하게 표현했으나, 지시문에 명시된 소지품 중 낡은 사진과 좌표 메모가 누락되어 아쉽습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 12
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 15,
   "B": 12
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "지정된 앵글과 '무릎을 모으고 앉은' 포즈를 정확히 구현했으며, 배낭과 사진, 메모 등 유지해야 할 소지품 디테일이 잘 표현되었습니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "배경과 구도는 훌륭하나, 쪼그려 앉은 자세가 지문과 다소 차이가 있고 사진과 메모 등 일부 소지품이 보이지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/groupbg_포구_접안부_ea2e3b.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영이 들고 있는 사진의 앞면(이미지)이 그녀의 시선 방향이 아닌 카메라를 향하고 있습니다.",
     "fix_en": "Rotate the photograph so the picture side faces Su-riyeong's eyes, showing only the blank back of the photo to the camera."
    },
    {
     "issue_ko": "배낭 주머니에 꽂힌 흰색 메모지에 가짜 텍스트가 적혀 있습니다 (텍스트 포함 금지 규칙 위반).",
     "fix_en": "Remove the text lines from the white note on the backpack, leaving it as a blank piece of paper."
    },
    {
     "issue_ko": "캐릭터 레퍼런스 사진에 있는 회색 비니를 착용하지 않았습니다.",
     "fix_en": "Add the grey beanie to Su-riyeong's head to match her wardrobe in the character reference exactly."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Match the layout-sketch pose by resting both forearms compactly around/on the drawn-in knees and holding the photograph close to the knees instead of extending one arm toward the sea.",
     "issue_ko": "수리영이 사진을 든 팔을 바다 쪽으로 뻗고 있어, 두 무릎을 모아 안은 채 팔을 무릎 위에 밀착시키는 레이아웃 스케치의 웅크린 자세와 다릅니다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Rotate the photograph so the picture side faces Su-riyeong's eyes, showing only the blank back of the photo to the camera.\n- Remove the text lines from the white note on the backpack, leaving it as a blank piece of paper.\n- Add the grey beanie to Su-riyeong's head to match her wardrobe in the character reference exactly.\n- Match the layout-sketch pose by resting both forearms compactly around/on the drawn-in knees and holding the photograph close to the knees instead of extending one arm toward the sea.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "스케치의 구도와 배경을 정확히 구현했으며, 캐릭터 레퍼런스의 비니와 의상, 소품 상태를 충실하게 반영했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "구도와 배경, 자세는 스케치와 잘 일치하지만, 캐릭터 레퍼런스에 있는 비니를 누락하여 감점되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "스케치의 구도와 배경을 정확히 구현했으며, 캐릭터 레퍼런스의 비니와 의상, 소품 상태를 충실하게 반영했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "구도와 배경, 자세는 스케치와 잘 일치하지만, 캐릭터 레퍼런스에 있는 비니를 누락하여 감점되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "캐릭터 레퍼런스의 복장(비니 포함)과 소지품을 정확히 반영했으며, 제시된 스케치의 구도와 배경을 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "배경과 인물의 자세 및 구도는 지시사항을 잘 따랐으나, 캐릭터 레퍼런스 이미지에 있는 비니가 누락되어 복장 일치도가 다소 떨어집니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "캐릭터 레퍼런스의 복장(비니 포함)과 소지품을 정확히 반영했으며, 제시된 스케치의 구도와 배경을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "배경과 인물의 자세 및 구도는 지시사항을 잘 따랐으나, 캐릭터 레퍼런스 이미지에 있는 비니가 누락되어 복장 일치도가 다소 떨어집니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S25sh2__bgfirst_bg.png",
   "bg_asset_id": "020c40ea-33de-4507-ab90-3eba86f7e749",
   "bg_record_key": "S25sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "포구_접안부",
   "groupbg_asset_id": "e334ab2c-4f6a-406a-ac03-afdb38df6db6"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S25sh3::variants": {
  "author_fp": "2e526ab00af48b2c",
  "author": {
   "variants": [
    {
     "approach_ko": "절제된 증거물 인서트처럼 화면과 젖은 암석의 질감을 고르게 또렷하게 포착한다.",
     "prompt_en": "Photograph it as a restrained, objective evidence insert: a perfectly square overhead viewpoint with the phone held geometrically stable at the center and the irregular rock border acting as a quiet natural frame. Use a neutral close-focus lens feel with broad, disciplined sharpness across the display, device edges, and wet stone texture; suppress distracting glare so the active screen remains immediately legible. Let the cool dusk ambience reveal fine moisture and wear while the display provides clean, contained illumination without exaggerated bloom."
    },
    {
     "approach_ko": "얕은 초점과 젖은 표면의 은은한 반사로 전화기의 긴박한 진동 순간을 촉각적으로 강조한다.",
     "prompt_en": "Treat the insert as an intimate, tactile macro moment, still looking straight down and keeping the phone centered, but with a very shallow focus plane locked precisely onto the active display. Allow the narrow surrounding rocks and the far edges of the device to soften rapidly, while muted screen light feathers gently across the nearest wet texture. Preserve crisp readability at the focal plane and suggest the arriving vibration through only the faintest natural edge tremor, retaining the heavy, fog-muted dusk atmosphere."
    }
   ]
  },
  "reused": false
 },
 "S25sh3": {
  "input_fingerprint": "a9d991540adf6841",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 갯바위 바닥에 놓인 수리영의 핸드폰 화면에 혜수의 이름과 부재중 전화 알림이 떠 있는 클로즈업.\n\nLOCATION (lock): On the wet rock surface at the fogbound harbor’s edge, beside the daughter’s resting place overlooking the sea. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Look straight down from close above the rocks, centering the phone while leaving a narrow, irregular border of stone around it for scale. The screen is aligned for immediate readability, presenting 혜수's name and the stack of missed-call notifications without showing either person physically.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: phone showing missed calls in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 수리영's phone (Screen active as the vibration alert arrives) — The active front screen faces directly upward toward camera, showing 혜수's name and multiple missed-call notifications; used as Isolated information detail and primary focal object; coastal rocks (Visible beneath and around the device); used as Thin environmental border around the phone.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted dusk ambience and the active phone display create restrained contrast within the gray, desaturated harbor palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her phone displays multiple missed calls from Hye-su, while the old photograph, folded maps in her backpack, and coordinate note remain in her possession.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 갯바위 바닥에 놓인 수리영의 핸드폰 화면에 혜수의 이름과 부재중 전화 알림이 떠 있는 클로즈업.\n\nLOCATION (lock): On the wet rock surface at the fogbound harbor’s edge, beside the daughter’s resting place overlooking the sea. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Look straight down from close above the rocks, centering the phone while leaving a narrow, irregular border of stone around it for scale. The screen is aligned for immediate readability, presenting 혜수's name and the stack of missed-call notifications without showing either person physically.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: phone showing missed calls in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 수리영's phone (Screen active as the vibration alert arrives) — The active front screen faces directly upward toward camera, showing 혜수's name and multiple missed-call notifications; used as Isolated information detail and primary focal object; coastal rocks (Visible beneath and around the device); used as Thin environmental border around the phone.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted dusk ambience and the active phone display create restrained contrast within the gray, desaturated harbor palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her phone displays multiple missed calls from Hye-su, while the old photograph, folded maps in her backpack, and coordinate note remain in her possession.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph it as a restrained, objective evidence insert: a perfectly square overhead viewpoint with the phone held geometrically stable at the center and the irregular rock border acting as a quiet natural frame. Use a neutral close-focus lens feel with broad, disciplined sharpness across the display, device edges, and wet stone texture; suppress distracting glare so the active screen remains immediately legible. Let the cool dusk ambience reveal fine moisture and wear while the display provides clean, contained illumination without exaggerated bloom.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 갯바위 바닥에 놓인 수리영의 핸드폰 화면에 혜수의 이름과 부재중 전화 알림이 떠 있는 클로즈업.\n\nLOCATION (lock): On the wet rock surface at the fogbound harbor’s edge, beside the daughter’s resting place overlooking the sea. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Look straight down from close above the rocks, centering the phone while leaving a narrow, irregular border of stone around it for scale. The screen is aligned for immediate readability, presenting 혜수's name and the stack of missed-call notifications without showing either person physically.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: phone showing missed calls in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 수리영's phone (Screen active as the vibration alert arrives) — The active front screen faces directly upward toward camera, showing 혜수's name and multiple missed-call notifications; used as Isolated information detail and primary focal object; coastal rocks (Visible beneath and around the device); used as Thin environmental border around the phone.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted dusk ambience and the active phone display create restrained contrast within the gray, desaturated harbor palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her phone displays multiple missed calls from Hye-su, while the old photograph, folded maps in her backpack, and coordinate note remain in her possession.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nTreat the insert as an intimate, tactile macro moment, still looking straight down and keeping the phone centered, but with a very shallow focus plane locked precisely onto the active display. Allow the narrow surrounding rocks and the far edges of the device to soften rapidly, while muted screen light feathers gently across the nearest wet texture. Preserve crisp readability at the focal plane and suggest the arriving vibration through only the faintest natural edge tremor, retaining the heavy, fog-muted dusk atmosphere.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 8,
   "A": 6
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "지시된 완벽한 수직(Straight down) 카메라 앵글을 정확히 구현하였고, 젖은 바위와 핸드폰 화면의 디테일이 훌륭합니다."
   },
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "요구사항을 전반적으로 충족했으나, 카메라 앵글이 완전히 수직으로 내려다보지 않고 비스듬하게 설정되어 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S25sh2_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "부재중 전화 알림 배너 안의 텍스트가 깨진/의미 없는 한글(예: '부테믕 간와')로 렌더링되었습니다.",
     "fix_en": "Change the gibberish text in the notification banners to correctly spell '부재중 전화' (missed call) in legible Korean."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Move the camera closer while maintaining the straight-down axis so the centered phone fills most of the frame, leaving only a narrow irregular border of rock around it.",
     "issue_ko": "휴대폰 주변의 바위가 지나치게 넓게 보여 계약의 디테일 인서트 클로즈업보다 프레이밍이 너무 넓다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the gibberish text in the notification banners to correctly spell '부재중 전화' (missed call) in legible Korean.\n- Move the camera closer while maintaining the straight-down axis so the centered phone fills most of the frame, leaving only a narrow irregular border of rock around it.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "젖은 바위 위 핸드폰 화면에 '혜수'의 이름과 부재중 전화 알림이 지시된 클로즈업 앵글로 정확히 연출되었으며, 텍스트 형태가 비교적 더 안정적입니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "앵글, 배경, 조명 등 모든 지시사항을 잘 따랐으나 화면 내 부재중 전화 알림 텍스트의 뭉개짐이 A보다 약간 더 심합니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "젖은 바위 위 핸드폰 화면에 '혜수'의 이름과 부재중 전화 알림이 지시된 클로즈업 앵글로 정확히 연출되었으며, 텍스트 형태가 비교적 더 안정적입니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "앵글, 배경, 조명 등 모든 지시사항을 잘 따랐으나 화면 내 부재중 전화 알림 텍스트의 뭉개짐이 A보다 약간 더 심합니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요청된 앵글과 피사체(핸드폰)의 스케일을 정확히 구현했으며, 혜수의 이름과 부재중 알림이 화면에 잘 나타남."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "A와 거의 동일한 이미지를 생성하였으며, 프레이밍과 화면 내용 등 프롬프트의 요구사항을 충실히 반영함."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요청된 앵글과 피사체(핸드폰)의 스케일을 정확히 구현했으며, 혜수의 이름과 부재중 알림이 화면에 잘 나타남."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "A와 거의 동일한 이미지를 생성하였으며, 프레이밍과 화면 내용 등 프롬프트의 요구사항을 충실히 반영함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh2"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S25sh6::variants": {
  "author_fp": "cb50f0c331bb6858",
  "author": {
   "variants": [
    {
     "approach_ko": "인우의 어깨 너머로 수리영의 서늘하고 절박한 올려다봄을 밀착 포착하는 친밀한 구도.",
     "prompt_en": "Frame an intimate over-In-woo’s-shoulder close shot from Su-riyeong’s eye height, with In-woo held as a near foreground edge and her upturned face in precise focus. Let her cool, desperate gaze carry the frame while the mooring and fishing boat fall into soft fog-muted depth. Use gentle gray dusk light and restrained contrast, preserving damp, tactile surface detail."
    },
    {
     "approach_ko": "측면의 넓은 투샷으로 두 사람의 거리와 밧줄, 안개 낀 계류장의 공간적 긴장을 강조하는 구도.",
     "prompt_en": "Photograph the encounter as a restrained lateral medium-wide two-shot at dock level, clearly showing Su-riyeong’s approach, lifted chin, and upward gaze toward In-woo within the harbor geometry. Use the dock line as a strong foreground element and layer the newly arrived fishing boat and foggy mooring behind them for spatial tension. Keep both figures legible in the cold, desaturated dusk, with soft fog-filtered light and realistic wet materials."
    }
   ]
  },
  "reused": false
 },
 "S25sh6": {
  "input_fingerprint": "b72a49a8a33f7a09",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 밧줄 앞 인우에게 다가선 수리영이 고개를 들어 서늘하고 간절한 눈빛으로 그를 올려다보는 구도.\n\nLOCATION (lock): At the foggy harbor mooring, beside the dock line of the newly arrived small fishing boat. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Gray dusk ambience filtered by dense fog keeps the exchange cold, desaturated, and low in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong approaches In-woo carrying the old photograph, her map-filled backpack, the coordinate note, and the phone used to leave Hye-su a message.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 밧줄 앞 인우에게 다가선 수리영이 고개를 들어 서늘하고 간절한 눈빛으로 그를 올려다보는 구도.\n\nLOCATION (lock): At the foggy harbor mooring, beside the dock line of the newly arrived small fishing boat. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Gray dusk ambience filtered by dense fog keeps the exchange cold, desaturated, and low in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong approaches In-woo carrying the old photograph, her map-filled backpack, the coordinate note, and the phone used to leave Hye-su a message.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nFrame an intimate over-In-woo’s-shoulder close shot from Su-riyeong’s eye height, with In-woo held as a near foreground edge and her upturned face in precise focus. Let her cool, desperate gaze carry the frame while the mooring and fishing boat fall into soft fog-muted depth. Use gentle gray dusk light and restrained contrast, preserving damp, tactile surface detail.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk, foggy after rain.\n\nSHOT TEXT (authoritative, Korean): 밧줄 앞 인우에게 다가선 수리영이 고개를 들어 서늘하고 간절한 눈빛으로 그를 올려다보는 구도.\n\nLOCATION (lock): At the foggy harbor mooring, beside the dock line of the newly arrived small fishing boat. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Gray dusk ambience filtered by dense fog keeps the exchange cold, desaturated, and low in contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong approaches In-woo carrying the old photograph, her map-filled backpack, the coordinate note, and the phone used to leave Hye-su a message.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nPhotograph the encounter as a restrained lateral medium-wide two-shot at dock level, clearly showing Su-riyeong’s approach, lifted chin, and upward gaze toward In-woo within the harbor geometry. Use the dock line as a strong foreground element and layer the newly arrived fishing boat and foggy mooring behind them for spatial tension. Keep both figures legible in the cold, desaturated dusk, with soft fog-filtered light and realistic wet materials.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 8,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "오버더숄더 구도를 통해 인물을 올려다보는 수리영의 간절한 눈빛을 효과적으로 담아냈으며 캐릭터 외형(긴 머리)도 잘 일치합니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "단순한 측면 구도로 인해 눈빛의 감정선이 약화되었고, 소지품이 주머니에 부자연스럽게 꽂혀 있으며 수리영의 머리 길이가 지시사항과 다릅니다."
   }
  ],
  "refs": [
   {
    "label": "STORYBOARD SKETCH — a bare thin-line staging guide, a REFERENCE ONLY: take from it ONLY the camera framing, figure placement, heading and size/depth order. It carries ZERO visual style — every texture, material, light and all realism come from the text and the photographic references. Never let any line-drawing quality leak into the output.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/lane_sketch_S25sh6.png"
   },
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S25sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   },
   {
    "label": "CHARACTER REFERENCE — 인우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:849126>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "어선 선체 측면에 글자가 적혀 있어 텍스트 포함을 금지한 지시를 위반했습니다.",
     "fix_en": "Remove all painted text and numbers from the side of the fishing boat."
    },
    {
     "issue_ko": "바닥 재질이 이전 샷 레퍼런스에 고정된 '어둡고 젖은 바위'가 아니라 스토리보드의 '콘크리트 부두'로 잘못 묘사되었습니다.",
     "fix_en": "Replace the concrete ground surface with the dark, wet, jagged rocks shown in the previous shot still."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove all painted text and numbers from the side of the fishing boat.\n- Replace the concrete ground surface with the dark, wet, jagged rocks shown in the previous shot still.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "오버 더 숄더 구도와 수리영의 시선 및 표정을 정확히 구현했으며, 부두의 환경과 안개 낀 조명도 사실적으로 표현되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "바위 질감이 씬에 부자연스럽게 합성되어 물리적으로 불가능한 공간감을 형성하는 치명적인 오류(콜라주/합성)가 있습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "오버 더 숄더 구도와 수리영의 시선 및 표정을 정확히 구현했으며, 부두의 환경과 안개 낀 조명도 사실적으로 표현되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "바위 질감이 씬에 부자연스럽게 합성되어 물리적으로 불가능한 공간감을 형성하는 치명적인 오류(콜라주/합성)가 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "이전 샷의 젖은 바위 질감을 장소 요소로 정확히 유지했으며, 텍스트 금지 지시를 충실히 따르고 프레이밍도 훌륭합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "바닥을 이전 샷의 바위가 아닌 콘크리트로 묘사하여 장소 일관성을 훼손했으며, 어선 측면에 텍스트가 포함되어 엄격한 금지 규칙을 위반했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 샷의 젖은 바위 질감을 장소 요소로 정확히 유지했으며, 텍스트 금지 지시를 충실히 따르고 프레이밍도 훌륭합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "바닥을 이전 샷의 바위가 아닌 콘크리트로 묘사하여 장소 일관성을 훼손했으며, 어선 측면에 텍스트가 포함되어 엄격한 금지 규칙을 위반했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 11
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "lane(map_marker): 스케치+prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh3"
  }
 },
 "S26sh1::bgfirst_bg": {
  "input_fingerprint": "655366a18209502d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 빛이 사라진 옥탑방 거실 한가운데서 핸드폰을 귀에 댄 혜수의 얼굴에 짙은 그림자가 드리운 안면 클로즈업.\n\nLOCATION (lock): In the darkened living room of the rooftop apartment at dusk, with little remaining light entering the desolate interior.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At Hyesu's eye level and directly before her, hold a static close frame in which her shadowed face occupies most of the center and the phone remains visible against one side of her head. Keep the room reduced to indistinct edge space, isolating her lowered eyeline and listening stillness without introducing a reverse angle.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading dusk ambient leaves Hyesu in deep, low-key shadow with a dark desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 빛이 사라진 옥탑방 거실 한가운데서 핸드폰을 귀에 댄 혜수의 얼굴에 짙은 그림자가 드리운 안면 클로즈업.\n\nLOCATION (lock): In the darkened living room of the rooftop apartment at dusk, with little remaining light entering the desolate interior.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At Hyesu's eye level and directly before her, hold a static close frame in which her shadowed face occupies most of the center and the phone remains visible against one side of her head. Keep the room reduced to indistinct edge space, isolating her lowered eyeline and listening stillness without introducing a reverse angle.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading dusk ambient leaves Hyesu in deep, low-key shadow with a dark desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S26sh1__bgfirst_bg.png",
  "asset_id": "cbdc6991-4f92-47a3-9bf8-0a487ad9a0ea",
  "input_asset_ids": [
   "8b573edd-01cd-4672-9d96-5072435a26ca",
   "2fdbb220-3a0b-43c0-8f34-cc10d5a59a49"
  ]
 },
 "S26sh1": {
  "input_fingerprint": "b210f7bcbe866f37",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 빛이 사라진 옥탑방 거실 한가운데서 핸드폰을 귀에 댄 혜수의 얼굴에 짙은 그림자가 드리운 안면 클로즈업.\n\nLOCATION (lock): In the darkened living room of the rooftop apartment at dusk, with little remaining light entering the desolate interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At Hyesu's eye level and directly before her, hold a static close frame in which her shadowed face occupies most of the center and the phone remains visible against one side of her head. Keep the room reduced to indistinct edge space, isolating her lowered eyeline and listening stillness without introducing a reverse angle.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading dusk ambient leaves Hyesu in deep, low-key shadow with a dark desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains ransacked and desolate, with its open window, overturned chairs, and the red circle left bare after the photograph was removed. Hye-su holds the phone playing Su-riyeong’s recorded message.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 빛이 사라진 옥탑방 거실 한가운데서 핸드폰을 귀에 댄 혜수의 얼굴에 짙은 그림자가 드리운 안면 클로즈업.\n\nLOCATION (lock): In the darkened living room of the rooftop apartment at dusk, with little remaining light entering the desolate interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At Hyesu's eye level and directly before her, hold a static close frame in which her shadowed face occupies most of the center and the phone remains visible against one side of her head. Keep the room reduced to indistinct edge space, isolating her lowered eyeline and listening stillness without introducing a reverse angle.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading dusk ambient leaves Hyesu in deep, low-key shadow with a dark desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains ransacked and desolate, with its open window, overturned chairs, and the red circle left bare after the photograph was removed. Hye-su holds the phone playing Su-riyeong’s recorded message.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 빛이 사라진 옥탑방 거실 한가운데서 핸드폰을 귀에 댄 혜수의 얼굴에 짙은 그림자가 드리운 안면 클로즈업.\n\nLOCATION (lock): In the darkened living room of the rooftop apartment at dusk, with little remaining light entering the desolate interior. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At Hyesu's eye level and directly before her, hold a static close frame in which her shadowed face occupies most of the center and the phone remains visible against one side of her head. Keep the room reduced to indistinct edge space, isolating her lowered eyeline and listening stillness without introducing a reverse angle.\n- FRAMING SCALE: close-up\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading dusk ambient leaves Hyesu in deep, low-key shadow with a dark desaturated palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The apartment remains ransacked and desolate, with its open window, overturned chairs, and the red circle left bare after the photograph was removed. Hye-su holds the phone playing Su-riyeong’s recorded message.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 혜수 (한국인 40대 초반 여성 얼굴, 성숙한 얼굴선, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S26sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S26sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:900092>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "배경과 인물 묘사는 우수하나, 요구된 안면 클로즈업보다 프레임이 넓어 '배경을 흐릿한 가장자리 공간으로 줄이라'는 지시를 위반했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "얼굴이 화면 대부분을 차지하는 타이트한 안면 클로즈업, 시선 처리, 아웃포커싱된 배경 등 프레이밍 지시를 정확히 따랐습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "배경과 인물 묘사는 우수하나, 요구된 안면 클로즈업보다 프레임이 넓어 '배경을 흐릿한 가장자리 공간으로 줄이라'는 지시를 위반했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "얼굴이 화면 대부분을 차지하는 타이트한 안면 클로즈업, 시선 처리, 아웃포커싱된 배경 등 프레이밍 지시를 정확히 따랐습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 안면 클로즈업 앵글을 정확히 구현하여 얼굴을 화면 중심에 채우고 배경을 흐릿하게 처리한 점이 훌륭합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 안면 클로즈업보다 화각이 넓고 배경이 너무 선명하게 드러나 '배경을 흐릿한 여백으로 두라'는 지침을 어겼습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 안면 클로즈업 앵글을 정확히 구현하여 얼굴을 화면 중심에 채우고 배경을 흐릿하게 처리한 점이 훌륭합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "요구된 안면 클로즈업보다 화각이 넓고 배경이 너무 선명하게 드러나 '배경을 흐릿한 여백으로 두라'는 지침을 어겼습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 9,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "배경과 인물 묘사는 우수하나, 요구된 안면 클로즈업보다 프레임이 넓어 '배경을 흐릿한 가장자리 공간으로 줄이라'는 지시를 위반했습니다."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "얼굴이 화면 대부분을 차지하는 타이트한 안면 클로즈업, 시선 처리, 아웃포커싱된 배경 등 프레이밍 지시를 정확히 따랐습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L06B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 혜수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:900092>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "핸드폰이 뒤집혀 있어 화면이 귀가 아닌 바깥쪽(카메라 쪽)을 향하고 있습니다.",
     "fix_en": "Flip the phone so its screen faces inward against her ear and the back of the phone faces outward."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Flip the phone so its screen faces inward against her ear and the back of the phone faces outward.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 안면 클로즈업, 시선 방향, 어두운 조명 및 통화 동작을 지시사항에 맞게 완벽히 구현했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "후보 A와 동일한 이미지로, 요구된 구도와 조명, 캐릭터의 디테일을 모두 정확하게 반영했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 안면 클로즈업, 시선 방향, 어두운 조명 및 통화 동작을 지시사항에 맞게 완벽히 구현했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "후보 A와 동일한 이미지로, 요구된 구도와 조명, 캐릭터의 디테일을 모두 정확하게 반영했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트에 명시된 안면 클로즈업, 짙은 그림자, 낮아진 시선 및 지정된 캐릭터의 외모와 공간적 배경을 매우 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "A와 동일한 이미지로, 프롬프트의 카메라 구도, 조명, 인물의 외양 및 전화기를 든 동작을 충실하게 반영했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "프롬프트에 명시된 안면 클로즈업, 짙은 그림자, 낮아진 시선 및 지정된 캐릭터의 외모와 공간적 배경을 매우 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "A와 동일한 이미지로, 프롬프트의 카메라 구도, 조명, 인물의 외양 및 전화기를 든 동작을 충실하게 반영했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S26sh1__bgfirst_bg.png",
   "bg_asset_id": "cbdc6991-4f92-47a3-9bf8-0a487ad9a0ea",
   "bg_record_key": "S26sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B02",
   "choice": "Candidate C",
   "confident": true,
   "reason_ko": "지문에서 '옥탑방 거실 한가운데'를 명시하고 있으므로, 어질러진 거실 공간 전체가 보이는 Candidate C가 가장 적합합니다.",
   "kept": "L06B02"
  }
 },
 "S26sh4::variants": {
  "author_fp": "1e3fbd97cfb87dae",
  "author": {
   "variants": [
    {
     "approach_ko": "정면 대칭과 깊은 초점으로 빈 현관 전체를 냉정한 정물처럼 포착한다.",
     "prompt_en": "Use rigid, nearly symmetrical architectural framing with the doorway planes held square and the narrow opening precisely centered. Keep the door, frame, and adjacent wall in clear deep focus, letting the empty geometry and surrounding negative space dominate. Render the minimal dusk ambience as soft, directionless illumination with compressed shadows and a cold, desaturated stillness."
    },
    {
     "approach_ko": "문 가장자리와 가느다란 틈에 선택적으로 초점을 모아 바람이 남긴 불안을 강조한다.",
     "prompt_en": "Concentrate attention on the thin door edge and the sliver of separation, with selective focus allowing the surrounding doorway and wall to recede gently while preserving their physical scale. Let the faint dusk ambient graze the worn metal edge and barely distinguish the opening from the lightless interior. Favor dense blacks, subtle surface texture, and restrained tonal separation for a tense, suspended still-life image."
    }
   ]
  },
  "reused": false
 },
 "S26sh4": {
  "input_fingerprint": "8005b729edbc7737",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어두운 현관문이 바람에 밀려 미세하게 틈새를 벌린 채 멈춘 스산한 정물 구도.\n\nLOCATION (lock): At the interior side of the rooftop apartment’s entrance, where the dark metal door shifts in the wind. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height inside the room, face the entrance squarely and hold a static medium view, placing the narrow opening on the central vertical axis while retaining enough adjacent doorway and wall for scale. The door occupies less than two-fifths of the frame, and no person intrudes into the final composition.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: slightly open entrance door in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 현관문 (Held slightly open after moving in the wind) — The room-facing side is presented nearly front-on, with its edge defining the narrow opening; used as Central static focal point and spatial endpoint of the scene.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The lightless room is rendered with minimal dusk ambient, restrained contrast, and a dark desaturated finish.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ransacked apartment remains desolate, and the red wall circle remains without its photograph as the unsecured entrance door rattles in the wind.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어두운 현관문이 바람에 밀려 미세하게 틈새를 벌린 채 멈춘 스산한 정물 구도.\n\nLOCATION (lock): At the interior side of the rooftop apartment’s entrance, where the dark metal door shifts in the wind. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height inside the room, face the entrance squarely and hold a static medium view, placing the narrow opening on the central vertical axis while retaining enough adjacent doorway and wall for scale. The door occupies less than two-fifths of the frame, and no person intrudes into the final composition.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: slightly open entrance door in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 현관문 (Held slightly open after moving in the wind) — The room-facing side is presented nearly front-on, with its edge defining the narrow opening; used as Central static focal point and spatial endpoint of the scene.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The lightless room is rendered with minimal dusk ambient, restrained contrast, and a dark desaturated finish.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ransacked apartment remains desolate, and the red wall circle remains without its photograph as the unsecured entrance door rattles in the wind.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse rigid, nearly symmetrical architectural framing with the doorway planes held square and the narrow opening precisely centered. Keep the door, frame, and adjacent wall in clear deep focus, letting the empty geometry and surrounding negative space dominate. Render the minimal dusk ambience as soft, directionless illumination with compressed shadows and a cold, desaturated stillness.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 어두운 현관문이 바람에 밀려 미세하게 틈새를 벌린 채 멈춘 스산한 정물 구도.\n\nLOCATION (lock): At the interior side of the rooftop apartment’s entrance, where the dark metal door shifts in the wind. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height inside the room, face the entrance squarely and hold a static medium view, placing the narrow opening on the central vertical axis while retaining enough adjacent doorway and wall for scale. The door occupies less than two-fifths of the frame, and no person intrudes into the final composition.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: slightly open entrance door in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 현관문 (Held slightly open after moving in the wind) — The room-facing side is presented nearly front-on, with its edge defining the narrow opening; used as Central static focal point and spatial endpoint of the scene.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The lightless room is rendered with minimal dusk ambient, restrained contrast, and a dark desaturated finish.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The ransacked apartment remains desolate, and the red wall circle remains without its photograph as the unsecured entrance door rattles in the wind.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nConcentrate attention on the thin door edge and the sliver of separation, with selective focus allowing the surrounding doorway and wall to recede gently while preserving their physical scale. Let the faint dusk ambient graze the worn metal edge and barely distinguish the opening from the lightless interior. Favor dense blacks, subtle surface texture, and restrained tonal separation for a tense, suspended still-life image.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 8,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "지시된 구도와 황혼의 조명을 잘 구현했으며, 빈 붉은 원의 디테일을 프롬프트에 맞게 정확히 표현했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "붉은 원 안에 사진(또는 종이) 형태가 남아 있어 '빈 붉은 원'이라는 구체적인 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S26sh1_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "문이 '미세하게 틈새를 벌린' 상태가 아니라 너무 넓게 열려 있으며, 틈새가 화면 중앙에 위치하지 않습니다.",
     "fix_en": "Close the door so it is only slightly open, leaving just a narrow vertical gap exactly in the center of the frame."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Close the door to a very narrow crack and position that crack precisely on the frame’s central vertical axis, keeping the door face nearly front-on and everything else unchanged.",
     "issue_ko": "문이 너무 크게 열려 있어 문짝 면이 비스듬히 보이며, 좁은 틈도 화면 중앙 수직축이 아니라 오른쪽에 놓여 있다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Close the door so it is only slightly open, leaving just a narrow vertical gap exactly in the center of the frame.\n- Close the door to a very narrow crack and position that crack precisely on the frame’s central vertical axis, keeping the door face nearly front-on and everything else unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "중앙에 살짝 열린 문, 빈 붉은 원, 스산한 저녁 분위기 등 프롬프트의 요구사항을 프레이밍과 함께 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "A와 동일한 이미지로, 요구된 카메라 구도와 환경 요소들을 완벽히 반영한 정물 구도를 보여줍니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "중앙에 살짝 열린 문, 빈 붉은 원, 스산한 저녁 분위기 등 프롬프트의 요구사항을 프레이밍과 함께 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "A와 동일한 이미지로, 요구된 카메라 구도와 환경 요소들을 완벽히 반영한 정물 구도를 보여줍니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 '어둡고 채도가 낮은(dark desaturated finish)' 조명 분위기를 더 충실히 구현하여 스산한 정물 구도의 느낌을 잘 살렸습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "구도와 프레임 구성은 매우 정확하나, 조명과 색감이 지시사항의 '최소한의 빛과 낮은 채도'에 비해 약간 밝고 선명합니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 '어둡고 채도가 낮은(dark desaturated finish)' 조명 분위기를 더 충실히 구현하여 스산한 정물 구도의 느낌을 잘 살렸습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "구도와 프레임 구성은 매우 정확하나, 조명과 색감이 지시사항의 '최소한의 빛과 낮은 채도'에 비해 약간 밝고 선명합니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 13
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S26sh1"
  },
  "lane_policy": "share_plan_prev_bgonly",
  "plate_select": {
   "candidates": {
    "A": "L06B01",
    "B": "L06B02",
    "C": "L06B03",
    "D": "L06B04",
    "E": "L06B05",
    "F": "L06B06",
    "G": "L06B07",
    "H": "L06B08"
   },
   "assigned": "L06B02",
   "choice": "Candidate B",
   "confident": true,
   "reason_ko": "\"어두운 현관문이 바람에 밀려 미세하게 틈새를 벌린 채\"라는 지문에 맞춰, 옥탑방 외부로 이어지는 짙은 색 철제 현관문이 열려 있는 공간입니다.",
   "kept": "L06B02"
  }
 },
 "S27sh1::bgfirst_bg": {
  "input_fingerprint": "b6827179738fd98b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 해질 무렵 황금빛이 스며드는 어선 갑판 위에 홀로 앉아 무릎 위에 펼쳐진 지도를 내려다보는 수리영의 정지 구도.\n\nLOCATION (lock): On the open deck of a small fishing boat at sea, sitting away from the wheelhouse in the fading golden light.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level and a moderate distance in front of 수리영, hold a static wide composition with her centered on the deck and the unfolded map clearly readable as a shape across her knees. Her seated figure remains the anchor while the surrounding deck and sea establish her isolation aboard the small fishing boat.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 소형어선 갑판 (수리영 is seated alone on it); used as Supports 수리영 and establishes the open working space around her; 바다 (The small fishing boat is afloat far from its departure point); used as Surrounding background that emphasizes the boat's distance from land; 펼쳐진 지도 (Unfolded across her knees) — Its marked face is angled upward toward 수리영 and remains obliquely visible to the camera; used as Visible across 수리영's knees as the immediate visual focus beneath her face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Golden dusk light reaches the deck and sea through a muted, desaturated grade with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 해질 무렵 황금빛이 스며드는 어선 갑판 위에 홀로 앉아 무릎 위에 펼쳐진 지도를 내려다보는 수리영의 정지 구도.\n\nLOCATION (lock): On the open deck of a small fishing boat at sea, sitting away from the wheelhouse in the fading golden light.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level and a moderate distance in front of 수리영, hold a static wide composition with her centered on the deck and the unfolded map clearly readable as a shape across her knees. Her seated figure remains the anchor while the surrounding deck and sea establish her isolation aboard the small fishing boat.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 소형어선 갑판 (수리영 is seated alone on it); used as Supports 수리영 and establishes the open working space around her; 바다 (The small fishing boat is afloat far from its departure point); used as Surrounding background that emphasizes the boat's distance from land; 펼쳐진 지도 (Unfolded across her knees) — Its marked face is angled upward toward 수리영 and remains obliquely visible to the camera; used as Visible across 수리영's knees as the immediate visual focus beneath her face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Golden dusk light reaches the deck and sea through a muted, desaturated grade with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S27sh1__bgfirst_bg.png",
  "asset_id": "158d12db-da1a-4533-bfe8-dad349f440a0",
  "input_asset_ids": [
   "a7f412bf-8b8d-43cf-9846-c16dcb8631b3",
   "a87bbc87-593d-4708-bf53-8728b60affbe"
  ]
 },
 "S27sh1": {
  "input_fingerprint": "46ed2cbde54d315b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 해질 무렵 황금빛이 스며드는 어선 갑판 위에 홀로 앉아 무릎 위에 펼쳐진 지도를 내려다보는 수리영의 정지 구도.\n\nLOCATION (lock): On the open deck of a small fishing boat at sea, sitting away from the wheelhouse in the fading golden light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level and a moderate distance in front of 수리영, hold a static wide composition with her centered on the deck and the unfolded map clearly readable as a shape across her knees. Her seated figure remains the anchor while the surrounding deck and sea establish her isolation aboard the small fishing boat.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 소형어선 갑판 (수리영 is seated alone on it); used as Supports 수리영 and establishes the open working space around her; 바다 (The small fishing boat is afloat far from its departure point); used as Surrounding background that emphasizes the boat's distance from land; 펼쳐진 지도 (Unfolded across her knees) — Its marked face is angled upward toward 수리영 and remains obliquely visible to the camera; used as Visible across 수리영's knees as the immediate visual focus beneath her face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Golden dusk light reaches the deck and sea through a muted, desaturated grade with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A printed map from Su-riyeong’s backpack is spread on her lap, while she retains the old photograph, the remaining folded maps, the coordinate note, and her phone aboard the boat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 해질 무렵 황금빛이 스며드는 어선 갑판 위에 홀로 앉아 무릎 위에 펼쳐진 지도를 내려다보는 수리영의 정지 구도.\n\nLOCATION (lock): On the open deck of a small fishing boat at sea, sitting away from the wheelhouse in the fading golden light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level and a moderate distance in front of 수리영, hold a static wide composition with her centered on the deck and the unfolded map clearly readable as a shape across her knees. Her seated figure remains the anchor while the surrounding deck and sea establish her isolation aboard the small fishing boat.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 소형어선 갑판 (수리영 is seated alone on it); used as Supports 수리영 and establishes the open working space around her; 바다 (The small fishing boat is afloat far from its departure point); used as Surrounding background that emphasizes the boat's distance from land; 펼쳐진 지도 (Unfolded across her knees) — Its marked face is angled upward toward 수리영 and remains obliquely visible to the camera; used as Visible across 수리영's knees as the immediate visual focus beneath her face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Golden dusk light reaches the deck and sea through a muted, desaturated grade with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A printed map from Su-riyeong’s backpack is spread on her lap, while she retains the old photograph, the remaining folded maps, the coordinate note, and her phone aboard the boat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 해질 무렵 황금빛이 스며드는 어선 갑판 위에 홀로 앉아 무릎 위에 펼쳐진 지도를 내려다보는 수리영의 정지 구도.\n\nLOCATION (lock): On the open deck of a small fishing boat at sea, sitting away from the wheelhouse in the fading golden light. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated eye level and a moderate distance in front of 수리영, hold a static wide composition with her centered on the deck and the unfolded map clearly readable as a shape across her knees. Her seated figure remains the anchor while the surrounding deck and sea establish her isolation aboard the small fishing boat.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 소형어선 갑판 (수리영 is seated alone on it); used as Supports 수리영 and establishes the open working space around her; 바다 (The small fishing boat is afloat far from its departure point); used as Surrounding background that emphasizes the boat's distance from land; 펼쳐진 지도 (Unfolded across her knees) — Its marked face is angled upward toward 수리영 and remains obliquely visible to the camera; used as Visible across 수리영's knees as the immediate visual focus beneath her face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Golden dusk light reaches the deck and sea through a muted, desaturated grade with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): A printed map from Su-riyeong’s backpack is spread on her lap, while she retains the old photograph, the remaining folded maps, the coordinate note, and her phone aboard the boat.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S27sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S27sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    },
    {
     "label": "PROP REFERENCE — 접힌 종이 해안 지도: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1283913>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L19B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    },
    {
     "label": "PROP REFERENCE — 접힌 종이 해안 지도: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1283913>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 샷 크기와 구도를 잘 따랐으며, 배낭, 휴대폰, 사진 등 명시된 소지품들을 갑판 위에 충실히 배치했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "지도의 앞면이 카메라를 향해 들려 있어 인물이 지도를 보는 시선 방향과 모순되며, 요구된 주변 소지품들이 대부분 누락되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 샷 크기와 구도를 잘 따랐으며, 배낭, 휴대폰, 사진 등 명시된 소지품들을 갑판 위에 충실히 배치했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "지도의 앞면이 카메라를 향해 들려 있어 인물이 지도를 보는 시선 방향과 모순되며, 요구된 주변 소지품들이 대부분 누락되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "와이드 샷의 구도와 해질녘의 조명 분위기를 잘 살렸으며, 배낭, 사진, 전화기 등 명시된 소품들을 갑판에 정확하게 배치해 지침을 훌륭히 따랐습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지도 하단에 의미 없는 문자가 생성되었고 인물 옆에 정체불명의 덩어리가 있으며, 휴대해야 할 소품들이 모두 누락되었습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "와이드 샷의 구도와 해질녘의 조명 분위기를 잘 살렸으며, 배낭, 사진, 전화기 등 명시된 소품들을 갑판에 정확하게 배치해 지침을 훌륭히 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지도 하단에 의미 없는 문자가 생성되었고 인물 옆에 정체불명의 덩어리가 있으며, 휴대해야 할 소품들이 모두 누락되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 7
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 14,
   "B": 7
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지정된 샷 크기와 구도를 잘 따랐으며, 배낭, 휴대폰, 사진 등 명시된 소지품들을 갑판 위에 충실히 배치했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "지도의 앞면이 카메라를 향해 들려 있어 인물이 지도를 보는 시선 방향과 모순되며, 요구된 주변 소지품들이 대부분 누락되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L19B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   },
   {
    "label": "PROP REFERENCE — 접힌 종이 해안 지도: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1283913>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영의 시선이 무릎 위의 지도를 향하지 않고 정면을 바라보고 있습니다.",
     "fix_en": "Tilt Su-riyeong's head and lower her gaze so she is looking directly at the map on her lap."
    },
    {
     "issue_ko": "무릎 위에 펼쳐진 지도의 디자인이 레퍼런스의 해안 지도(그리드 선, 지형 모양)와 전혀 일치하지 않습니다.",
     "fix_en": "Replace the map's surface image to match the precise grid layout and coastal landmass shapes of the provided prop reference."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Move Su-riyeong lower and slightly closer to the camera, enlarging her seated figure and the map across her knees to match the layout sketch while keeping her centered and preserving the background framing.",
     "issue_ko": "수리영이 레이아웃 스케치보다 화면 위쪽의 먼 지점에 작게 배치되어 있으며, 무릎 위 지도도 함께 너무 작게 보인다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Tilt Su-riyeong's head and lower her gaze so she is looking directly at the map on her lap.\n- Replace the map's surface image to match the precise grid layout and coastal landmass shapes of the provided prop reference.\n- Move Su-riyeong lower and slightly closer to the camera, enlarging her seated figure and the map across her knees to match the layout sketch while keeping her centered and preserving the background framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "배경, 의상, 소품 배치는 전반적으로 정확히 구현되었으나, 지도를 내려다봐야 한다는 프롬프트의 시선 지시를 어기고 정면을 응시한 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지도를 내려다보는 시선은 맞으나, 인물의 하반신(다리)이 누락되고 지도가 갑판 바닥에 평면적으로 붙어있는 물리적 구조 오류(Hard Violation)가 발생했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "배경, 의상, 소품 배치는 전반적으로 정확히 구현되었으나, 지도를 내려다봐야 한다는 프롬프트의 시선 지시를 어기고 정면을 응시한 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지도를 내려다보는 시선은 맞으나, 인물의 하반신(다리)이 누락되고 지도가 갑판 바닥에 평면적으로 붙어있는 물리적 구조 오류(Hard Violation)가 발생했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "스케치에 지정된 양반다리 자세와 전체적인 구도를 훌륭하게 구현했으나, 펼쳐진 지도가 주어진 소품 레퍼런스와 다르게 렌더링되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "하반신이 갑판 아래로 사라지거나 잘린 듯한 물리적으로 불가능한 형태의 심각한 오류가 있으며, 지도 소품 또한 레퍼런스와 일치하지 않습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "스케치에 지정된 양반다리 자세와 전체적인 구도를 훌륭하게 구현했으나, 펼쳐진 지도가 주어진 소품 레퍼런스와 다르게 렌더링되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "하반신이 갑판 아래로 사라지거나 잘린 듯한 물리적으로 불가능한 형태의 심각한 오류가 있으며, 지도 소품 또한 레퍼런스와 일치하지 않습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S27sh1__bgfirst_bg.png",
   "bg_asset_id": "158d12db-da1a-4533-bfe8-dad349f440a0",
   "bg_record_key": "S27sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S27sh5::variants": {
  "author_fp": "a53e572296c78247",
  "author": {
   "variants": [
    {
     "approach_ko": "정면에 가까운 고정 클로즈업으로 인우의 굳은 표정과 아래로 박힌 시선을 압박감 있게 포착한다.",
     "prompt_en": "Hold a nearly frontal, eye-level close-up with In-woo’s face dominating the center and the photograph reduced to a soft, partial cue along the lower edge. Keep the focal plane firmly on his fixed downward eyes and rigid facial muscles, with the deck dissolving into subdued blur. Let the fading golden dusk fall softly across one side of his face, rendering his shock through drained color, restrained contrast, and complete physical stillness."
    },
    {
     "approach_ko": "비스듬한 대각 구도와 전경의 사진 가장자리로 인우의 당황한 얼굴을 갇힌 듯 강조한다.",
     "prompt_en": "Use the slight frontal offset to build an asymmetric diagonal close-up, placing In-woo’s blanched face heavily in the frame while the small, out-of-focus edge of the photograph rises beneath his eyeline. Crop tightly enough that his locked gaze, tense mouth, and frozen jaw carry the entire beat, while the shallow-focus deck recedes as a narrow contextual layer. Shape the subdued dusk as gentle grazing light across his features, with muted warmth around the edges and cooler, desaturated facial tones intensifying the stunned reaction."
    }
   ]
  },
  "reused": false
 },
 "S27sh5": {
  "input_fingerprint": "faabf9f911ea74f3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 사진에 시선을 고정한 채 안색이 백지장처럼 굳어진 인우의 당황한 안면.\n\nLOCATION (lock): On the open deck of the small fishing boat, where the boatman examines the old photograph in the evening light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut to 인우 at eye level from a slight frontal offset, tightening until his blanched face occupies most of the frame while the photograph remains only as a small lower-edge cue. Use shallow focus to hold his fixed downward eyeline and rigid expression, leaving the deck context subdued behind him.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사진 (Held in front of 인우 for inspection) — The image-bearing front face tilts up toward 인우, with only part of it visible to the camera; used as A small lower-frame eyeline cue beneath 인우's face; 소형어선 갑판 (Visible only outside the shallow focal plane); used as Soft contextual background behind the close reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading golden dusk is kept subdued and desaturated, allowing 인우's pale reaction to register without brightening the surrounding deck.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the weathered deck, golden dusk light, sea atmosphere, and ongoing presence of the unfolded map from the reference. Exclude the seated young woman from the frame; isolate the young man's stunned face as he stares at the photograph.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is showing In-woo the same old photograph taken from the red circle; her map-filled backpack, coordinate note, and phone remain aboard with her.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 사진에 시선을 고정한 채 안색이 백지장처럼 굳어진 인우의 당황한 안면.\n\nLOCATION (lock): On the open deck of the small fishing boat, where the boatman examines the old photograph in the evening light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut to 인우 at eye level from a slight frontal offset, tightening until his blanched face occupies most of the frame while the photograph remains only as a small lower-edge cue. Use shallow focus to hold his fixed downward eyeline and rigid expression, leaving the deck context subdued behind him.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사진 (Held in front of 인우 for inspection) — The image-bearing front face tilts up toward 인우, with only part of it visible to the camera; used as A small lower-frame eyeline cue beneath 인우's face; 소형어선 갑판 (Visible only outside the shallow focal plane); used as Soft contextual background behind the close reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading golden dusk is kept subdued and desaturated, allowing 인우's pale reaction to register without brightening the surrounding deck.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the weathered deck, golden dusk light, sea atmosphere, and ongoing presence of the unfolded map from the reference. Exclude the seated young woman from the frame; isolate the young man's stunned face as he stares at the photograph.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is showing In-woo the same old photograph taken from the red circle; her map-filled backpack, coordinate note, and phone remain aboard with her.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nHold a nearly frontal, eye-level close-up with In-woo’s face dominating the center and the photograph reduced to a soft, partial cue along the lower edge. Keep the focal plane firmly on his fixed downward eyes and rigid facial muscles, with the deck dissolving into subdued blur. Let the fading golden dusk fall softly across one side of his face, rendering his shock through drained color, restrained contrast, and complete physical stillness.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 사진에 시선을 고정한 채 안색이 백지장처럼 굳어진 인우의 당황한 안면.\n\nLOCATION (lock): On the open deck of the small fishing boat, where the boatman examines the old photograph in the evening light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Cut to 인우 at eye level from a slight frontal offset, tightening until his blanched face occupies most of the frame while the photograph remains only as a small lower-edge cue. Use shallow focus to hold his fixed downward eyeline and rigid expression, leaving the deck context subdued behind him.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사진 (Held in front of 인우 for inspection) — The image-bearing front face tilts up toward 인우, with only part of it visible to the camera; used as A small lower-frame eyeline cue beneath 인우's face; 소형어선 갑판 (Visible only outside the shallow focal plane); used as Soft contextual background behind the close reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Fading golden dusk is kept subdued and desaturated, allowing 인우's pale reaction to register without brightening the surrounding deck.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the weathered deck, golden dusk light, sea atmosphere, and ongoing presence of the unfolded map from the reference. Exclude the seated young woman from the frame; isolate the young man's stunned face as he stares at the photograph.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong is showing In-woo the same old photograph taken from the red circle; her map-filled backpack, coordinate note, and phone remain aboard with her.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse the slight frontal offset to build an asymmetric diagonal close-up, placing In-woo’s blanched face heavily in the frame while the small, out-of-focus edge of the photograph rises beneath his eyeline. Crop tightly enough that his locked gaze, tense mouth, and frozen jaw carry the entire beat, while the shallow-focus deck recedes as a narrow contextual layer. Shape the subdued dusk as gentle grazing light across his features, with muted warmth around the edges and cooler, desaturated facial tones intensifying the stunned reaction.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 10,
   "B": 8
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 10,
    "verdict_ko": "지정된 클로즈업 프레이밍을 정확히 구현하였으며, 이전 샷의 노을 지는 배경 조명과 창백해진 인물의 피부톤(백지장처럼 굳어진 안색) 대비를 훌륭하게 표현함."
   },
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "프레이밍과 표정은 지시사항을 잘 따랐으나, 배경의 노을빛(황금빛 해질녘)이 너무 어둡고 회색조로 렌더링되어 이전 샷과의 조명 일관성이 떨어짐."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S27sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 인우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:849126>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "사진의 앞면(이미지)이 인우의 시선이 아닌 카메라를 향하고 있어 사물 방향 지시를 위반했습니다.",
     "fix_en": "Tilt the photograph so its image-bearing side faces In-woo's downward gaze, exposing mostly its back or top edge to the camera."
    },
    {
     "issue_ko": "프롬프트에서 유지하라고 지시한 '펼쳐진 지도'가 배경의 갑판 위에 보이지 않습니다.",
     "fix_en": "Add the unfolded map on the green deck in the background, matching its position from the previous shot still."
    }
   ]
  },
  "composition_critique": {
   "issues": []
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Tilt the photograph so its image-bearing side faces In-woo's downward gaze, exposing mostly its back or top edge to the camera.\n- Add the unfolded map on the green deck in the background, matching its position from the previous shot still.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 클로즈업 앵글과 인물의 굳은 표정을 훌륭하게 구현했으며, 사진의 앞면 일부가 하단에 자연스럽게 노출되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물 뒤에 지도가 물리적으로 불가능한 평면 그래픽(스티커) 형태로 합성되어 심각한 규정 위반(Hard Violation)입니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 클로즈업 앵글과 인물의 굳은 표정을 훌륭하게 구현했으며, 사진의 앞면 일부가 하단에 자연스럽게 노출되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물 뒤에 지도가 물리적으로 불가능한 평면 그래픽(스티커) 형태로 합성되어 심각한 규정 위반(Hard Violation)입니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요청된 클로즈업 프레임과 시선 처리, 창백한 표정을 훌륭하게 구현했으며 배경도 어색함 없이 자연스럽습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "인물의 어깨 뒤로 지도가 허공에 둥둥 떠 있는 심각한 물리적, 시각적 오류(Hard Violation)가 발생했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요청된 클로즈업 프레임과 시선 처리, 창백한 표정을 훌륭하게 구현했으며 배경도 어색함 없이 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "인물의 어깨 뒤로 지도가 허공에 둥둥 떠 있는 심각한 물리적, 시각적 오류(Hard Violation)가 발생했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 16,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh1"
  }
 },
 "S28sh4::bgfirst_bg": {
  "input_fingerprint": "0ab058d42f9d875c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수리영의 검지손가락 끝이 낡은 사진 속 한 아이의 얼굴을 꾹 짚고 있는 클로즈업.\n\nLOCATION (lock): Inside the small fishing boat’s wheelhouse, over the old photograph being held between the two passengers.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height, look down at a shallow angle as the camera completes its inward move toward 수리영's index fingertip pressing the child's pictured face. Keep the fingertip and face together near center, with the photograph occupying less than two-fifths of the frame and its edges plus part of 수리영's hand preserving physical scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사진 (Held between 수리영 and 인우 while 수리영 identifies the child) — The image-bearing front face is aimed up toward the lens, clearly showing the child's face under the fingertip; used as Primary evidence surface beneath 수리영's fingertip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low dusk ambient in the wheelhouse is treated with desaturated low-key contrast, keeping the photographed face and fingertip legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수리영의 검지손가락 끝이 낡은 사진 속 한 아이의 얼굴을 꾹 짚고 있는 클로즈업.\n\nLOCATION (lock): Inside the small fishing boat’s wheelhouse, over the old photograph being held between the two passengers.\n\nTIME OF DAY (lock): dusk.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height, look down at a shallow angle as the camera completes its inward move toward 수리영's index fingertip pressing the child's pictured face. Keep the fingertip and face together near center, with the photograph occupying less than two-fifths of the frame and its edges plus part of 수리영's hand preserving physical scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사진 (Held between 수리영 and 인우 while 수리영 identifies the child) — The image-bearing front face is aimed up toward the lens, clearly showing the child's face under the fingertip; used as Primary evidence surface beneath 수리영's fingertip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low dusk ambient in the wheelhouse is treated with desaturated low-key contrast, keeping the photographed face and fingertip legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. No text or watermarks.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S28sh4__bgfirst_bg.png",
  "asset_id": "74303897-861d-49f5-a444-677e1f7df882",
  "input_asset_ids": [
   "1c300a5d-6535-4a9f-af27-4651aa00fcb0",
   "5ea06a22-449a-4ff7-88f1-7c32512ca98d"
  ]
 },
 "S28sh4": {
  "input_fingerprint": "f2fe8accc2d158a4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 수리영의 검지손가락 끝이 낡은 사진 속 한 아이의 얼굴을 꾹 짚고 있는 클로즈업.\n\nLOCATION (lock): Inside the small fishing boat’s wheelhouse, over the old photograph being held between the two passengers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height, look down at a shallow angle as the camera completes its inward move toward 수리영's index fingertip pressing the child's pictured face. Keep the fingertip and face together near center, with the photograph occupying less than two-fifths of the frame and its edges plus part of 수리영's hand preserving physical scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사진 (Held between 수리영 and 인우 while 수리영 identifies the child) — The image-bearing front face is aimed up toward the lens, clearly showing the child's face under the fingertip; used as Primary evidence surface beneath 수리영's fingertip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low dusk ambient in the wheelhouse is treated with desaturated low-key contrast, keeping the photographed face and fingertip legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong still holds the old photograph and points directly to the child she identifies as Minsuk; her maps and other belongings remain aboard.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 수리영의 검지손가락 끝이 낡은 사진 속 한 아이의 얼굴을 꾹 짚고 있는 클로즈업.\n\nLOCATION (lock): Inside the small fishing boat’s wheelhouse, over the old photograph being held between the two passengers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height, look down at a shallow angle as the camera completes its inward move toward 수리영's index fingertip pressing the child's pictured face. Keep the fingertip and face together near center, with the photograph occupying less than two-fifths of the frame and its edges plus part of 수리영's hand preserving physical scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사진 (Held between 수리영 and 인우 while 수리영 identifies the child) — The image-bearing front face is aimed up toward the lens, clearly showing the child's face under the fingertip; used as Primary evidence surface beneath 수리영's fingertip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low dusk ambient in the wheelhouse is treated with desaturated low-key contrast, keeping the photographed face and fingertip legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong still holds the old photograph and points directly to the child she identifies as Minsuk; her maps and other belongings remain aboard.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 수리영의 검지손가락 끝이 낡은 사진 속 한 아이의 얼굴을 꾹 짚고 있는 클로즈업.\n\nLOCATION (lock): Inside the small fishing boat’s wheelhouse, over the old photograph being held between the two passengers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height, look down at a shallow angle as the camera completes its inward move toward 수리영's index fingertip pressing the child's pictured face. Keep the fingertip and face together near center, with the photograph occupying less than two-fifths of the frame and its edges plus part of 수리영's hand preserving physical scale.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 사진 (Held between 수리영 and 인우 while 수리영 identifies the child) — The image-bearing front face is aimed up toward the lens, clearly showing the child's face under the fingertip; used as Primary evidence surface beneath 수리영's fingertip.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low dusk ambient in the wheelhouse is treated with desaturated low-key contrast, keeping the photographed face and fingertip legible.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Su-riyeong still holds the old photograph and points directly to the child she identifies as Minsuk; her maps and other belongings remain aboard.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S28sh4__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/conti/conti_S28sh4.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    },
    {
     "label": "PROP REFERENCE — 두 한국인 어린 소녀가 찍힌 낡은 사진: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1242961>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L20B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:853090>"
    },
    {
     "label": "PROP REFERENCE — 두 한국인 어린 소녀가 찍힌 낡은 사진: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1242961>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지시된 클로즈업보다 넓은 미디엄 샷으로 인물과 배경을 과하게 노출했으나, 두 사람이 사진을 함께 들고 손가락으로 아이를 짚는 핵심 액션과 공간감을 사실적으로 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "장소 레퍼런스의 구도를 그대로 복사하지 말라는 지시를 어겼으며, 1인칭 시점으로 처리하여 두 승객이 사진을 함께 들고 있다는 설정을 누락했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지시된 클로즈업보다 넓은 미디엄 샷으로 인물과 배경을 과하게 노출했으나, 두 사람이 사진을 함께 들고 손가락으로 아이를 짚는 핵심 액션과 공간감을 사실적으로 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "장소 레퍼런스의 구도를 그대로 복사하지 말라는 지시를 어겼으며, 1인칭 시점으로 처리하여 두 승객이 사진을 함께 들고 있다는 설정을 누락했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지시된 클로즈업보다 프레임이 넓게(미디엄 샷) 잡혀 아쉽지만, 두 사람이 사진을 함께 들고 있는 연출과 레퍼런스 소품(사진 속 두 소녀)을 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "전경의 손과 배경 공간의 물리적 비율이 맞지 않는 콜라주(Priority 1 위반) 형태이며, 레퍼런스 사진 속 두 소녀 중 한 명이 누락되는 치명적인 소품 오류가 있습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지시된 클로즈업보다 프레임이 넓게(미디엄 샷) 잡혀 아쉽지만, 두 사람이 사진을 함께 들고 있는 연출과 레퍼런스 소품(사진 속 두 소녀)을 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "전경의 손과 배경 공간의 물리적 비율이 맞지 않는 콜라주(Priority 1 위반) 형태이며, 레퍼런스 사진 속 두 소녀 중 한 명이 누락되는 치명적인 소품 오류가 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 12,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "지시된 클로즈업보다 넓은 미디엄 샷으로 인물과 배경을 과하게 노출했으나, 두 사람이 사진을 함께 들고 손가락으로 아이를 짚는 핵심 액션과 공간감을 사실적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "장소 레퍼런스의 구도를 그대로 복사하지 말라는 지시를 어겼으며, 1인칭 시점으로 처리하여 두 승객이 사진을 함께 들고 있다는 설정을 누락했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/episodes/c0e2bace-39ca-47de-88f7-99463c1696bc/images/background_chain/L20B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:853090>"
   },
   {
    "label": "PROP REFERENCE — 두 한국인 어린 소녀가 찍힌 낡은 사진: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1242961>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "검지손가락 끝이 사진 속 아이의 얼굴을 꾹 짚지 않고 허공에 뜬 채 가리키고만 있습니다.",
     "fix_en": "Make the index fingertip physically touch and press firmly against the child's face on the photograph's surface."
    },
    {
     "issue_ko": "낡은 사진 속 아이들과 건물의 크기가 소품 레퍼런스와 다르게 사진 내에서 너무 작게 축소되어 있습니다.",
     "fix_en": "Restore the exact scale and layout of the girls and the building inside the old photograph to match the provided prop reference perfectly."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Move the camera inward to a true close-up at hand height, enlarging the photograph to just under two-fifths of the frame while keeping the fingertip and the child's face together near center and cropping out the passengers' faces and most of their bodies.",
     "issue_ko": "현재 프레임은 승객들의 얼굴과 상반신 및 조타실을 넓게 담은 미디엄 숏으로, 사진과 손가락이 지나치게 작다. 지시된 손 높이의 클로즈업이 아니다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Make the index fingertip physically touch and press firmly against the child's face on the photograph's surface.\n- Restore the exact scale and layout of the girls and the building inside the old photograph to match the provided prop reference perfectly.\n- Move the camera inward to a true close-up at hand height, enlarging the photograph to just under two-fifths of the frame while keeping the fingertip and the child's face together near center and cropping out the passengers' faces and most of their bodies.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 클로즈업 앵글과 피사체 크기를 더 잘 구현했으며, 소품(낡은 사진)의 디테일과 텍스트가 레퍼런스와 매우 흡사하게 묘사되었습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "클로즈업 샷이라기보다 미디엄 샷에 가깝게 프레이밍이 넓게 잡혔으며, 사진 속 디테일(간판 등)이 다소 왜곡되었습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 클로즈업 앵글과 피사체 크기를 더 잘 구현했으며, 소품(낡은 사진)의 디테일과 텍스트가 레퍼런스와 매우 흡사하게 묘사되었습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "클로즈업 샷이라기보다 미디엄 샷에 가깝게 프레이밍이 넓게 잡혔으며, 사진 속 디테일(간판 등)이 다소 왜곡되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "레이아웃 스케치에 지정된 오른쪽 인물이 프레임에서 누락되었고, 가리키는 손의 형태가 해부학적으로 부자연스럽습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "레이아웃 스케치의 두 인물 배치와 구도를 정확히 반영했으며, 사진 소품과 배경의 디테일을 충실하게 구현했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "레이아웃 스케치에 지정된 오른쪽 인물이 프레임에서 누락되었고, 가리키는 손의 형태가 해부학적으로 부자연스럽습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "레이아웃 스케치의 두 인물 배치와 구도를 정확히 반영했으며, 사진 소품과 배경의 디테일을 충실하게 구현했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 12
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S28sh4__bgfirst_bg.png",
   "bg_asset_id": "74303897-861d-49f5-a444-677e1f7df882",
   "bg_record_key": "S28sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S28sh6::variants": {
  "author_fp": "98b15b91ecc7f1b4",
  "author": {
   "variants": [
    {
     "approach_ko": "수평과 깊은 초점으로 공간 전체를 냉정한 증거물처럼 드러내는 정적인 광각 접근.",
     "prompt_en": "Use a restrained wide-angle feel with level, rectilinear geometry and deep focus, turning the wheelhouse into a stark, observational tableau. Keep the foreground profile narrow but clearly defined while the layered wear, brush texture, and dense ring pattern remain sharply readable across the wall. Let the dim dusk exposure fall softly and evenly, preserving subdued shadow detail and isolating the crimson through color contrast rather than heightened brightness."
    },
    {
     "approach_ko": "압축된 원근과 벽면 초점으로 인우의 굳은 어깨를 흐릿한 장벽처럼 만드는 밀실적 접근.",
     "prompt_en": "Use a subtly compressed perspective that visually presses the foreground figure against the wall and makes the interior feel airless. Hold critical focus on the painted surface while allowing the near head and shoulder to fall gently soft, their rigid contour acting as a dark, narrow frame for the reveal. Shape the low dusk ambience with restrained lateral falloff so the wall’s age and raised brushstrokes emerge unevenly from the gloom without making the red unnaturally vivid."
    }
   ]
  },
  "reused": false
 },
 "S28sh6": {
  "input_fingerprint": "4b8c33f3fbeb9bec",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 굳어버린 인우의 어깨 너머 조타실 낡은 벽면 위로, 피처럼 붉은 도료로 칠해진 수많은 원들이 빼곡하게 채워진 스산한 시점 쇼트.\n\nLOCATION (lock): Inside the aging wheelhouse, looking past the boatman’s shoulder toward a wall crowded with painted red circles. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At standing eye level immediately behind 인우's shoulder, settle the track into a firm over-the-shoulder wide composition aimed across the wheelhouse. His rigid head and shoulder form a narrow foreground block along one side, while the wall face bearing the numerous red circles remains fully legible across the midground and background.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 조타실 벽면의 수많은 원 (Numerous red circles cover the visible wall face) — The interior wall face visible to camera is densely covered with numerous circles painted in blood-like red pigment; used as Principal reveal beyond 인우's foreground shoulder; 인우의 어깨 윤곽 (Rigid and motionless at the endpoint of the track) — The back and shoulder face the camera while his front is directed away toward the wheelhouse interior; used as Foreground occlusion that creates the over-the-shoulder reveal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low dusk ambient leaves the wheelhouse desaturated and oppressive, with the painted red circles providing the only restrained crimson emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wheelhouse walls remain densely covered with numerous circles, including a conspicuous red one. Su-riyeong still has the old photograph and her travel belongings with her in the wheelhouse.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 굳어버린 인우의 어깨 너머 조타실 낡은 벽면 위로, 피처럼 붉은 도료로 칠해진 수많은 원들이 빼곡하게 채워진 스산한 시점 쇼트.\n\nLOCATION (lock): Inside the aging wheelhouse, looking past the boatman’s shoulder toward a wall crowded with painted red circles. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At standing eye level immediately behind 인우's shoulder, settle the track into a firm over-the-shoulder wide composition aimed across the wheelhouse. His rigid head and shoulder form a narrow foreground block along one side, while the wall face bearing the numerous red circles remains fully legible across the midground and background.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 조타실 벽면의 수많은 원 (Numerous red circles cover the visible wall face) — The interior wall face visible to camera is densely covered with numerous circles painted in blood-like red pigment; used as Principal reveal beyond 인우's foreground shoulder; 인우의 어깨 윤곽 (Rigid and motionless at the endpoint of the track) — The back and shoulder face the camera while his front is directed away toward the wheelhouse interior; used as Foreground occlusion that creates the over-the-shoulder reveal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low dusk ambient leaves the wheelhouse desaturated and oppressive, with the painted red circles providing the only restrained crimson emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wheelhouse walls remain densely covered with numerous circles, including a conspicuous red one. Su-riyeong still has the old photograph and her travel belongings with her in the wheelhouse.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained wide-angle feel with level, rectilinear geometry and deep focus, turning the wheelhouse into a stark, observational tableau. Keep the foreground profile narrow but clearly defined while the layered wear, brush texture, and dense ring pattern remain sharply readable across the wall. Let the dim dusk exposure fall softly and evenly, preserving subdued shadow detail and isolating the crimson through color contrast rather than heightened brightness.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 굳어버린 인우의 어깨 너머 조타실 낡은 벽면 위로, 피처럼 붉은 도료로 칠해진 수많은 원들이 빼곡하게 채워진 스산한 시점 쇼트.\n\nLOCATION (lock): Inside the aging wheelhouse, looking past the boatman’s shoulder toward a wall crowded with painted red circles. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At standing eye level immediately behind 인우's shoulder, settle the track into a firm over-the-shoulder wide composition aimed across the wheelhouse. His rigid head and shoulder form a narrow foreground block along one side, while the wall face bearing the numerous red circles remains fully legible across the midground and background.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 조타실 벽면의 수많은 원 (Numerous red circles cover the visible wall face) — The interior wall face visible to camera is densely covered with numerous circles painted in blood-like red pigment; used as Principal reveal beyond 인우's foreground shoulder; 인우의 어깨 윤곽 (Rigid and motionless at the endpoint of the track) — The back and shoulder face the camera while his front is directed away toward the wheelhouse interior; used as Foreground occlusion that creates the over-the-shoulder reveal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low dusk ambient leaves the wheelhouse desaturated and oppressive, with the painted red circles providing the only restrained crimson emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The wheelhouse walls remain densely covered with numerous circles, including a conspicuous red one. Su-riyeong still has the old photograph and her travel belongings with her in the wheelhouse.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a subtly compressed perspective that visually presses the foreground figure against the wall and makes the interior feel airless. Hold critical focus on the painted surface while allowing the near head and shoulder to fall gently soft, their rigid contour acting as a dark, narrow frame for the reveal. Shape the low dusk ambience with restrained lateral falloff so the wall’s age and raised brushstrokes emerge unevenly from the gloom without making the red unnaturally vivid.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "B": 8,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "이전 샷의 조타실 내부 구조와 기기 배치를 완벽하게 유지하면서, 요구된 어깨 너머 구도와 붉은 원들을 충실히 구현했습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "어깨 너머 구도는 잡았으나, 이전 샷과 조타실 내부 구조 및 조종 장치 배치가 전혀 일치하지 않는 치명적 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S28sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 인우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:849126>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "중앙 조타석 측면 하단에 속이 붉게 채워진 원이 그려져 있습니다. (지시사항: 그려진 원은 속이 빈 고리 형태를 유지해야 하며, 꽉 찬 원반 형태가 되어서는 안 됨)",
     "fix_en": "Convert the solid red painted circle on the side of the white central cabinet into an open red ring."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reframe In-woo as a narrow foreground edge along the left side, reducing his head-and-shoulder screen area while preserving the over-the-shoulder viewpoint.",
     "issue_ko": "인우의 머리와 어깨가 화면 왼쪽을 지나치게 넓게 차지해, 한쪽의 좁은 전경 블록이어야 한다는 구도를 벗어난다."
    },
    {
     "fix_en": "Redirect the camera axis slightly toward the circle-covered wall so that the wall reads as the dominant, fully legible midground-background reveal rather than the windows and helm.",
     "issue_ko": "카메라 축이 원들로 뒤덮인 벽면보다 전면 창과 조타 장치 쪽에 치우쳐, 붉은 원 벽면이 중·후경의 주된 공개 대상으로 충분히 크게 드러나지 않는다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Convert the solid red painted circle on the side of the white central cabinet into an open red ring.\n- Reframe In-woo as a narrow foreground edge along the left side, reducing his head-and-shoulder screen area while preserving the over-the-shoulder viewpoint.\n- Redirect the camera axis slightly toward the circle-covered wall so that the wall reads as the dominant, fully legible midground-background reveal rather than the windows and helm.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "어깨 너머의 시점 쇼트 구도를 정확히 구현했으며, 이전 컷의 조타실 공간 구조와 창문 위치 등을 훼손하지 않고 유지했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "원들을 빼곡히 채우기 위해 조타실 중앙에 기존 레퍼런스에 없던 거대한 수납장/벽면을 임의로 생성하여 장소의 일관성을 크게 훼손했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "어깨 너머의 시점 쇼트 구도를 정확히 구현했으며, 이전 컷의 조타실 공간 구조와 창문 위치 등을 훼손하지 않고 유지했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "원들을 빼곡히 채우기 위해 조타실 중앙에 기존 레퍼런스에 없던 거대한 수납장/벽면을 임의로 생성하여 장소의 일관성을 크게 훼손했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "기준 이미지의 조타실 구조(창문, 조향 장치 등)를 정확히 유지하면서 요구된 어깨 너머 구도와 붉은 원 형태를 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "조타실 정면 창문 위치에 거대한 벽돌/수납장을 임의로 생성하여 기준 이미지의 공간 구조를 심각하게 훼손했습니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "기준 이미지의 조타실 구조(창문, 조향 장치 등)를 정확히 유지하면서 요구된 어깨 너머 구도와 붉은 원 형태를 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "조타실 정면 창문 위치에 거대한 벽돌/수납장을 임의로 생성하여 기준 이미지의 공간 구조를 심각하게 훼손했습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S28sh4"
  }
 },
 "S28sh9::variants": {
  "author_fp": "61ef7abb2f0b974d",
  "author": {
   "variants": [
    {
     "approach_ko": "얕은 심도로 이마와 어깨의 접촉점, 감긴 눈, 무너진 체중을 섬세하게 고립시키는 친밀한 관찰.",
     "prompt_en": "Use a gently compressed, intimate lens feel with very shallow depth of field, holding the forehead-to-shoulder contact and her closed three-quarter face in the same delicate plane while his farther features fall softly away. Let broad, muted dusk ambience skim the contours of skin, dark hair, and creased clothing, with the wheelhouse reduced to subdued, weathered traces at the frame edges. Favor quiet tonal transitions and palpable bodily weight, making the frozen contact feel fragile and final."
    },
    {
     "approach_ko": "조금 더 깊은 초점과 근거리 원근감으로 두 사람의 무게 관계와 조타실 가장자리의 압박감을 함께 드러내는 관찰적 화면.",
     "prompt_en": "Use a restrained close-focus, slightly wider lens feel with deeper focus, preserving natural facial proportions while making the cramped wheelhouse margins press closely around both figures. Keep the centered contact point geometrically firm, and emphasize the diagonal flow of her collapse from left into his braced shoulder on the right. Render the low dusk light with firmer texture across hair, fabric, worn interior surfaces, and the unchanged red ring marks at the margins, creating a stark observational stillness rather than a softened portrait."
    }
   ]
  },
  "reused": false
 },
 "S28sh9": {
  "input_fingerprint": "3e336c5ab924a706",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 눈을 감고 의식을 잃은 수리영의 이마가 인우의 어깨 위에 툭 떨어진 채 멈춘 정지 구도.\n\nLOCATION (lock): Inside the cramped wheelhouse beside the controls, where the unconscious woman collapses against the boatman. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside both figures, hold a static close composition with the point where 수리영's forehead meets 인우's shoulder centered in the frame. 수리영 folds in from the left with her closed face visible in three-quarter view, while 인우 braces on the right and remains partially turned toward her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground; 인우 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 조타실 내부 (Visible only at the margins of the tightened frame); used as Minimal edge context that keeps the contact grounded inside the wheelhouse.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted dusk ambient holds both figures in desaturated low-key contrast, emphasizing the stillness of 수리영's unconscious face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the cramped weathered wheelhouse, dim dusk lighting, and numerous red circles covering the wall from the reference. Keep the young man's clothing consistent, but exclude the prior over-the-shoulder viewpoint and the photograph close-up; show the unconscious young woman collapsed onto his shoulder.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her unconscious body slumps against Inwoo with her eyes closed and her forehead resting on his shoulder, which supports her upper body. The source does not specify the exact positions of her arms or legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her unconscious body slumps against Inwoo with her eyes closed and her forehead resting on his shoulder, which supports her upper body. The old photograph, printed maps, backpack, coordinate note, and phone remain aboard the boat, and the wheelhouse circles remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 눈을 감고 의식을 잃은 수리영의 이마가 인우의 어깨 위에 툭 떨어진 채 멈춘 정지 구도.\n\nLOCATION (lock): Inside the cramped wheelhouse beside the controls, where the unconscious woman collapses against the boatman. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside both figures, hold a static close composition with the point where 수리영's forehead meets 인우's shoulder centered in the frame. 수리영 folds in from the left with her closed face visible in three-quarter view, while 인우 braces on the right and remains partially turned toward her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground; 인우 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 조타실 내부 (Visible only at the margins of the tightened frame); used as Minimal edge context that keeps the contact grounded inside the wheelhouse.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted dusk ambient holds both figures in desaturated low-key contrast, emphasizing the stillness of 수리영's unconscious face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the cramped weathered wheelhouse, dim dusk lighting, and numerous red circles covering the wall from the reference. Keep the young man's clothing consistent, but exclude the prior over-the-shoulder viewpoint and the photograph close-up; show the unconscious young woman collapsed onto his shoulder.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her unconscious body slumps against Inwoo with her eyes closed and her forehead resting on his shoulder, which supports her upper body. The source does not specify the exact positions of her arms or legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her unconscious body slumps against Inwoo with her eyes closed and her forehead resting on his shoulder, which supports her upper body. The old photograph, printed maps, backpack, coordinate note, and phone remain aboard the boat, and the wheelhouse circles remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a gently compressed, intimate lens feel with very shallow depth of field, holding the forehead-to-shoulder contact and her closed three-quarter face in the same delicate plane while his farther features fall softly away. Let broad, muted dusk ambience skim the contours of skin, dark hair, and creased clothing, with the wheelhouse reduced to subdued, weathered traces at the frame edges. Favor quiet tonal transitions and palpable bodily weight, making the frozen contact feel fragile and final.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): dusk.\n\nSHOT TEXT (authoritative, Korean): 눈을 감고 의식을 잃은 수리영의 이마가 인우의 어깨 위에 툭 떨어진 채 멈춘 정지 구도.\n\nLOCATION (lock): Inside the cramped wheelhouse beside the controls, where the unconscious woman collapses against the boatman. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside both figures, hold a static close composition with the point where 수리영's forehead meets 인우's shoulder centered in the frame. 수리영 folds in from the left with her closed face visible in three-quarter view, while 인우 braces on the right and remains partially turned toward her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 수리영 in the middle-left of the frame, midground; 인우 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 조타실 내부 (Visible only at the margins of the tightened frame); used as Minimal edge context that keeps the contact grounded inside the wheelhouse.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Muted dusk ambient holds both figures in desaturated low-key contrast, emphasizing the stillness of 수리영's unconscious face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the cramped weathered wheelhouse, dim dusk lighting, and numerous red circles covering the wall from the reference. Keep the young man's clothing consistent, but exclude the prior over-the-shoulder viewpoint and the photograph close-up; show the unconscious young woman collapsed onto his shoulder.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Her unconscious body slumps against Inwoo with her eyes closed and her forehead resting on his shoulder, which supports her upper body. The source does not specify the exact positions of her arms or legs.\n\nIMMOBILE BODIES OBEY GRAVITY: a person who is dead or unconscious\nexerts NO muscular effort. Every part of their body — head, torso,\narms, hands, fingers, legs — rests fully on whatever supports it\n(floor, wall, furniture, their own lap) and hangs or slumps with\ngravity. NEVER show any part of an immobile person's body lifted,\nraised, held up in the air, or posed as if presenting something:\nan object in their grip stays clenched in a hand that itself lies\nfallen on a support — the hand does not hold the object up. If the\ncanonical pose leaves a body part unspecified, resolve it as the\nmost gravity-compliant, fully-supported position.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above\nwins): when no pose contract, immobility contract or shot-text\ndirection says otherwise, people read as alive in mid-moment —\nbelievable weight shift, hands engaged with something real, gaze on\nthe target the SHOT TEXT implies; avoid a stiff attention stance (feet\ntogether, arms hanging straight down) and a blank stare into the lens.\nIf the SHOT TEXT or any POSE/IMMOBILE contract above stages stillness,\ndeath, sleep, unconsciousness, restraint, drill or an explicit\ndirect-to-camera look, follow THAT exactly — this clause never\noverrides it.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Her unconscious body slumps against Inwoo with her eyes closed and her forehead resting on his shoulder, which supports her upper body. The old photograph, printed maps, backpack, coordinate note, and phone remain aboard the boat, and the wheelhouse circles remain unchanged.\n\nPEOPLE: the SHOT TEXT alone decides whether any person is visible in this shot. IF a person appears, they must be one of: 수리영 (한국인 20대 초반 여성 얼굴, 타원형의 젊은 얼굴, 긴 검은 머리카락); 인우 (한국인 20대 중반 남성 얼굴, 수려한 이목구비, 자연스러운 짙은색 머리카락) — never anyone else, and never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person's profile above.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a restrained close-focus, slightly wider lens feel with deeper focus, preserving natural facial proportions while making the cramped wheelhouse margins press closely around both figures. Keep the centered contact point geometrically firm, and emphasize the diagonal flow of her collapse from left into his braced shoulder on the right. Render the low dusk light with firmer texture across hair, fabric, worn interior surfaces, and the unchanged red ring marks at the margins, creating a stark observational stillness rather than a softened portrait.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "요구된 클로즈업 프레이밍과 타이트한 화면 구성을 훌륭히 구현했으나, 수리영의 이마가 어깨가 아닌 뺨에 닿아 있고 비니가 누락되었습니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "인물들의 자세와 복장 디테일은 정확하지만, 명시된 클로즈업 지시를 위반하고 화면을 미디엄 샷으로 넓게 잡아 우선순위에서 밀립니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/28ac6c06-e1df-4fba-a789-6cd55401a4d8/images/c0e2bace-39ca-47de-88f7-99463c1696bc/scene/recipe/S28sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 수리영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1167875>"
   },
   {
    "label": "CHARACTER REFERENCE — 인우: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:849126>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "수리영의 이마가 인우의 어깨가 아닌 얼굴(뺨/관자놀이)에 닿아 있습니다.",
     "fix_en": "Lower Suri-young's head so her forehead rests firmly on In-woo's shoulder, removing all contact with his face."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Reposition Inwoo and Suriyoung so her forehead rests directly on the top of his shoulder rather than against his neck or upper chest, preserving everything else.",
     "issue_ko": "수리영의 이마가 인우의 어깨가 아니라 목과 윗가슴 쪽에 닿아 있어 지정된 접촉 구도가 어긋난다."
    },
    {
     "fix_en": "Shift the camera laterally and reframe both figures at the same midground depth so Inwoo’s back does not dominate as a foreground over-the-shoulder mass.",
     "issue_ko": "인우의 등과 어깨가 전경에서 지나치게 크게 차지해 두 인물이 모두 중경에 놓이는 밀착 투샷 대신 오버숄더 구도처럼 보인다."
    },
    {
     "fix_en": "Recenter the frame precisely on the forehead-to-shoulder contact point while keeping Suriyoung middle-left and Inwoo middle-right.",
     "issue_ko": "이마와 어깨의 접촉점이 화면 중앙보다 오른쪽에 치우쳐 있다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Lower Suri-young's head so her forehead rests firmly on In-woo's shoulder, removing all contact with his face.\n- Reposition Inwoo and Suriyoung so her forehead rests directly on the top of his shoulder rather than against his neck or upper chest, preserving everything else.\n- Shift the camera laterally and reframe both figures at the same midground depth so Inwoo’s back does not dominate as a foreground over-the-shoulder mass.\n- Recenter the frame precisely on the forehead-to-shoulder contact point while keeping Suriyoung middle-left and Inwoo middle-right.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "수리영의 이마가 인우의 어깨에 닿아 있는 핵심 구도를 정확히 구현하였고, 지정된 조타실 배경과 인물 배치를 잘 따랐습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "수리영의 이마가 인우의 어깨가 아닌 얼굴(뺨)에 기대어 있어 '이마가 어깨에 떨어진' 주요 지시사항을 위반했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "수리영의 이마가 인우의 어깨에 닿아 있는 핵심 구도를 정확히 구현하였고, 지정된 조타실 배경과 인물 배치를 잘 따랐습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "수리영의 이마가 인우의 어깨가 아닌 얼굴(뺨)에 기대어 있어 '이마가 어깨에 떨어진' 주요 지시사항을 위반했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "인우가 고개를 들고 지탱하는 자세를 유지하며, 수리영이 의식을 잃고 어깨와 목 부근에 기대어 있는 모습을 비교적 충실하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "인우가 눈을 감고 이마를 맞대는 지시되지 않은 제스처를 취해, 의식을 잃고 쓰러진 정지 구도라는 지문과 어긋납니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "인우가 고개를 들고 지탱하는 자세를 유지하며, 수리영이 의식을 잃고 어깨와 목 부근에 기대어 있는 모습을 비교적 충실하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "인우가 눈을 감고 이마를 맞대는 지시되지 않은 제스처를 취해, 의식을 잃고 쓰러진 정지 구도라는 지문과 어긋납니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S28sh6"
  }
 },
 "S29sh3::variants": {
  "author_fp": "04228ddb32049c6a",
  "author": {
   "variants": [
    {
     "approach_ko": "망원 압축감과 정면 대칭으로 안개가 배를 삼키는 불가해한 경계를 평면적이고 정지된 이미지로 강조한다.",
     "prompt_en": "Use restrained long-lens compression while preserving the locked waterline-wide framing and exact centered scale, flattening the sea, bow, and opaque fog boundary into a severe frontal composition. Let the fog diffuse the sunset into a nearly shadowless, desaturated golden-gray field, with only the weathered forward hull and a narrow band of subdued sea retaining tactile contrast. Hold the geometry rigidly symmetrical and eerily motionless, emphasizing the abrupt erasure of the boat’s farther outline."
    },
    {
     "approach_ko": "넓은 원근감과 전경 수면의 깊이로 뱃머리가 거대한 안개 벽을 통과하는 물리적 문턱을 강조한다.",
     "prompt_en": "Use a moderately wide, naturalistic lens feel at the fixed waterline position, preserving the prescribed centered bow placement while giving the lower-frame sea stronger spatial depth toward the midground boat. Render the fog bank as a massive soft-edged volume rather than a flat backdrop, with fading sunset light grazing the visible hull and glinting faintly across the water before being swallowed by the dense haze. Keep foreground texture crisp enough to establish physical distance, then let contrast and detail fall away rapidly at the boat’s embedded half."
    }
   ]
  },
  "reused": false
 },
 "S29sh3": {
  "input_fingerprint": "335691712371997a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): sunset, dense fog.\n\nSHOT TEXT (authoritative, Korean): 짙고 불투명한 안개 띠 속으로 낡은 소형어선의 뱃머리가 절반쯤 파묻힌 기괴한 정지 찰나.\n\nLOCATION (lock): On the open sea at sunset, at the boundary where the small fishing boat enters a dense, opaque bank of fog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At waterline height a short distance ahead of the boat, hold a static wide view centered precisely on its forward axis as the bow penetrates the fog band. Keep the bow below two-fifths of the image, with its forward half still readable against the sea while the remaining form disappears into the dense boundary.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: boat bow halfway within fog in the middle-center of the frame, midground; dense fog band in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 소형어선 뱃머리 (Continuing slowly into the fog band) — The forward face points toward the camera on the boat's central axis, with roughly half of the bow embedded in fog; used as Centered scale anchor crossing the visual threshold; 짙은 안개 띠 (Dense and opaque); used as Opaque boundary that erases the boat's farther outline and the separation between sea and distance; 바다 (The boat remains afloat as it enters the fog); used as Lower-frame spatial reference at camera height.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun casts fading golden light across the sea, muted and desaturated as the dense fog obscures the boat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same small boat carries the still-unconscious Su-riyeong and her belongings into the dense fog; nothing establishes that she has recovered or left In-woo’s vessel.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): sunset, dense fog.\n\nSHOT TEXT (authoritative, Korean): 짙고 불투명한 안개 띠 속으로 낡은 소형어선의 뱃머리가 절반쯤 파묻힌 기괴한 정지 찰나.\n\nLOCATION (lock): On the open sea at sunset, at the boundary where the small fishing boat enters a dense, opaque bank of fog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At waterline height a short distance ahead of the boat, hold a static wide view centered precisely on its forward axis as the bow penetrates the fog band. Keep the bow below two-fifths of the image, with its forward half still readable against the sea while the remaining form disappears into the dense boundary.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: boat bow halfway within fog in the middle-center of the frame, midground; dense fog band in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 소형어선 뱃머리 (Continuing slowly into the fog band) — The forward face points toward the camera on the boat's central axis, with roughly half of the bow embedded in fog; used as Centered scale anchor crossing the visual threshold; 짙은 안개 띠 (Dense and opaque); used as Opaque boundary that erases the boat's farther outline and the separation between sea and distance; 바다 (The boat remains afloat as it enters the fog); used as Lower-frame spatial reference at camera height.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun casts fading golden light across the sea, muted and desaturated as the dense fog obscures the boat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same small boat carries the still-unconscious Su-riyeong and her belongings into the dense fog; nothing establishes that she has recovered or left In-woo’s vessel.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse restrained long-lens compression while preserving the locked waterline-wide framing and exact centered scale, flattening the sea, bow, and opaque fog boundary into a severe frontal composition. Let the fog diffuse the sunset into a nearly shadowless, desaturated golden-gray field, with only the weathered forward hull and a narrow band of subdued sea retaining tactile contrast. Hold the geometry rigidly symmetrical and eerily motionless, emphasizing the abrupt erasure of the boat’s farther outline.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — coastal Incheon, South Korea, 2026; all people are Korean and Korean-speaking unless stated. TIME OF DAY (lock): sunset, dense fog.\n\nSHOT TEXT (authoritative, Korean): 짙고 불투명한 안개 띠 속으로 낡은 소형어선의 뱃머리가 절반쯤 파묻힌 기괴한 정지 찰나.\n\nLOCATION (lock): On the open sea at sunset, at the boundary where the small fishing boat enters a dense, opaque bank of fog. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At waterline height a short distance ahead of the boat, hold a static wide view centered precisely on its forward axis as the bow penetrates the fog band. Keep the bow below two-fifths of the image, with its forward half still readable against the sea while the remaining form disappears into the dense boundary.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: boat bow halfway within fog in the middle-center of the frame, midground; dense fog band in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 소형어선 뱃머리 (Continuing slowly into the fog band) — The forward face points toward the camera on the boat's central axis, with roughly half of the bow embedded in fog; used as Centered scale anchor crossing the visual threshold; 짙은 안개 띠 (Dense and opaque); used as Opaque boundary that erases the boat's farther outline and the separation between sea and distance; 바다 (The boat remains afloat as it enters the fog); used as Lower-frame spatial reference at camera height.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun casts fading golden light across the sea, muted and desaturated as the dense fog obscures the boat.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE AS A LIVE-ACTION SHOT: the Korean shot\ntext may describe characters metaphorically, figuratively or with\nexaggeration. Photograph what a real movie camera would actually\nrecord on a physical set — exaggerated or figurative impressions\nbecome realistic staging choices (lighting, distance, angle,\nwardrobe), not literal fantasy imagery.\nEVERY CHARACTER IS A HUMAN BEING: unless the story explicitly\nfeatures non-human or virtual beings (as in science fiction or\nfantasy), every character — however indirectly, vaguely or\nfiguratively the text describes them — IS a real human. When the\ntext gives no direct visual description of a person, IMAGINE one\nand still show them as a concrete, fully-formed human being:\nalways render the human form to the maximum extent — build,\nposture, face, hands, clothing — never reduce a person to a shape,\nblob, solid silhouette or abstract mass.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: when the text describes a\nperson's eyes, face or presence emotionally or figuratively (vacant,\nhollow, dazed, burning, lifeless gaze and the like), it describes an\nACTOR'S PERFORMANCE captured by a real camera — realize it ONLY\nthrough gaze direction, focus, eyelids, facial muscles, stillness\nand posture. Every living person's eyes remain anatomically normal\nhuman eyes with a natural iris and pupil, natural sclera and normal\nproportions; NEVER whiten, blank out, cloud over, glow, enlarge or\notherwise alter eyeballs, skin or anatomy — unless the story\nexplicitly declares that being non-human or supernatural in form.\n\nPROPS FACE THE RIGHT WAY: every handheld or used object must be\noriented exactly as its real-world use requires. A person reading,\nwatching or operating something (a phone, a photograph, a paper,\nany device) has its functional side — screen, front, page — facing\nTHEIR OWN eyes; the camera then sees whatever side the staging\ngeometry implies (often its back). Show the functional side to the\ncamera ONLY when the shot text itself stages it toward the viewer.\nNever flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: any mark the text describes as drawn,\npainted or traced (a circle, a line, a symbol) is a STROKE sitting on\nthe surface — an outline whose interior still shows the underlying\nsurface (wall, skin, paper). Render its stated shape faithfully: a\ndrawn circle stays an open ring of brush-width, never filled into a\nsolid disc, unless the text explicitly says it is filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same small boat carries the still-unconscious Su-riyeong and her belongings into the dense fog; nothing establishes that she has recovered or left In-woo’s vessel.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nNo text, captions, watermarks or annotations anywhere.\n\nVISUAL APPROACH — this take (camera, framing, composition):\nUse a moderately wide, naturalistic lens feel at the fixed waterline position, preserving the prescribed centered bow placement while giving the lower-frame sea stronger spatial depth toward the midground boat. Render the fog bank as a massive soft-edged volume rather than a flat backdrop, with fading sunset light grazing the visible hull and glinting faintly across the water before being swallowed by the dense haze. Keep foreground texture crisp enough to establish physical distance, then let contrast and detail fall away rapidly at the boat’s embedded half.\n\nPRIORITY OF INSTRUCTIONS: the VISUAL APPROACH above governs camera,\nframing and composition ONLY. Every other instruction in this prompt —\nlocation, time of day, people and their identity, poses, props,\ncarried state, physical rules, and the no-text rule — is binding\nexactly as written earlier. If the VISUAL APPROACH appears to conflict\nwith any of them, the earlier instruction wins."
  },
  "totals": {
   "A": 3,
   "B": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "배 위에서 앞을 내다보는 시점으로 연출되어, 배 앞 수면에서 뱃머리 정면을 바라보라는 카메라 구도 지시를 완전히 위반함."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "카메라가 배 앞 수면 높이에서 뱃머리 정면을 바라보며 선체 뒷부분이 안개에 파묻히는 지정된 카메라 구도와 비율을 정확히 구현함."
   }
  ],
  "refs": [],
  "critique": {
   "issues": [
    {
     "issue_ko": "뱃머리가 화면 세로 중앙에 위치해 '화면의 2/5 아래에 배치'하라는 지시를 위반함.",
     "fix_en": "Move the boat downwards so the bow sits entirely below the two-fifths vertical mark of the frame."
    },
    {
     "issue_ko": "선체 우측에 글자(텍스트)가 새겨져 있음 (텍스트 금지 지시 위반).",
     "fix_en": "Remove the lettering from the right side of the boat's hull, leaving only the bare weathered surface."
    }
   ]
  },
  "composition_critique": {
   "issues": [
    {
     "fix_en": "Extend the opaque fog boundary across roughly the farther half of the bow so that only its forward half remains readable against the sea, preserving the centered frontal axis, scale, and waterline framing.",
     "issue_ko": "뱃머리 전체가 안개 앞에서 선명하게 드러나 있어, 뱃머리의 약 절반이 안개 경계에 파묻혀 사라져야 하는 구도가 구현되지 않았다."
    }
   ]
  },
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Move the boat downwards so the bow sits entirely below the two-fifths vertical mark of the frame.\n- Remove the lettering from the right side of the boat's hull, leaving only the bare weathered surface.\n- Extend the opaque fog boundary across roughly the farther half of the bow so that only its forward half remains readable against the sea, preserving the centered frontal axis, scale, and waterline framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. No text, captions or\nwatermarks anywhere.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "정면 구도와 수면 높이의 카메라 워크를 잘 따랐으며, 배의 나머지 부분이 불투명한 안개 속으로 완전히 사라지는 찰나를 완벽하게 묘사했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 구도와 조명은 훌륭하나, 안개 너머로 선박의 뒷부분(마스트 등)이 희미하게 보여 '불투명한 안개'와 '나머지 형태가 사라짐'이라는 조건을 완전히 충족하지 못했습니다."
     }
    ]
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "정면 구도와 수면 높이의 카메라 워크를 잘 따랐으며, 배의 나머지 부분이 불투명한 안개 속으로 완전히 사라지는 찰나를 완벽하게 묘사했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 구도와 조명은 훌륭하나, 안개 너머로 선박의 뒷부분(마스트 등)이 희미하게 보여 '불투명한 안개'와 '나머지 형태가 사라짐'이라는 조건을 완전히 충족하지 못했습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 수면 높이의 앵글과 안개 속으로 사라지는 소형어선의 구도를 잘 살렸으며 불필요한 피사체가 없습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트에 명시되지 않은 정체불명의 대형 선박 실루엣이 배경에 임의로 추가되어 치명적인 감점 요소가 됩니다."
     }
    ]
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 수면 높이의 앵글과 안개 속으로 사라지는 소형어선의 구도를 잘 살렸으며 불필요한 피사체가 없습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트에 명시되지 않은 정체불명의 대형 선박 실루엣이 배경에 임의로 추가되어 치명적인 감점 요소가 됩니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "policy": "fix_rejudge_v1_flip"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 }
}