Add files using upload-large-folder tool

cutrope/data/rope-097.json

@@ -0,0 +1,164 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-097",
+  "level_file": "data/task/rope-097.json",
+  "video": {
+    "path": "data/video/rope-097.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bouncer": 1,
+      "bubble": 2,
+      "candy": 1,
+      "grab_or_rope_anchor": 3,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 164,
+        "y": 266
+      },
+      {
+        "name": 2,
+        "x": 236,
+        "y": 73
+      },
+      {
+        "name": 100,
+        "x": 154,
+        "y": 129,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 283,
+        "y": 269,
+        "length": 95,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 3,
+        "x": 77,
+        "y": 382,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 258,
+        "y": 378,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 235,
+        "y": 163,
+        "timeout": -1
+      },
+      {
+        "name": 82,
+        "x": 307,
+        "y": 385,
+        "angle": 90,
+        "size": 2
+      },
+      {
+        "name": 54,
+        "x": 77,
+        "y": 173
+      },
+      {
+        "name": 56,
+        "x": 203,
+        "y": 382,
+        "group": 0,
+        "angle": 0
+      },
+      {
+        "name": 56,
+        "x": 79,
+        "y": 104,
+        "group": 0,
+        "angle": 90
+      },
+      {
+        "name": 100,
+        "x": 34,
+        "y": 274,
+        "length": 95,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 80,
+        "y": 374
+      }
+    ],
+    "mirroredFrom": "06-23.json"
+  },
+  "reference_solution": "cut_rope 0,2\npop_bubble 0 when candy_y < 340",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:39:56.417Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,2\npop_bubble 0 when candy_y < 340",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,2\npop_bubble 0 when candy_y < 340",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-097\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 3\n  - bubble: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-098.json

@@ -0,0 +1,124 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-098",
+  "level_file": "data/task/rope-098.json",
+  "video": {
+    "path": "data/video/rope-098.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 140,
+        "y": 165
+      },
+      {
+        "name": 56,
+        "x": 140,
+        "y": 403,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 247,
+        "y": 402,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 3,
+        "x": 249,
+        "y": 53,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 143,
+        "y": 298,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 251,
+        "y": 296,
+        "timeout": -1
+      },
+      {
+        "name": 52,
+        "x": 246,
+        "y": 131
+      },
+      {
+        "name": 100,
+        "x": 248,
+        "y": 239,
+        "length": 80,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 247,
+        "y": 128
+      }
+    ],
+    "levelId": "04-13",
+    "levelEditedAt": "2026-05-05T15:41:29.028Z",
+    "mirroredFrom": "04-13.json"
+  },
+  "reference_solution": "cut_rope 0 \npop_bubble 0 when candy_y < 140\ncut_rope 0 when candy_y > 570",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:41:57.174Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 \npop_bubble 0 when candy_y < 140\ncut_rope 0 when candy_y > 570",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \npop_bubble 0 when candy_y < 140\ncut_rope 0 when candy_y > 570",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-098\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n  - bubble: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-099.json

@@ -0,0 +1,122 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-099",
+  "level_file": "data/task/rope-099.json",
+  "video": {
+    "path": "data/video/rope-099.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 2,
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "pump": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 54,
+        "x": 237,
+        "y": 124
+      },
+      {
+        "name": 54,
+        "x": 237,
+        "y": 319
+      },
+      {
+        "name": 55,
+        "x": 162,
+        "y": 126,
+        "angle": 0
+      },
+      {
+        "name": 2,
+        "x": 383,
+        "y": 420
+      },
+      {
+        "name": 52,
+        "x": 236,
+        "y": 324
+      },
+      {
+        "name": 3,
+        "x": 387,
+        "y": 317,
+        "timeout": 8
+      },
+      {
+        "name": 3,
+        "x": 380,
+        "y": 135,
+        "timeout": 9
+      },
+      {
+        "name": 3,
+        "x": 236,
+        "y": 124,
+        "timeout": 7
+      },
+      {
+        "name": 100,
+        "x": 237,
+        "y": 434,
+        "length": 80,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "levelId": "02-19",
+    "levelEditedAt": "2026-05-05T14:21:35.907Z",
+    "mirroredFrom": "02-19.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 770,720,60\nactivate_pump 0 when candy_y < 300\npop_bubble 0 when candy_y < 100",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:21:53.945Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 770,720,60\nactivate_pump 0 when candy_y < 300\npop_bubble 0 when candy_y < 100",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 770,720,60\nactivate_pump 0 when candy_y < 300\npop_bubble 0 when candy_y < 100",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-099\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n  - bubble: 2\n  - pump: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-100.json

@@ -0,0 +1,133 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-100",
+  "level_file": "data/task/rope-100.json",
+  "video": {
+    "path": "data/video/rope-100.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 3,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 402,
+        "y": 351
+      },
+      {
+        "name": 2,
+        "x": 234,
+        "y": 61
+      },
+      {
+        "name": 54,
+        "x": 247,
+        "y": 410
+      },
+      {
+        "name": 100,
+        "x": 330,
+        "y": 128,
+        "length": 210,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 224,
+        "y": 126,
+        "length": 275,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 54,
+        "x": 325,
+        "y": 418
+      },
+      {
+        "name": 54,
+        "x": 403,
+        "y": 412
+      },
+      {
+        "name": 3,
+        "x": 404,
+        "y": 287,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 235,
+        "y": 193,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 326,
+        "y": 418,
+        "timeout": -1
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0 \npop_bubble 2 when candy_y < 400",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T13:12:36.555Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 \npop_bubble 2 when candy_y < 400",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \npop_bubble 2 when candy_y < 400",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-100\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 3\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-101.json

@@ -0,0 +1,131 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-101",
+  "level_file": "data/task/rope-101.json",
+  "video": {
+    "path": "data/video/rope-101.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "pump": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 100,
+        "x": 125,
+        "y": 258,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 40,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 3,
+        "x": 126,
+        "y": 371,
+        "timeout": 10
+      },
+      {
+        "name": 54,
+        "x": 127,
+        "y": 378
+      },
+      {
+        "name": 2,
+        "x": 124,
+        "y": 71
+      },
+      {
+        "name": 3,
+        "x": 205,
+        "y": 83,
+        "timeout": -1
+      },
+      {
+        "name": 55,
+        "x": 36,
+        "y": 146,
+        "angle": 0
+      },
+      {
+        "name": 3,
+        "x": 31,
+        "y": 76,
+        "timeout": -1
+      },
+      {
+        "name": 52,
+        "x": 125,
+        "y": 140
+      },
+      {
+        "name": 100,
+        "x": 126,
+        "y": 34,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "levelId": "02-22",
+    "levelEditedAt": "2026-05-05T14:38:58.835Z",
+    "mirroredFrom": "02-22.json"
+  },
+  "reference_solution": "activate_pump 0 \ncut_rope 1 when candy_near 730,250,50 times 2\ncut_rope 0 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:41:38.128Z"
+  },
+  "answer": {
+    "commands": "activate_pump 0 \ncut_rope 1 when candy_near 730,250,50 times 2\ncut_rope 0 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "activate_pump 0 \ncut_rope 1 when candy_near 730,250,50 times 2\ncut_rope 0 when candy_still for 0.3",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-101\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 1\n  - pump: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-102.json

@@ -0,0 +1,171 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-102",
+  "level_file": "data/task/rope-102.json",
+  "video": {
+    "path": "data/video/rope-102.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bouncer": 2,
+      "bubble": 2,
+      "candy": 1,
+      "grab_or_rope_anchor": 3,
+      "object_103": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 157,
+        "y": 41
+      },
+      {
+        "name": 52,
+        "x": 154,
+        "y": 174
+      },
+      {
+        "name": 100,
+        "x": 258,
+        "y": 161,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 50,
+        "y": 160,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 301,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 198,
+        "y": 397,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 155,
+        "y": 109,
+        "timeout": 9
+      },
+      {
+        "name": 3,
+        "x": 110,
+        "y": 402,
+        "timeout": -1
+      },
+      {
+        "name": 81,
+        "x": 63,
+        "y": 407,
+        "angle": 90,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 260,
+        "y": 288,
+        "angle": 90,
+        "size": 1
+      },
+      {
+        "name": 54,
+        "x": 106,
+        "y": 398
+      },
+      {
+        "name": 54,
+        "x": 154,
+        "y": 173
+      },
+      {
+        "name": 103,
+        "x": 38,
+        "y": 212,
+        "angle": 20,
+        "Pj": 12
+      }
+    ],
+    "levelId": "06-17",
+    "levelEditedAt": "2026-05-05T17:31:14.946Z"
+  },
+  "reference_solution": "cut_rope 1 \npop_bubble 1 when candy_still for 0.3 and candy_near 960,370,60\ncut_rope 0 when candy_y > 660\ncut_rope 2 when candy_y < 380",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:32:03.745Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 \npop_bubble 1 when candy_still for 0.3 and candy_near 960,370,60\ncut_rope 0 when candy_y > 660\ncut_rope 2 when candy_y < 380",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 \npop_bubble 1 when candy_still for 0.3 and candy_near 960,370,60\ncut_rope 0 when candy_y > 660\ncut_rope 2 when candy_y < 380",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-102\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 3\n  - bubble: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-103.json

@@ -0,0 +1,135 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-103",
+  "level_file": "data/task/rope-103.json",
+  "video": {
+    "path": "data/video/rope-103.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 4,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "special": 1,
+        "ropePhysicsSpeed": 1
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 162,
+        "y": 187
+      },
+      {
+        "name": 2,
+        "x": 58,
+        "y": 362
+      },
+      {
+        "name": 100,
+        "x": 159,
+        "y": 315,
+        "length": 93,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 31,
+        "y": 186,
+        "length": 105,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 158,
+        "y": 57,
+        "length": 93,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 287,
+        "y": 186,
+        "length": 105,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 3,
+        "x": 161,
+        "y": 226,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 288,
+        "y": 312,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 434,
+        "timeout": -1
+      }
+    ],
+    "mirroredFrom": "01-04.json"
+  },
+  "reference_solution": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_x > 1200",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:54:15.133Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_x > 1200",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_x > 1200",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-103\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 4\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-104.json

@@ -0,0 +1,128 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-104",
+  "level_file": "data/task/rope-104.json",
+  "video": {
+    "path": "data/video/rope-104.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 2,
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 388,
+        "y": 284
+      },
+      {
+        "name": 56,
+        "x": 228,
+        "y": 395,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 384,
+        "y": 394,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 3,
+        "x": 230,
+        "y": 60,
+        "timeout": -1
+      },
+      {
+        "name": 2,
+        "x": 388,
+        "y": 122
+      },
+      {
+        "name": 3,
+        "x": 229,
+        "y": 179,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 230,
+        "y": 119,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 387,
+        "y": 219,
+        "length": 30,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 388,
+        "y": 343
+      },
+      {
+        "name": 54,
+        "x": 230,
+        "y": 256
+      }
+    ],
+    "mirroredFrom": "04-07.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 1110,640,60\npop_bubble 0 when candy_y < 450\npop_bubble 1 when candy_y < 100",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:44:30.241Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 1110,640,60\npop_bubble 0 when candy_y < 450\npop_bubble 1 when candy_y < 100",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 1110,640,60\npop_bubble 0 when candy_y < 450\npop_bubble 1 when candy_y < 100",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-104\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n  - bubble: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-105.json

@@ -0,0 +1,184 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-105",
+  "level_file": "data/task/rope-105.json",
+  "video": {
+    "path": "data/video/rope-105.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": true,
+    "special": null,
+    "object_counts": {
+      "bouncer": 2,
+      "bubble": 1,
+      "grab_or_rope_anchor": 4,
+      "left_candy": 1,
+      "right_candy": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "twoParts": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 51,
+        "x": 285,
+        "y": 188
+      },
+      {
+        "name": 50,
+        "x": 28,
+        "y": 188
+      },
+      {
+        "name": 100,
+        "x": 287,
+        "y": 95,
+        "length": 50,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 30,
+        "y": 95,
+        "length": 50,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 287,
+        "y": 384,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 30,
+        "y": 384,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 160,
+        "y": 435
+      },
+      {
+        "name": 3,
+        "x": 27,
+        "y": 143,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 286,
+        "y": 146,
+        "timeout": -1
+      },
+      {
+        "name": 81,
+        "x": 30,
+        "y": 440,
+        "angle": 195,
+        "size": 1
+      },
+      {
+        "name": 3,
+        "x": 160,
+        "y": 157,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 160,
+        "y": 329
+      },
+      {
+        "name": 81,
+        "x": 286,
+        "y": 436,
+        "angle": 165,
+        "size": 1
+      }
+    ],
+    "levelId": "09-14",
+    "levelEditedAt": "2026-05-05T18:39:38.480Z",
+    "mirroredFrom": "09-14.json"
+  },
+  "reference_solution": "cut_rope 2,3 \ncut_rope 0,1  \npop_bubble 0 when candy_y < 300",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:39:48.755Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2,3 \ncut_rope 0,1  \npop_bubble 0 when candy_y < 300",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2,3 \ncut_rope 0,1  \npop_bubble 0 when candy_y < 300",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-105\n- canvas_size: 1920 x 1080\n- two_parts (split candy): True\n- object_counts:\n  - target: 1\n  - star: 3\n  - left split candy: 1\n  - right split candy: 1\n  - rope/grab: 4\n  - bubble: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Split candy: the level starts with left_candy and right_candy halves. Before they merge, use left_candy_* and right_candy_* conditions. After they touch and merge, use candy_* for the complete candy.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-106.json

@@ -0,0 +1,155 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-106",
+  "level_file": "data/task/rope-106.json",
+  "video": {
+    "path": "data/video/rope-106.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 5,
+      "candy": 1,
+      "grab_or_rope_anchor": 3,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 54,
+        "x": 210,
+        "y": 315
+      },
+      {
+        "name": 54,
+        "x": 311,
+        "y": 420
+      },
+      {
+        "name": 100,
+        "x": 312,
+        "y": 318,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 80,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 54,
+        "x": 311,
+        "y": 210
+      },
+      {
+        "name": 3,
+        "x": 403,
+        "y": 167,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 216,
+        "y": 314,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 410,
+        "y": 313,
+        "timeout": -1
+      },
+      {
+        "name": 2,
+        "x": 318,
+        "y": 41
+      },
+      {
+        "name": 54,
+        "x": 239,
+        "y": 242
+      },
+      {
+        "name": 54,
+        "x": 239,
+        "y": 387
+      },
+      {
+        "name": 100,
+        "x": 231,
+        "y": 35,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 52,
+        "x": 315,
+        "y": 121
+      },
+      {
+        "name": 100,
+        "x": 416,
+        "y": 35,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 1 \ncut_rope 2 when candy_near 1250,360,60 times 2\ncut_rope 0 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:01:54.229Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 \ncut_rope 2 when candy_near 1250,360,60 times 2\ncut_rope 0 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 \ncut_rope 2 when candy_near 1250,360,60 times 2\ncut_rope 0 when candy_still for 0.3",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-106\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 3\n  - bubble: 5\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-107.json

@@ -0,0 +1,156 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-107",
+  "level_file": "data/task/rope-107.json",
+  "video": {
+    "path": "data/video/rope-107.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 5,
+      "candy": 1,
+      "grab_or_rope_anchor": 3,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 54,
+        "x": 430,
+        "y": 315
+      },
+      {
+        "name": 54,
+        "x": 329,
+        "y": 420
+      },
+      {
+        "name": 100,
+        "x": 328,
+        "y": 318,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 80,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 329,
+        "y": 210
+      },
+      {
+        "name": 3,
+        "x": 237,
+        "y": 167,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 424,
+        "y": 314,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 230,
+        "y": 313,
+        "timeout": -1
+      },
+      {
+        "name": 2,
+        "x": 322,
+        "y": 41
+      },
+      {
+        "name": 54,
+        "x": 401,
+        "y": 242
+      },
+      {
+        "name": 54,
+        "x": 401,
+        "y": 387
+      },
+      {
+        "name": 100,
+        "x": 409,
+        "y": 35,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 52,
+        "x": 325,
+        "y": 121
+      },
+      {
+        "name": 100,
+        "x": 224,
+        "y": 35,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "mirroredFrom": "02-13.json"
+  },
+  "reference_solution": "cut_rope 1 \ncut_rope 2 when candy_near 670,360,60 times 2\ncut_rope 0 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:01:54.229Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 \ncut_rope 2 when candy_near 670,360,60 times 2\ncut_rope 0 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 \ncut_rope 2 when candy_near 670,360,60 times 2\ncut_rope 0 when candy_still for 0.3",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-107\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 3\n  - bubble: 5\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-108.json

@@ -0,0 +1,124 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-108",
+  "level_file": "data/task/rope-108.json",
+  "video": {
+    "path": "data/video/rope-108.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "pump": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 290,
+        "y": 435
+      },
+      {
+        "name": 52,
+        "x": 290,
+        "y": 203
+      },
+      {
+        "name": 3,
+        "x": 356,
+        "y": 195,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 225,
+        "y": 197,
+        "timeout": -1
+      },
+      {
+        "name": 55,
+        "x": 365,
+        "y": 198,
+        "angle": 180
+      },
+      {
+        "name": 3,
+        "x": 290,
+        "y": 384,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 291,
+        "y": 26,
+        "length": 140,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 292,
+        "y": 341,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ],
+    "levelId": "02-15",
+    "levelEditedAt": "2026-05-05T14:10:19.581Z"
+  },
+  "reference_solution": "activate_pump 0 when candy_still for 0.3 and candy_near 900,460,60\ncut_rope 0 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:12:23.144Z"
+  },
+  "answer": {
+    "commands": "activate_pump 0 when candy_still for 0.3 and candy_near 900,460,60\ncut_rope 0 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "activate_pump 0 when candy_still for 0.3 and candy_near 900,460,60\ncut_rope 0 when candy_still for 0.3",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-108\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - pump: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-109.json

@@ -0,0 +1,182 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-109",
+  "level_file": "data/task/rope-109.json",
+  "video": {
+    "path": "data/video/rope-109.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": null,
+    "object_counts": {
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 160,
+        "y": 52
+      },
+      {
+        "name": 52,
+        "x": 157,
+        "y": 367
+      },
+      {
+        "name": 3,
+        "x": 161,
+        "y": 262,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 161,
+        "y": 145,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 161,
+        "y": 203,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 159,
+        "y": 463,
+        "length": 65,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 157,
+        "y": 365
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 166,
+        "y": 155,
+        "locale": "ru",
+        "text": "Эти пилы поворачиваются вместе",
+        "width": 150
+      },
+      {
+        "name": 13,
+        "x": 238,
+        "y": 138,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "w": [
+      {
+        "name": 4,
+        "x": 166,
+        "y": 161,
+        "locale": "en",
+        "text": "These blades rotate together",
+        "width": 150
+      },
+      {
+        "name": 13,
+        "x": 239,
+        "y": 147,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 125,
+        "y": 189,
+        "locale": "de",
+        "text": "Diese Sägeblätter drehen sich gemeinsam",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 215,
+        "y": 177,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 172,
+        "y": 162,
+        "locale": "fr",
+        "text": "Ces lames tournent ensemble",
+        "width": 140
+      },
+      {
+        "name": 13,
+        "x": 236,
+        "y": 150,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "mirroredFrom": "09-02.json"
+  },
+  "reference_solution": "cut_rope 0",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:23:46.516Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-109\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n  - bubble: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-110.json

@@ -0,0 +1,206 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-110",
+  "level_file": "data/task/rope-110.json",
+  "video": {
+    "path": "data/video/rope-110.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 2,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "object_4": 1,
+      "object_8": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 2,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 217,
+        "y": 413
+      },
+      {
+        "name": 100,
+        "x": 50,
+        "y": 56,
+        "length": 55,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 58,
+        "length": 130,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 52,
+        "x": 47,
+        "y": 150
+      },
+      {
+        "name": 3,
+        "x": 151,
+        "y": 384,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 121,
+        "y": 217,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 51,
+        "y": 217,
+        "timeout": -1
+      },
+      {
+        "name": 56,
+        "x": 127,
+        "y": 388,
+        "group": 0,
+        "angle": 0
+      },
+      {
+        "name": 56,
+        "x": 194,
+        "y": 218,
+        "group": 0,
+        "angle": -180
+      },
+      {
+        "name": 4,
+        "x": 51,
+        "y": 247,
+        "locale": "en",
+        "text": "Drop the candy into the sock and it will fall out from the other one",
+        "width": 220
+      },
+      {
+        "name": 8,
+        "x": 160,
+        "y": 236,
+        "locale": "en",
+        "angle": 180,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 8,
+        "x": 160,
+        "y": 237,
+        "locale": "fr",
+        "angle": 180,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 52,
+        "y": 246,
+        "locale": "fr",
+        "text": "Dépose le bonbon dans la chaussette et il sortira de l'autre chaussette",
+        "width": 160
+      }
+    ],
+    "de": [
+      {
+        "name": 8,
+        "x": 161,
+        "y": 237,
+        "locale": "de",
+        "angle": 180,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 54,
+        "y": 245,
+        "locale": "de",
+        "text": "Wirf die Süßigkeiten in den Strumpf und sie kommt aus dem anderen wieder heraus",
+        "width": 200
+      }
+    ],
+    "ru": [
+      {
+        "name": 8,
+        "x": 160,
+        "y": 236,
+        "locale": "ru",
+        "angle": 180,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 52,
+        "y": 242,
+        "locale": "ru",
+        "text": "Киньте леденец в один из носков, и он вылетит из другого",
+        "width": 160
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T05:23:54.654Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-110\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-111.json

@@ -0,0 +1,154 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-111",
+  "level_file": "data/task/rope-111.json",
+  "video": {
+    "path": "data/video/rope-111.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 2,
+    "object_counts": {
+      "bubble": 3,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "pump": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 2,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 248,
+        "y": 65
+      },
+      {
+        "name": 100,
+        "x": 183,
+        "y": 219,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 32,
+        "y": 218,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 52,
+        "x": 113,
+        "y": 317
+      },
+      {
+        "name": 3,
+        "x": 108,
+        "y": 244,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 254,
+        "y": 252,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 170,
+        "y": 122,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 113,
+        "y": 182
+      },
+      {
+        "name": 54,
+        "x": 109,
+        "y": 242
+      },
+      {
+        "name": 54,
+        "x": 256,
+        "y": 310
+      },
+      {
+        "name": 55,
+        "x": 51,
+        "y": 160,
+        "angle": 0
+      },
+      {
+        "name": 56,
+        "x": 107,
+        "y": 408,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 254,
+        "y": 371,
+        "group": 0,
+        "angle": 270
+      }
+    ],
+    "mirroredFrom": "00-11.json"
+  },
+  "reference_solution": "cut_rope 0,1 \npop_bubble 2 when candy_y < 400\nactivate_pump 0 when candy_y < 350",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T09:36:55.902Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1 \npop_bubble 2 when candy_y < 400\nactivate_pump 0 when candy_y < 350",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1 \npop_bubble 2 when candy_y < 400\nactivate_pump 0 when candy_y < 350",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-111\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 3\n  - pump: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-112.json

@@ -0,0 +1,201 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-112",
+  "level_file": "data/task/rope-112.json",
+  "video": {
+    "path": "data/video/rope-112.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 3,
+      "object_13": 1,
+      "object_4": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 69,
+        "y": 203
+      },
+      {
+        "name": 100,
+        "x": 52,
+        "y": 70,
+        "length": 90,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 162,
+        "y": 69,
+        "length": 170,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 264,
+        "y": 419
+      },
+      {
+        "name": 3,
+        "x": 54,
+        "y": 250,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 53,
+        "y": 369,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 280,
+        "y": 250,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 275,
+        "y": 69,
+        "length": 320,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 4,
+        "x": 133,
+        "y": 285,
+        "locale": "en",
+        "text": "Collect as many stars as you can",
+        "width": 150
+      },
+      {
+        "name": 13,
+        "x": 113,
+        "y": 321,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "ru": [
+      {
+        "name": 13,
+        "x": 94,
+        "y": 315,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 103,
+        "y": 278,
+        "locale": "ru",
+        "text": "Старайтесь собирать звезды леденцом",
+        "width": 210
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 137,
+        "y": 285,
+        "locale": "fr",
+        "text": "Collecte autant d'étoiles que tu le peux",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 113,
+        "y": 321,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 126,
+        "y": 285,
+        "locale": "de",
+        "text": "Sammle möglichst viele Sternchen",
+        "width": 170
+      },
+      {
+        "name": 13,
+        "x": 106,
+        "y": 321,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0\ncut_rope 1 when candy_near 700,580,50 times 2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:51:22.171Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0\ncut_rope 1 when candy_near 700,580,50 times 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0\ncut_rope 1 when candy_near 700,580,50 times 2",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-112\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 3\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-113.json

@@ -0,0 +1,192 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-113",
+  "level_file": "data/task/rope-113.json",
+  "video": {
+    "path": "data/video/rope-113.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "object_4": 1,
+      "object_8": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 248,
+        "y": 384
+      },
+      {
+        "name": 56,
+        "x": 59,
+        "y": 367,
+        "group": 0,
+        "angle": -90
+      },
+      {
+        "name": 56,
+        "x": 254,
+        "y": 114,
+        "group": 0,
+        "angle": 90
+      },
+      {
+        "name": 52,
+        "x": 55,
+        "y": 173
+      },
+      {
+        "name": 100,
+        "x": 56,
+        "y": 69,
+        "length": 90,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": 0,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 4,
+        "x": 76,
+        "y": 209,
+        "locale": "en",
+        "text": "Drop the candy into the magic hat and it will fall out from the other one",
+        "width": 160
+      },
+      {
+        "name": 8,
+        "x": 118,
+        "y": 347,
+        "locale": "en",
+        "angle": 0,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 3,
+        "x": 56,
+        "y": 328,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 259,
+        "y": 177,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 250,
+        "y": 319,
+        "timeout": -1
+      }
+    ],
+    "fr": [
+      {
+        "name": 8,
+        "x": 118,
+        "y": 347,
+        "locale": "fr",
+        "angle": 0,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 76,
+        "y": 201,
+        "locale": "fr",
+        "text": "Dépose le bonbon dans le chapeau magique et il tombera de l'autre chapeau",
+        "width": 160
+      }
+    ],
+    "de": [
+      {
+        "name": 8,
+        "x": 118,
+        "y": 348,
+        "locale": "de",
+        "angle": 0,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 72,
+        "y": 197,
+        "locale": "de",
+        "text": "Wirf den Bonbon in den magischen Hut und er kommt aus einem anderen wieder raus",
+        "width": 160
+      }
+    ],
+    "ru": [
+      {
+        "name": 8,
+        "x": 122,
+        "y": 354,
+        "locale": "ru",
+        "angle": 0,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 80,
+        "y": 230,
+        "locale": "ru",
+        "text": "Киньте конфету в одну из волшебных шляп, и она вылетит из другой",
+        "width": 160
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:23:40.902Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-113\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-114.json

@@ -0,0 +1,201 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-114",
+  "level_file": "data/task/rope-114.json",
+  "video": {
+    "path": "data/video/rope-114.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 2,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "object_13": 1,
+      "object_4": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 256,
+        "y": 310
+      },
+      {
+        "name": 2,
+        "x": 76,
+        "y": 77
+      },
+      {
+        "name": 100,
+        "x": 254,
+        "y": 248,
+        "length": 60,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 54,
+        "x": 255,
+        "y": 424
+      },
+      {
+        "name": 100,
+        "x": 177,
+        "y": 159,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 75,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 54,
+        "x": 82,
+        "y": 211
+      },
+      {
+        "name": 3,
+        "x": 255,
+        "y": 425,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 257,
+        "y": 205,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 81,
+        "y": 211,
+        "timeout": -1
+      },
+      {
+        "name": 4,
+        "x": 185,
+        "y": 336,
+        "locale": "en",
+        "text": "You can skip the level in the pause menu",
+        "width": 140
+      },
+      {
+        "name": 13,
+        "x": 120,
+        "y": 322,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 137,
+        "y": 312,
+        "locale": "ru",
+        "text": "Вы можете пропустить уровень через меню паузы",
+        "width": 140
+      },
+      {
+        "name": 13,
+        "x": 202,
+        "y": 297,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 135,
+        "y": 337,
+        "locale": "fr",
+        "text": "Tu peux passer le niveau via le menu Pause",
+        "width": 140
+      },
+      {
+        "name": 13,
+        "x": 200,
+        "y": 323,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 141,
+        "y": 318,
+        "locale": "de",
+        "text": "Im Pause-Menü kannst Du den Level überspringen",
+        "width": 170
+      },
+      {
+        "name": 13,
+        "x": 224,
+        "y": 303,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "mirroredFrom": "01-14.json"
+  },
+  "reference_solution": "cut_rope 0 \npop_bubble 0 when candy_y < 300",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T13:13:28.866Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 \npop_bubble 0 when candy_y < 300",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \npop_bubble 0 when candy_y < 300",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-114\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-115.json

@@ -0,0 +1,130 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-115",
+  "level_file": "data/task/rope-115.json",
+  "video": {
+    "path": "data/video/rope-115.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 3,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 280,
+        "y": 221
+      },
+      {
+        "name": 100,
+        "x": 361,
+        "y": 281,
+        "length": 95,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 363,
+        "y": 120,
+        "length": 95,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 2,
+        "x": 253,
+        "y": 387
+      },
+      {
+        "name": 3,
+        "x": 481,
+        "y": 318,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 363,
+        "y": 403,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 415,
+        "y": 223,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 108,
+        "y": 169,
+        "length": 140,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "mirroredFrom": "01-03.json"
+  },
+  "reference_solution": "cut_rope 2 \ncut_rope 1 when candy_near 1200,500,40",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:52:21.093Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2 \ncut_rope 1 when candy_near 1200,500,40",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2 \ncut_rope 1 when candy_near 1200,500,40",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-115\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 3\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-116.json

@@ -0,0 +1,186 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-116",
+  "level_file": "data/task/rope-116.json",
+  "video": {
+    "path": "data/video/rope-116.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bouncer": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 41,
+        "y": 333
+      },
+      {
+        "name": 52,
+        "x": 113,
+        "y": 149
+      },
+      {
+        "name": 100,
+        "x": 194,
+        "y": 55,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 3,
+        "x": 171,
+        "y": 234,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 106,
+        "y": 286,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 264,
+        "y": 225,
+        "timeout": -1
+      },
+      {
+        "name": 81,
+        "x": 282,
+        "y": 261,
+        "angle": 140,
+        "size": 1
+      }
+    ],
+    "w": [
+      {
+        "name": 8,
+        "x": 56,
+        "y": 322,
+        "locale": "en",
+        "angle": 90,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 78,
+        "y": 302,
+        "locale": "en",
+        "text": "Candy will bounce away from this platform",
+        "width": 120
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 81,
+        "y": 302,
+        "locale": "ru",
+        "text": "Леденец отскакивает от таких платформ",
+        "width": 100
+      },
+      {
+        "name": 8,
+        "x": 56,
+        "y": 323,
+        "locale": "ru",
+        "angle": 90,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 8,
+        "x": 65,
+        "y": 323,
+        "locale": "de",
+        "angle": 90,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 91,
+        "y": 297,
+        "locale": "de",
+        "text": "Von dieser Plattform prallt der Bonbon ab",
+        "width": 90
+      }
+    ],
+    "fr": [
+      {
+        "name": 8,
+        "x": 57,
+        "y": 323,
+        "locale": "fr",
+        "angle": 90,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 79,
+        "y": 301,
+        "locale": "fr",
+        "text": "Le bonbon rebondira hors de cette plateforme",
+        "width": 120
+      }
+    ],
+    "mirroredFrom": "06-01.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_x > 1080",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:10:40.704Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_x > 1080",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_x > 1080",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-116\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-117.json

@@ -0,0 +1,170 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-117",
+  "level_file": "data/task/rope-117.json",
+  "video": {
+    "path": "data/video/rope-117.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 7,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "special": 1,
+        "ropePhysicsSpeed": 1
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 162,
+        "y": 430
+      },
+      {
+        "name": 100,
+        "x": 101,
+        "y": 271,
+        "length": 100,
+        "wheel": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 36,
+        "y": 148,
+        "length": 100,
+        "wheel": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 291,
+        "y": 149,
+        "length": 100,
+        "wheel": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 227,
+        "y": 273,
+        "length": 100,
+        "wheel": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 52,
+        "x": 159,
+        "y": 159
+      },
+      {
+        "name": 3,
+        "x": 96,
+        "y": 379,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 29,
+        "y": 246,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 225,
+        "y": 377,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 98,
+        "y": 79,
+        "length": 90,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 226,
+        "y": 80,
+        "length": 90,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 162,
+        "y": 34,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 6,5 \ncut_rope 4 when candy_x < 700\ncut_rope 1 when candy_still for 0.3\ncut_rope 0 when candy_still for 0.3 \ncut_rope 3 when candy_near 1070,870,60 times 2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T13:07:13.482Z"
+  },
+  "answer": {
+    "commands": "cut_rope 6,5 \ncut_rope 4 when candy_x < 700\ncut_rope 1 when candy_still for 0.3\ncut_rope 0 when candy_still for 0.3 \ncut_rope 3 when candy_near 1070,870,60 times 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 6,5 \ncut_rope 4 when candy_x < 700\ncut_rope 1 when candy_still for 0.3\ncut_rope 0 when candy_still for 0.3 \ncut_rope 3 when candy_near 1070,870,60 times 2",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-117\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 7\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-118.json

@@ -0,0 +1,134 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-118",
+  "level_file": "data/task/rope-118.json",
+  "video": {
+    "path": "data/video/rope-118.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 4,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "special": 1,
+        "ropePhysicsSpeed": 1
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 158,
+        "y": 187
+      },
+      {
+        "name": 2,
+        "x": 262,
+        "y": 362
+      },
+      {
+        "name": 100,
+        "x": 161,
+        "y": 315,
+        "length": 93,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 289,
+        "y": 186,
+        "length": 105,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 162,
+        "y": 57,
+        "length": 93,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 33,
+        "y": 186,
+        "length": 105,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 226,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 32,
+        "y": 312,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 161,
+        "y": 434,
+        "timeout": -1
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_x < 720",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:54:15.133Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_x < 720",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_x < 720",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-118\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 4\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-119.json

@@ -0,0 +1,123 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-119",
+  "level_file": "data/task/rope-119.json",
+  "video": {
+    "path": "data/video/rope-119.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 2,
+    "object_counts": {
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 2,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 162,
+        "y": 430
+      },
+      {
+        "name": 54,
+        "x": 159,
+        "y": 320
+      },
+      {
+        "name": 100,
+        "x": 97,
+        "y": 122,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 52,
+        "x": 159,
+        "y": 231
+      },
+      {
+        "name": 3,
+        "x": 161,
+        "y": 75,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 162,
+        "y": 170,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 275,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 222,
+        "y": 122,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ],
+    "levelId": "00-15",
+    "levelEditedAt": "2026-05-05T10:12:06.544Z"
+  },
+  "reference_solution": "cut_rope 0,1 \npop_bubble 0 when candy_y < 190",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:12:53.720Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1 \npop_bubble 0 when candy_y < 190",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1 \npop_bubble 0 when candy_y < 190",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-119\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-120.json

@@ -0,0 +1,133 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-120",
+  "level_file": "data/task/rope-120.json",
+  "video": {
+    "path": "data/video/rope-120.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "gravity_button": 1,
+      "pump": 3,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 55,
+        "x": 600,
+        "y": 398,
+        "angle": 180
+      },
+      {
+        "name": 55,
+        "x": 483,
+        "y": 98,
+        "angle": 180
+      },
+      {
+        "name": 55,
+        "x": 316,
+        "y": 398,
+        "angle": 180
+      },
+      {
+        "name": 52,
+        "x": 482,
+        "y": 301
+      },
+      {
+        "name": 54,
+        "x": 479,
+        "y": 391
+      },
+      {
+        "name": 100,
+        "x": 478,
+        "y": 164,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 2,
+        "x": 89,
+        "y": 64
+      },
+      {
+        "name": 3,
+        "x": 354,
+        "y": 94,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 89,
+        "y": 166,
+        "timeout": 20
+      },
+      {
+        "name": 3,
+        "x": 172,
+        "y": 399,
+        "timeout": -1
+      },
+      {
+        "name": 53,
+        "x": 261,
+        "y": 228
+      }
+    ],
+    "mirroredFrom": "08-05.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 1320,670,60\nactivate_pump 0 when candy_y > 880\nactivate_pump 0 when candy_y > 890\ntoggle_gravity when candy_x < 1080\ntoggle_gravity when candy_x < 1060\nactivate_pump 1 when candy_y < 240\nactivate_pump 1\ntoggle_gravity when candy_y < 50\nactivate_pump 2 when candy_y > 850\nactivate_pump 2 when candy_y > 910\ntoggle_gravity when candy_y > 1000",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:02:46.931Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 1320,670,60\nactivate_pump 0 when candy_y > 880\nactivate_pump 0 when candy_y > 890\ntoggle_gravity when candy_x < 1080\ntoggle_gravity when candy_x < 1060\nactivate_pump 1 when candy_y < 240\nactivate_pump 1\ntoggle_gravity when candy_y < 50\nactivate_pump 2 when candy_y > 850\nactivate_pump 2 when candy_y > 910\ntoggle_gravity when candy_y > 1000",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 1320,670,60\nactivate_pump 0 when candy_y > 880\nactivate_pump 0 when candy_y > 890\ntoggle_gravity when candy_x < 1080\ntoggle_gravity when candy_x < 1060\nactivate_pump 1 when candy_y < 240\nactivate_pump 1\ntoggle_gravity when candy_y < 50\nactivate_pump 2 when candy_y > 850\nactivate_pump 2 when candy_y > 910\ntoggle_gravity when candy_y > 1000",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-120\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n  - bubble: 1\n  - pump: 3\n  - gravity switch: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n- Gravity switch: `toggle_gravity` reverses or rotates gravity. Often triggered after the candy crosses a coordinate threshold.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-121.json

@@ -0,0 +1,199 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-121",
+  "level_file": "data/task/rope-121.json",
+  "video": {
+    "path": "data/video/rope-121.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "gravity_button": 1,
+      "object_4": 1,
+      "object_8": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 303,
+        "y": 214
+      },
+      {
+        "name": 100,
+        "x": 383,
+        "y": 175,
+        "length": 65,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 215,
+        "y": 175,
+        "length": 65,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 395,
+        "y": 386
+      },
+      {
+        "name": 3,
+        "x": 431,
+        "y": 81,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 302,
+        "y": 85,
+        "timeout": -1
+      },
+      {
+        "name": 53,
+        "x": 150,
+        "y": 392
+      },
+      {
+        "name": 4,
+        "x": 215,
+        "y": 331,
+        "locale": "en",
+        "text": "Toggle the gravity with this button",
+        "width": 140
+      },
+      {
+        "name": 8,
+        "x": 222,
+        "y": 429,
+        "locale": "en",
+        "angle": 35,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 3,
+        "x": 302,
+        "y": 273,
+        "timeout": -1
+      }
+    ],
+    "ru": [
+      {
+        "name": 8,
+        "x": 219,
+        "y": 417,
+        "locale": "ru",
+        "angle": 40,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 229,
+        "y": 328,
+        "locale": "ru",
+        "text": "Переключайте гравитацию с помощью этой кнопки",
+        "width": 140
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 180,
+        "y": 338,
+        "locale": "fr",
+        "text": "Active et désactive la gravité avec ce bouton",
+        "width": 200
+      },
+      {
+        "name": 8,
+        "x": 211,
+        "y": 407,
+        "locale": "fr",
+        "angle": 35,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 8,
+        "x": 207,
+        "y": 425,
+        "locale": "de",
+        "angle": 55,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 208,
+        "y": 342,
+        "locale": "de",
+        "text": "Hier kannst Du die Schwerkraft ein oder aus schalten",
+        "width": 140
+      }
+    ],
+    "levelId": "08-01",
+    "levelEditedAt": "2026-05-05T17:53:03.499Z"
+  },
+  "reference_solution": "toggle_gravity \ncut_rope 1 when candy_still for 0.3 \ntoggle_gravity when candy_near 1200,190,60\ncut_rope 0 when candy_near 1050,650,60 times 2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:53:13.567Z"
+  },
+  "answer": {
+    "commands": "toggle_gravity \ncut_rope 1 when candy_still for 0.3 \ntoggle_gravity when candy_near 1200,190,60\ncut_rope 0 when candy_near 1050,650,60 times 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "toggle_gravity \ncut_rope 1 when candy_still for 0.3 \ntoggle_gravity when candy_near 1200,190,60\ncut_rope 0 when candy_near 1050,650,60 times 2",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-121\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - gravity switch: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Gravity switch: `toggle_gravity` reverses or rotates gravity. Often triggered after the candy crosses a coordinate threshold.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-122.json

@@ -0,0 +1,165 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-122",
+  "level_file": "data/task/rope-122.json",
+  "video": {
+    "path": "data/video/rope-122.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 2,
+    "object_counts": {
+      "bubble": 2,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "pump": 1,
+      "sock": 4,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 2,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 275,
+        "y": 76
+      },
+      {
+        "name": 55,
+        "x": 48,
+        "y": 329,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 50,
+        "y": 102,
+        "group": 1,
+        "angle": 90
+      },
+      {
+        "name": 56,
+        "x": 159,
+        "y": 399,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 54,
+        "x": 164,
+        "y": 160
+      },
+      {
+        "name": 56,
+        "x": 273,
+        "y": 398,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 167,
+        "y": 105,
+        "group": 1,
+        "angle": 90
+      },
+      {
+        "name": 3,
+        "x": 280,
+        "y": 159,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 162,
+        "y": 161,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 49,
+        "y": 119,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 80,
+        "y": 144,
+        "length": 30,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 52,
+        "x": 50,
+        "y": 182
+      },
+      {
+        "name": 54,
+        "x": 278,
+        "y": 158
+      },
+      {
+        "name": 100,
+        "x": 19,
+        "y": 143,
+        "length": 30,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "levelId": "00-22",
+    "levelEditedAt": "2026-05-05T10:43:19.434Z",
+    "mirroredFrom": "00-22.json"
+  },
+  "reference_solution": "cut_rope 0,1 \nactivate_pump 0 until candy_y < 300\npop_bubble 0 when candy_y < 300",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:45:53.636Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1 \nactivate_pump 0 until candy_y < 300\npop_bubble 0 when candy_y < 300",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1 \nactivate_pump 0 until candy_y < 300\npop_bubble 0 when candy_y < 300",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-122\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 2\n  - pump: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-123.json

@@ -0,0 +1,154 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-123",
+  "level_file": "data/task/rope-123.json",
+  "video": {
+    "path": "data/video/rope-123.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "sock": 4,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 56,
+        "x": 35,
+        "y": 340,
+        "group": 0,
+        "angle": 290
+      },
+      {
+        "name": 56,
+        "x": 101,
+        "y": 373,
+        "group": 1,
+        "angle": 290
+      },
+      {
+        "name": 56,
+        "x": 220,
+        "y": 373,
+        "group": 0,
+        "angle": 250
+      },
+      {
+        "name": 56,
+        "x": 290,
+        "y": 340,
+        "group": 1,
+        "angle": 250
+      },
+      {
+        "name": 52,
+        "x": 162,
+        "y": 164
+      },
+      {
+        "name": 100,
+        "x": 106,
+        "y": 47,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 228,
+        "y": 50,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 2,
+        "x": 86,
+        "y": 103
+      },
+      {
+        "name": 3,
+        "x": 165,
+        "y": 329,
+        "timeout": 15
+      },
+      {
+        "name": 3,
+        "x": 138,
+        "y": 293,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 261,
+        "y": 300,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 161,
+        "y": 332
+      }
+    ],
+    "levelId": "09-07",
+    "levelEditedAt": "2026-05-05T18:27:32.492Z"
+  },
+  "reference_solution": "cut_rope 1 when candy_still for 0.3 and candy_near 970,380,60\ncut_rope 0 when candy_x < 930",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:27:46.690Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 when candy_still for 0.3 and candy_near 970,380,60\ncut_rope 0 when candy_x < 930",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 when candy_still for 0.3 and candy_near 970,380,60\ncut_rope 0 when candy_x < 930",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-123\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-124.json

@@ -0,0 +1,202 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-124",
+  "level_file": "data/task/rope-124.json",
+  "video": {
+    "path": "data/video/rope-124.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 3,
+      "object_13": 1,
+      "object_4": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 251,
+        "y": 203
+      },
+      {
+        "name": 100,
+        "x": 268,
+        "y": 70,
+        "length": 90,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 158,
+        "y": 69,
+        "length": 170,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 2,
+        "x": 56,
+        "y": 419
+      },
+      {
+        "name": 3,
+        "x": 266,
+        "y": 250,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 267,
+        "y": 369,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 40,
+        "y": 250,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 45,
+        "y": 69,
+        "length": 320,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 4,
+        "x": 187,
+        "y": 285,
+        "locale": "en",
+        "text": "Collect as many stars as you can",
+        "width": 150
+      },
+      {
+        "name": 13,
+        "x": 207,
+        "y": 321,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "ru": [
+      {
+        "name": 13,
+        "x": 94,
+        "y": 315,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 103,
+        "y": 278,
+        "locale": "ru",
+        "text": "Старайтесь собирать звезды леденцом",
+        "width": 210
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 137,
+        "y": 285,
+        "locale": "fr",
+        "text": "Collecte autant d'étoiles que tu le peux",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 113,
+        "y": 321,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 126,
+        "y": 285,
+        "locale": "de",
+        "text": "Sammle möglichst viele Sternchen",
+        "width": 170
+      },
+      {
+        "name": 13,
+        "x": 106,
+        "y": 321,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "mirroredFrom": "01-02.json"
+  },
+  "reference_solution": "cut_rope 0\ncut_rope 1 when candy_near 1220,580,50 times 2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:51:22.171Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0\ncut_rope 1 when candy_near 1220,580,50 times 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0\ncut_rope 1 when candy_near 1220,580,50 times 2",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-124\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 3\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-125.json

@@ -0,0 +1,142 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-125",
+  "level_file": "data/task/rope-125.json",
+  "video": {
+    "path": "data/video/rope-125.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 2,
+    "object_counts": {
+      "bubble": 2,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 2,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 159,
+        "y": 429
+      },
+      {
+        "name": 56,
+        "x": 60,
+        "y": 343,
+        "group": 0,
+        "angle": 290
+      },
+      {
+        "name": 56,
+        "x": 262,
+        "y": 168,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 3,
+        "x": 259,
+        "y": 50,
+        "timeout": -1
+      },
+      {
+        "name": 52,
+        "x": 91,
+        "y": 200
+      },
+      {
+        "name": 3,
+        "x": 160,
+        "y": 122,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 250,
+        "y": 288
+      },
+      {
+        "name": 3,
+        "x": 250,
+        "y": 289,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 160,
+        "y": 81,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 158,
+        "y": 289,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 40,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 160,
+        "y": 367
+      }
+    ],
+    "mirroredFrom": "00-21.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_x > 1090\npop_bubble 0 when candy_y < 100\ncut_rope 1 when candy_still for 0.3\npop_bubble 1 when candy_near 970,250,50",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:36:45.808Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_x > 1090\npop_bubble 0 when candy_y < 100\ncut_rope 1 when candy_still for 0.3\npop_bubble 1 when candy_near 970,250,50",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_x > 1090\npop_bubble 0 when candy_y < 100\ncut_rope 1 when candy_still for 0.3\npop_bubble 1 when candy_near 970,250,50",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-125\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-126.json

@@ -0,0 +1,149 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-126",
+  "level_file": "data/task/rope-126.json",
+  "video": {
+    "path": "data/video/rope-126.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "pump": 2,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 112,
+        "y": 352
+      },
+      {
+        "name": 100,
+        "x": 50,
+        "y": 275,
+        "length": 70,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 177,
+        "y": 278,
+        "length": 70,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 271,
+        "y": 90
+      },
+      {
+        "name": 56,
+        "x": 270,
+        "y": 384,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 49,
+        "y": 97,
+        "group": 0,
+        "angle": 90
+      },
+      {
+        "name": 55,
+        "x": 112,
+        "y": 418,
+        "angle": 270
+      },
+      {
+        "name": 3,
+        "x": 115,
+        "y": 230,
+        "timeout": -1
+      },
+      {
+        "name": 55,
+        "x": 113,
+        "y": 305,
+        "angle": 270
+      },
+      {
+        "name": 3,
+        "x": 48,
+        "y": 188,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 185,
+        "y": 359,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 47,
+        "y": 188
+      }
+    ],
+    "levelId": "04-06",
+    "levelEditedAt": "2026-05-05T15:30:05.369Z"
+  },
+  "reference_solution": "activate_pump 0 when candy_still for 0.3 and candy_near 850,810,60\nactivate_pump 1 when candy_y < 660\ncut_rope 2\ncut_rope 0 when candy_y > 800\ncut_rope 1 when candy_x > 1110",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:30:18.013Z"
+  },
+  "answer": {
+    "commands": "activate_pump 0 when candy_still for 0.3 and candy_near 850,810,60\nactivate_pump 1 when candy_y < 660\ncut_rope 2\ncut_rope 0 when candy_y > 800\ncut_rope 1 when candy_x > 1110",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "activate_pump 0 when candy_still for 0.3 and candy_near 850,810,60\nactivate_pump 1 when candy_y < 660\ncut_rope 2\ncut_rope 0 when candy_y > 800\ncut_rope 1 when candy_x > 1110",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-126\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 1\n  - pump: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-127.json

@@ -0,0 +1,200 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-127",
+  "level_file": "data/task/rope-127.json",
+  "video": {
+    "path": "data/video/rope-127.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "gravity_button": 1,
+      "object_4": 1,
+      "object_8": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 337,
+        "y": 214
+      },
+      {
+        "name": 100,
+        "x": 257,
+        "y": 175,
+        "length": 65,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 425,
+        "y": 175,
+        "length": 65,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 2,
+        "x": 245,
+        "y": 386
+      },
+      {
+        "name": 3,
+        "x": 209,
+        "y": 81,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 338,
+        "y": 85,
+        "timeout": -1
+      },
+      {
+        "name": 53,
+        "x": 490,
+        "y": 392
+      },
+      {
+        "name": 4,
+        "x": 425,
+        "y": 331,
+        "locale": "en",
+        "text": "Toggle the gravity with this button",
+        "width": 140
+      },
+      {
+        "name": 8,
+        "x": 418,
+        "y": 429,
+        "locale": "en",
+        "angle": 145,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 3,
+        "x": 338,
+        "y": 273,
+        "timeout": -1
+      }
+    ],
+    "ru": [
+      {
+        "name": 8,
+        "x": 219,
+        "y": 417,
+        "locale": "ru",
+        "angle": 40,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 229,
+        "y": 328,
+        "locale": "ru",
+        "text": "Переключайте гравитацию с помощью этой кнопки",
+        "width": 140
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 180,
+        "y": 338,
+        "locale": "fr",
+        "text": "Active et désactive la gravité avec ce bouton",
+        "width": 200
+      },
+      {
+        "name": 8,
+        "x": 211,
+        "y": 407,
+        "locale": "fr",
+        "angle": 35,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 8,
+        "x": 207,
+        "y": 425,
+        "locale": "de",
+        "angle": 55,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 208,
+        "y": 342,
+        "locale": "de",
+        "text": "Hier kannst Du die Schwerkraft ein oder aus schalten",
+        "width": 140
+      }
+    ],
+    "levelId": "08-01",
+    "levelEditedAt": "2026-05-05T17:53:03.499Z",
+    "mirroredFrom": "08-01.json"
+  },
+  "reference_solution": "toggle_gravity \ncut_rope 1 when candy_still for 0.3 \ntoggle_gravity when candy_near 720,190,60\ncut_rope 0 when candy_near 870,650,60 times 2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:53:13.567Z"
+  },
+  "answer": {
+    "commands": "toggle_gravity \ncut_rope 1 when candy_still for 0.3 \ntoggle_gravity when candy_near 720,190,60\ncut_rope 0 when candy_near 870,650,60 times 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "toggle_gravity \ncut_rope 1 when candy_still for 0.3 \ntoggle_gravity when candy_near 720,190,60\ncut_rope 0 when candy_near 870,650,60 times 2",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-127\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - gravity switch: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Gravity switch: `toggle_gravity` reverses or rotates gravity. Often triggered after the candy crosses a coordinate threshold.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-128.json

@@ -0,0 +1,184 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-128",
+  "level_file": "data/task/rope-128.json",
+  "video": {
+    "path": "data/video/rope-128.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": true,
+    "special": 1,
+    "object_counts": {
+      "grab_or_rope_anchor": 6,
+      "left_candy": 1,
+      "right_candy": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 100,
+        "x": 259,
+        "y": 64,
+        "length": 200,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 158,
+        "y": 64,
+        "length": 30,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 57,
+        "y": 63,
+        "length": 200,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 55,
+        "y": 384,
+        "length": 200,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 160,
+        "y": 385,
+        "length": 30,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 261,
+        "y": 385,
+        "length": 200,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 51,
+        "x": 156,
+        "y": 302
+      },
+      {
+        "name": 50,
+        "x": 154,
+        "y": 157
+      },
+      {
+        "name": 2,
+        "x": 161,
+        "y": 444
+      },
+      {
+        "name": 3,
+        "x": 192,
+        "y": 335,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 192,
+        "y": 128,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 163,
+        "y": 228,
+        "timeout": -1
+      }
+    ],
+    "mirroredFrom": "09-17.json"
+  },
+  "reference_solution": "cut_rope 4 when right_candy_still for 0.3 \ncut_rope 3 when candy_still for 0.3 and candy_near 960,330,60\ncut_rope 1 when candy_still for 0.3 and candy_near 1000,310,60\ncut_rope 2 when candy_still for 0.3 and candy_near 950,620,60\ncut_rope 5 when candy_x > 1370\ncut_rope 0 when candy_near 1120,700,50 times 3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:46:01.088Z"
+  },
+  "answer": {
+    "commands": "cut_rope 4 when right_candy_still for 0.3 \ncut_rope 3 when candy_still for 0.3 and candy_near 960,330,60\ncut_rope 1 when candy_still for 0.3 and candy_near 1000,310,60\ncut_rope 2 when candy_still for 0.3 and candy_near 950,620,60\ncut_rope 5 when candy_x > 1370\ncut_rope 0 when candy_near 1120,700,50 times 3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 4 when right_candy_still for 0.3 \ncut_rope 3 when candy_still for 0.3 and candy_near 960,330,60\ncut_rope 1 when candy_still for 0.3 and candy_near 1000,310,60\ncut_rope 2 when candy_still for 0.3 and candy_near 950,620,60\ncut_rope 5 when candy_x > 1370\ncut_rope 0 when candy_near 1120,700,50 times 3",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-128\n- canvas_size: 1920 x 1080\n- two_parts (split candy): True\n- object_counts:\n  - target: 1\n  - star: 3\n  - left split candy: 1\n  - right split candy: 1\n  - rope/grab: 6\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Split candy: the level starts with left_candy and right_candy halves. Before they merge, use left_candy_* and right_candy_* conditions. After they touch and merge, use candy_* for the complete candy.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-129.json

@@ -0,0 +1,146 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-129",
+  "level_file": "data/task/rope-129.json",
+  "video": {
+    "path": "data/video/rope-129.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": true,
+    "special": 1,
+    "object_counts": {
+      "grab_or_rope_anchor": 4,
+      "left_candy": 1,
+      "right_candy": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 162,
+        "y": 424
+      },
+      {
+        "name": 50,
+        "x": 117,
+        "y": 204
+      },
+      {
+        "name": 51,
+        "x": 239,
+        "y": 207
+      },
+      {
+        "name": 3,
+        "x": 124,
+        "y": 292,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 254,
+        "y": 298,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 180,
+        "y": 360,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 237,
+        "y": 72,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 110,
+        "y": 78,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 171,
+        "y": 140,
+        "length": 136,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 168,
+        "y": 143,
+        "length": 154,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ],
+    "generatorTemplate": "split_drop"
+  },
+  "reference_solution": "cut_rope 0,1 \ncut_rope 2 when candy_still for 0.3 and candy_near 1000,700,60\ncut_rope 3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T19:04:09.649Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1 \ncut_rope 2 when candy_still for 0.3 and candy_near 1000,700,60\ncut_rope 3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1 \ncut_rope 2 when candy_still for 0.3 and candy_near 1000,700,60\ncut_rope 3",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-129\n- canvas_size: 1920 x 1080\n- two_parts (split candy): True\n- object_counts:\n  - target: 1\n  - star: 3\n  - left split candy: 1\n  - right split candy: 1\n  - rope/grab: 4\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Split candy: the level starts with left_candy and right_candy halves. Before they merge, use left_candy_* and right_candy_* conditions. After they touch and merge, use candy_* for the complete candy.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-130.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-130",
+  "level_file": "data/task/rope-130.json",
+  "video": {
+    "path": "data/video/rope-130.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 4,
+      "gravity_button": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 160,
+        "y": 218
+      },
+      {
+        "name": 2,
+        "x": 232,
+        "y": 382
+      },
+      {
+        "name": 53,
+        "x": 61,
+        "y": 402
+      },
+      {
+        "name": 100,
+        "x": 281,
+        "y": 206,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 44,
+        "y": 204,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 163,
+        "y": 87,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 162,
+        "y": 322,
+        "length": 245,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 157,
+        "y": 179,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 320,
+        "y": 330,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 278,
+        "y": 95,
+        "timeout": -1
+      }
+    ],
+    "levelId": "08-12",
+    "levelEditedAt": "2026-05-05T18:13:11.996Z"
+  },
+  "reference_solution": "cut_rope 1,2 \ntoggle_gravity when candy_x < 980\ncut_rope 3 when candy_x < 1030\ntoggle_gravity when candy_x > 1070\ncut_rope 0 when candy_near 1300,760,50",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:13:17.137Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1,2 \ntoggle_gravity when candy_x < 980\ncut_rope 3 when candy_x < 1030\ntoggle_gravity when candy_x > 1070\ncut_rope 0 when candy_near 1300,760,50",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1,2 \ntoggle_gravity when candy_x < 980\ncut_rope 3 when candy_x < 1030\ntoggle_gravity when candy_x > 1070\ncut_rope 0 when candy_near 1300,760,50",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-130\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 4\n  - gravity switch: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Gravity switch: `toggle_gravity` reverses or rotates gravity. Often triggered after the candy crosses a coordinate threshold.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-131.json

@@ -0,0 +1,231 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-131",
+  "level_file": "data/task/rope-131.json",
+  "video": {
+    "path": "data/video/rope-131.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 2,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 4,
+      "object_13": 1,
+      "object_4": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 2,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 257,
+        "y": 296
+      },
+      {
+        "name": 52,
+        "x": 158,
+        "y": 292
+      },
+      {
+        "name": 100,
+        "x": 291,
+        "y": 397,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 26,
+        "y": 192,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 291,
+        "y": 192,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 26,
+        "y": 399,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 159,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 40,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 160,
+        "y": 359,
+        "timeout": -1
+      },
+      {
+        "name": 56,
+        "x": 155,
+        "y": 401,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 72,
+        "y": 292,
+        "group": 0,
+        "angle": 0
+      },
+      {
+        "name": 13,
+        "x": 247,
+        "y": 97,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 238,
+        "y": 63,
+        "locale": "en",
+        "text": "You can cut several ropes at once",
+        "width": 200
+      }
+    ],
+    "ru": [
+      {
+        "name": 13,
+        "x": 40,
+        "y": 96,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 60,
+        "y": 50,
+        "locale": "ru",
+        "text": "Вы можете перерезать несколько веревок одновременно",
+        "width": 240
+      }
+    ],
+    "fr": [
+      {
+        "name": 13,
+        "x": 39,
+        "y": 78,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 58,
+        "y": 46,
+        "locale": "fr",
+        "text": "Vous pouvez couper plusieurs cordes à la fois",
+        "width": 250
+      }
+    ],
+    "de": [
+      {
+        "name": 13,
+        "x": 46,
+        "y": 83,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 57,
+        "y": 46,
+        "locale": "de",
+        "text": "Du kannst mehrere Seile auf einmal durchschneiden",
+        "width": 250
+      }
+    ],
+    "mirroredFrom": "00-08.json"
+  },
+  "reference_solution": "cut_rope 0\ncut_rope 2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T05:33:53.017Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0\ncut_rope 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0\ncut_rope 2",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-131\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 4\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-132.json

@@ -0,0 +1,139 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-132",
+  "level_file": "data/task/rope-132.json",
+  "video": {
+    "path": "data/video/rope-132.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 4,
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 56,
+        "x": 334,
+        "y": 417,
+        "group": 1,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 171,
+        "y": 189,
+        "group": 1,
+        "angle": 350
+      },
+      {
+        "name": 2,
+        "x": 424,
+        "y": 179
+      },
+      {
+        "name": 3,
+        "x": 258,
+        "y": 179,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 338,
+        "y": 336,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 215,
+        "y": 180,
+        "timeout": -1
+      },
+      {
+        "name": 52,
+        "x": 345,
+        "y": 90
+      },
+      {
+        "name": 100,
+        "x": 341,
+        "y": 213,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 54,
+        "x": 343,
+        "y": 88
+      },
+      {
+        "name": 54,
+        "x": 338,
+        "y": 335
+      },
+      {
+        "name": 54,
+        "x": 341,
+        "y": 254
+      },
+      {
+        "name": 54,
+        "x": 340,
+        "y": 164
+      }
+    ],
+    "mirroredFrom": "09-24.json"
+  },
+  "reference_solution": "pop_bubble 0 when candy_still for 0.3 and candy_near 1010,150,60\npop_bubble 3 when candy_y > 310\ncut_rope 0 when candy_y > 370\npop_bubble 2 when candy_y > 570\npop_bubble 1 when candy_y < 180",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:53:54.627Z"
+  },
+  "answer": {
+    "commands": "pop_bubble 0 when candy_still for 0.3 and candy_near 1010,150,60\npop_bubble 3 when candy_y > 310\ncut_rope 0 when candy_y > 370\npop_bubble 2 when candy_y > 570\npop_bubble 1 when candy_y < 180",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "pop_bubble 0 when candy_still for 0.3 and candy_near 1010,150,60\npop_bubble 3 when candy_y > 310\ncut_rope 0 when candy_y > 370\npop_bubble 2 when candy_y > 570\npop_bubble 1 when candy_y < 180",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-132\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n  - bubble: 4\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-133.json

@@ -0,0 +1,150 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-133",
+  "level_file": "data/task/rope-133.json",
+  "video": {
+    "path": "data/video/rope-133.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": true,
+    "special": 1,
+    "object_counts": {
+      "bouncer": 3,
+      "grab_or_rope_anchor": 2,
+      "left_candy": 1,
+      "right_candy": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 272,
+        "y": 369
+      },
+      {
+        "name": 3,
+        "x": 41,
+        "y": 400,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 194,
+        "y": 397,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 118,
+        "y": 333,
+        "timeout": -1
+      },
+      {
+        "name": 82,
+        "x": 164,
+        "y": 440,
+        "angle": 180,
+        "size": 2
+      },
+      {
+        "name": 82,
+        "x": 266,
+        "y": 440,
+        "angle": 180,
+        "size": 2
+      },
+      {
+        "name": 81,
+        "x": 37,
+        "y": 435,
+        "angle": 188,
+        "size": 1
+      },
+      {
+        "name": 100,
+        "x": 189,
+        "y": 35,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 51,
+        "x": 188,
+        "y": 137
+      },
+      {
+        "name": 100,
+        "x": 44,
+        "y": 36,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 50,
+        "x": 42,
+        "y": 139
+      }
+    ],
+    "mirroredFrom": "09-10.json"
+  },
+  "reference_solution": "cut_rope 0,1",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:32:17.053Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-133\n- canvas_size: 1920 x 1080\n- two_parts (split candy): True\n- object_counts:\n  - target: 1\n  - star: 3\n  - left split candy: 1\n  - right split candy: 1\n  - rope/grab: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Split candy: the level starts with left_candy and right_candy halves. Before they merge, use left_candy_* and right_candy_* conditions. After they touch and merge, use candy_* for the complete candy.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-134.json

@@ -0,0 +1,137 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-134",
+  "level_file": "data/task/rope-134.json",
+  "video": {
+    "path": "data/video/rope-134.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": true,
+    "special": 1,
+    "object_counts": {
+      "bubble": 1,
+      "grab_or_rope_anchor": 2,
+      "left_candy": 1,
+      "pump": 1,
+      "right_candy": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 51,
+        "x": 338,
+        "y": 324
+      },
+      {
+        "name": 50,
+        "x": 403,
+        "y": 210
+      },
+      {
+        "name": 2,
+        "x": 319,
+        "y": 397
+      },
+      {
+        "name": 100,
+        "x": 405,
+        "y": 107,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 344,
+        "y": 227,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 401,
+        "y": 390
+      },
+      {
+        "name": 55,
+        "x": 232,
+        "y": 323,
+        "angle": 0
+      },
+      {
+        "name": 3,
+        "x": 333,
+        "y": 153,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 323,
+        "y": 110,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 400,
+        "y": 391,
+        "timeout": -1
+      }
+    ],
+    "levelId": "05-23",
+    "levelEditedAt": "2026-05-05T16:49:48.000Z",
+    "mirroredFrom": "05-23.json"
+  },
+  "reference_solution": "cut_rope 0 \nactivate_pump 0 when left_candy_near 1140,1000,50\ncut_rope 0 when candy_still for 0.3\npop_bubble 0 when candy_still for 0.3 and candy_near 1010,270,60\ncut_rope 1 when candy_y > 290",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T16:50:45.570Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 \nactivate_pump 0 when left_candy_near 1140,1000,50\ncut_rope 0 when candy_still for 0.3\npop_bubble 0 when candy_still for 0.3 and candy_near 1010,270,60\ncut_rope 1 when candy_y > 290",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \nactivate_pump 0 when left_candy_near 1140,1000,50\ncut_rope 0 when candy_still for 0.3\npop_bubble 0 when candy_still for 0.3 and candy_near 1010,270,60\ncut_rope 1 when candy_y > 290",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-134\n- canvas_size: 1920 x 1080\n- two_parts (split candy): True\n- object_counts:\n  - target: 1\n  - star: 3\n  - left split candy: 1\n  - right split candy: 1\n  - rope/grab: 2\n  - bubble: 1\n  - pump: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Split candy: the level starts with left_candy and right_candy halves. Before they merge, use left_candy_* and right_candy_* conditions. After they touch and merge, use candy_* for the complete candy.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-135.json

@@ -0,0 +1,245 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-135",
+  "level_file": "data/task/rope-135.json",
+  "video": {
+    "path": "data/video/rope-135.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 2,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 6,
+      "object_13": 1,
+      "object_4": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 2,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 253,
+        "y": 420
+      },
+      {
+        "name": 52,
+        "x": 159,
+        "y": 153
+      },
+      {
+        "name": 3,
+        "x": 158,
+        "y": 357,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 82,
+        "y": 358,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 238,
+        "y": 360,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 79,
+        "y": 220,
+        "length": 75,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 246,
+        "y": 99,
+        "length": 75,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 74,
+        "y": 101,
+        "length": 75,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 243,
+        "y": 225,
+        "length": 75,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 160,
+        "y": 48,
+        "length": 75,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 160,
+        "y": 270,
+        "length": 75,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 4,
+        "x": 85,
+        "y": 274,
+        "locale": "en",
+        "text": "Collect as many stars as you can",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 65,
+        "y": 307,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "ru": [
+      {
+        "name": 13,
+        "x": 60,
+        "y": 315,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 73,
+        "y": 278,
+        "locale": "ru",
+        "text": "Старайтесь собирать звезды леденцом",
+        "width": 210
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 97,
+        "y": 280,
+        "locale": "fr",
+        "text": "Recueillez le plus d'étoiles possible",
+        "width": 150
+      },
+      {
+        "name": 13,
+        "x": 83,
+        "y": 316,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 13,
+        "x": 66,
+        "y": 311,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 86,
+        "y": 275,
+        "locale": "de",
+        "text": "Sammle möglichst viele Sternchen",
+        "width": 170
+      }
+    ],
+    "levelId": "00-02",
+    "levelEditedAt": "2026-05-05T04:09:02.662Z"
+  },
+  "reference_solution": "cut_rope 1\ncut_rope 2\ncut_rope 4\ncut_rope 3 \ncut_rope 0 when candy_near 780,780,50\ncut_rope 5 when candy_near 1100,800,50",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T04:09:11.578Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1\ncut_rope 2\ncut_rope 4\ncut_rope 3 \ncut_rope 0 when candy_near 780,780,50\ncut_rope 5 when candy_near 1100,800,50",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1\ncut_rope 2\ncut_rope 4\ncut_rope 3 \ncut_rope 0 when candy_near 780,780,50\ncut_rope 5 when candy_near 1100,800,50",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-135\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 6\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-136.json

@@ -0,0 +1,154 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-136",
+  "level_file": "data/task/rope-136.json",
+  "video": {
+    "path": "data/video/rope-136.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bouncer": 1,
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "pump": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 100,
+        "x": 214,
+        "y": 25,
+        "length": 90,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 382,
+        "y": 26,
+        "length": 90,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 52,
+        "x": 302,
+        "y": 116
+      },
+      {
+        "name": 82,
+        "x": 299,
+        "y": 223,
+        "angle": 210,
+        "size": 2
+      },
+      {
+        "name": 56,
+        "x": 199,
+        "y": 141,
+        "group": 0,
+        "angle": 0
+      },
+      {
+        "name": 56,
+        "x": 373,
+        "y": 361,
+        "group": 0,
+        "angle": 180
+      },
+      {
+        "name": 2,
+        "x": 414,
+        "y": 178
+      },
+      {
+        "name": 3,
+        "x": 297,
+        "y": 362,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 321,
+        "y": 185,
+        "timeout": -1
+      },
+      {
+        "name": 55,
+        "x": 207,
+        "y": 354,
+        "angle": 0
+      },
+      {
+        "name": 54,
+        "x": 295,
+        "y": 417
+      },
+      {
+        "name": 3,
+        "x": 300,
+        "y": 71,
+        "timeout": -1
+      }
+    ],
+    "levelId": "06-16",
+    "levelEditedAt": "2026-05-05T17:30:39.650Z",
+    "mirroredFrom": "06-16.json"
+  },
+  "reference_solution": "cut_rope 1 \nactivate_pump 0 times 2 when candy_near 820,800,50\npop_bubble 0 when candy_y < 130",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:30:47.382Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 \nactivate_pump 0 times 2 when candy_near 820,800,50\npop_bubble 0 when candy_y < 130",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 \nactivate_pump 0 times 2 when candy_near 820,800,50\npop_bubble 0 when candy_y < 130",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-136\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 1\n  - pump: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-137.json

@@ -0,0 +1,197 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-137",
+  "level_file": "data/task/rope-137.json",
+  "video": {
+    "path": "data/video/rope-137.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bouncer": 7,
+      "candy": 1,
+      "grab_or_rope_anchor": 4,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 81,
+        "x": 596,
+        "y": 315,
+        "angle": -5,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 514,
+        "y": 318,
+        "angle": -1,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 432,
+        "y": 319,
+        "angle": -1,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 352,
+        "y": 320,
+        "angle": -1,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 277,
+        "y": 321,
+        "angle": -1,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 199,
+        "y": 322,
+        "angle": -1,
+        "size": 1
+      },
+      {
+        "name": 2,
+        "x": 57,
+        "y": 289
+      },
+      {
+        "name": 52,
+        "x": 570,
+        "y": 200
+      },
+      {
+        "name": 100,
+        "x": 574,
+        "y": 77,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 3,
+        "x": 196,
+        "y": 261,
+        "timeout": 9
+      },
+      {
+        "name": 3,
+        "x": 353,
+        "y": 255,
+        "timeout": 7
+      },
+      {
+        "name": 3,
+        "x": 500,
+        "y": 257,
+        "timeout": -1
+      },
+      {
+        "name": 81,
+        "x": 125,
+        "y": 323,
+        "angle": -1,
+        "size": 1
+      },
+      {
+        "name": 100,
+        "x": 473,
+        "y": 344,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 316,
+        "y": 342,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 162,
+        "y": 342,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 1540,480,60\ncut_rope 1 when candy_x < 1260\ncut_rope 2 when candy_x < 890",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:29:01.485Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 1540,480,60\ncut_rope 1 when candy_x < 1260\ncut_rope 2 when candy_x < 890",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 1540,480,60\ncut_rope 1 when candy_x < 1260\ncut_rope 2 when candy_x < 890",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-137\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 4\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-138.json

@@ -0,0 +1,204 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-138",
+  "level_file": "data/task/rope-138.json",
+  "video": {
+    "path": "data/video/rope-138.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 1,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "object_13": 1,
+      "object_4": 1,
+      "pump": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 224,
+        "y": 199
+      },
+      {
+        "name": 100,
+        "x": 223,
+        "y": 113,
+        "length": 40,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 2,
+        "x": 518,
+        "y": 432
+      },
+      {
+        "name": 3,
+        "x": 483,
+        "y": 332,
+        "timeout": 17
+      },
+      {
+        "name": 3,
+        "x": 222,
+        "y": 64,
+        "timeout": 4
+      },
+      {
+        "name": 3,
+        "x": 388,
+        "y": 165,
+        "timeout": 9
+      },
+      {
+        "name": 54,
+        "x": 222,
+        "y": 348
+      },
+      {
+        "name": 55,
+        "x": 225,
+        "y": 442,
+        "angle": -90
+      },
+      {
+        "name": 100,
+        "x": 304,
+        "y": 66,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 70,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 4,
+        "x": 176,
+        "y": 215,
+        "locale": "en",
+        "text": "Some stars require fast actions to be collected",
+        "width": 250
+      },
+      {
+        "name": 13,
+        "x": 167,
+        "y": 252,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "ru": [
+      {
+        "name": 13,
+        "x": 65,
+        "y": 259,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 80,
+        "y": 213,
+        "locale": "ru",
+        "text": "Некоторые звезды пропадают если вы не достаточно быстры",
+        "width": 200
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 46,
+        "y": 213,
+        "locale": "fr",
+        "text": "Il est nécessaire d'agir rapidement pour collecter certaines étoiles",
+        "width": 240
+      },
+      {
+        "name": 13,
+        "x": 40,
+        "y": 256,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 60,
+        "y": 207,
+        "locale": "de",
+        "text": "Bei einigen Sternchen musst Du beim Einsammeln schnell sein",
+        "width": 250
+      },
+      {
+        "name": 13,
+        "x": 43,
+        "y": 251,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "levelId": "02-17",
+    "levelEditedAt": "2026-05-05T14:18:10.989Z"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 740,420,60\npop_bubble 0 when candy_y < 170\ncut_rope 1 when candy_x > 1160",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:18:22.894Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 740,420,60\npop_bubble 0 when candy_y < 170\ncut_rope 1 when candy_x > 1160",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 740,420,60\npop_bubble 0 when candy_y < 170\ncut_rope 1 when candy_x > 1160",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-138\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 1\n  - pump: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n- Pumps: activating a pump pushes nearby objects. `activate_pump N`, optionally with `times`/`every`/`until` modifiers for repeated activation.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-139.json

@@ -0,0 +1,188 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-139",
+  "level_file": "data/task/rope-139.json",
+  "video": {
+    "path": "data/video/rope-139.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "object_13": 1,
+      "object_4": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 390,
+        "y": 125
+      },
+      {
+        "name": 2,
+        "x": 322,
+        "y": 431
+      },
+      {
+        "name": 100,
+        "x": 391,
+        "y": 46,
+        "length": 40,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 323,
+        "y": 225,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": 80,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 4,
+        "x": 429,
+        "y": 300,
+        "locale": "en",
+        "text": "Cut the rope before the spider reaches the candy",
+        "width": 240
+      },
+      {
+        "name": 13,
+        "x": 446,
+        "y": 335,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 3,
+        "x": 372,
+        "y": 223,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 230,
+        "y": 287,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 323,
+        "y": 377,
+        "timeout": -1
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 48,
+        "y": 271,
+        "locale": "ru",
+        "text": "Перережьте веревку прежде чем паучок доберется до леденца",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 35,
+        "y": 317,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 56,
+        "y": 299,
+        "locale": "fr",
+        "text": "Coupe la corde avant que l'araignée n'atteigne le bonbon",
+        "width": 240
+      },
+      {
+        "name": 13,
+        "x": 39,
+        "y": 334,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 79,
+        "y": 287,
+        "locale": "de",
+        "text": "Schneide das Seil durch, bevor die Spinne am Bonbon ist",
+        "width": 170
+      },
+      {
+        "name": 13,
+        "x": 150,
+        "y": 276,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "mirroredFrom": "02-09.json"
+  },
+  "reference_solution": "cut_rope 0 \ncut_rope 1 when candy_near 850,780,60 times 2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T13:52:31.316Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 \ncut_rope 1 when candy_near 850,780,60 times 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \ncut_rope 1 when candy_near 850,780,60 times 2",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-139\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-140.json

@@ -0,0 +1,119 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-140",
+  "level_file": "data/task/rope-140.json",
+  "video": {
+    "path": "data/video/rope-140.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "gravity_button": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "width": 320,
+        "height": 480,
+        "gridSize": 32
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 138,
+        "y": 108
+      },
+      {
+        "name": 52,
+        "x": 136,
+        "y": 290
+      },
+      {
+        "name": 3,
+        "x": 129,
+        "y": 386,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 130,
+        "y": 201,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 138,
+        "y": 143,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 183,
+        "y": 205,
+        "length": 115,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 89,
+        "y": 212,
+        "length": 104,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 53,
+        "x": 268,
+        "y": 268
+      }
+    ],
+    "generatorTemplate": "gravity_lift_2",
+    "mirroredFrom": "18-01.json"
+  },
+  "reference_solution": "cut_rope 0,1\ntoggle_gravity when candy_y > 850",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:59:32.809Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1\ntoggle_gravity when candy_y > 850",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1\ntoggle_gravity when candy_y > 850",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-140\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - gravity switch: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Gravity switch: `toggle_gravity` reverses or rotates gravity. Often triggered after the candy crosses a coordinate threshold.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-141.json

@@ -0,0 +1,141 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-141",
+  "level_file": "data/task/rope-141.json",
+  "video": {
+    "path": "data/video/rope-141.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": true,
+    "special": 1,
+    "object_counts": {
+      "bubble": 1,
+      "grab_or_rope_anchor": 3,
+      "left_candy": 1,
+      "right_candy": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 51,
+        "x": 209,
+        "y": 198
+      },
+      {
+        "name": 100,
+        "x": 99,
+        "y": 198,
+        "length": 30,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 304,
+        "y": 197,
+        "length": 30,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 49,
+        "y": 548
+      },
+      {
+        "name": 50,
+        "x": 54,
+        "y": 547
+      },
+      {
+        "name": 2,
+        "x": 104,
+        "y": 42
+      },
+      {
+        "name": 3,
+        "x": 103,
+        "y": 138,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 79,
+        "y": 162,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 53,
+        "y": 188,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 101,
+        "y": 365,
+        "length": 60,
+        "wheel": false,
+        "gun": false,
+        "radius": 50,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 1 when candy_still for 0.3 and candy_near 0,0,60\ncut_rope 2 when candy_still for 0.3 and candy_near 820,590,60\ncut_rope 0 when candy_still for 0.3 and candy_near 820,300,60",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:55:39.203Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 when candy_still for 0.3 and candy_near 0,0,60\ncut_rope 2 when candy_still for 0.3 and candy_near 820,590,60\ncut_rope 0 when candy_still for 0.3 and candy_near 820,300,60",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 when candy_still for 0.3 and candy_near 0,0,60\ncut_rope 2 when candy_still for 0.3 and candy_near 820,590,60\ncut_rope 0 when candy_still for 0.3 and candy_near 820,300,60",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-141\n- canvas_size: 1920 x 1080\n- two_parts (split candy): True\n- object_counts:\n  - target: 1\n  - star: 3\n  - left split candy: 1\n  - right split candy: 1\n  - rope/grab: 3\n  - bubble: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Split candy: the level starts with left_candy and right_candy halves. Before they merge, use left_candy_* and right_candy_* conditions. After they touch and merge, use candy_* for the complete candy.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-142.json

@@ -0,0 +1,138 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-142",
+  "level_file": "data/task/rope-142.json",
+  "video": {
+    "path": "data/video/rope-142.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 4,
+      "candy": 1,
+      "grab_or_rope_anchor": 1,
+      "sock": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 56,
+        "x": 306,
+        "y": 417,
+        "group": 1,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 469,
+        "y": 189,
+        "group": 1,
+        "angle": 190
+      },
+      {
+        "name": 2,
+        "x": 216,
+        "y": 179
+      },
+      {
+        "name": 3,
+        "x": 382,
+        "y": 179,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 302,
+        "y": 336,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 425,
+        "y": 180,
+        "timeout": -1
+      },
+      {
+        "name": 52,
+        "x": 295,
+        "y": 90
+      },
+      {
+        "name": 100,
+        "x": 299,
+        "y": 213,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 54,
+        "x": 297,
+        "y": 88
+      },
+      {
+        "name": 54,
+        "x": 302,
+        "y": 335
+      },
+      {
+        "name": 54,
+        "x": 299,
+        "y": 254
+      },
+      {
+        "name": 54,
+        "x": 300,
+        "y": 164
+      }
+    ]
+  },
+  "reference_solution": "pop_bubble 0 when candy_still for 0.3 and candy_near 910,150,60\npop_bubble 3 when candy_y > 310\ncut_rope 0 when candy_y > 370\npop_bubble 2 when candy_y > 570\npop_bubble 1 when candy_y < 180",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:53:54.627Z"
+  },
+  "answer": {
+    "commands": "pop_bubble 0 when candy_still for 0.3 and candy_near 910,150,60\npop_bubble 3 when candy_y > 310\ncut_rope 0 when candy_y > 370\npop_bubble 2 when candy_y > 570\npop_bubble 1 when candy_y < 180",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "pop_bubble 0 when candy_still for 0.3 and candy_near 910,150,60\npop_bubble 3 when candy_y > 310\ncut_rope 0 when candy_y > 370\npop_bubble 2 when candy_y > 570\npop_bubble 1 when candy_y < 180",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-142\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 1\n  - bubble: 4\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-143.json

@@ -0,0 +1,137 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-143",
+  "level_file": "data/task/rope-143.json",
+  "video": {
+    "path": "data/video/rope-143.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 640,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "bubble": 5,
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 640,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "special": 1,
+        "ropePhysicsSpeed": 1
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 439,
+        "y": 373
+      },
+      {
+        "name": 52,
+        "x": 108,
+        "y": 150
+      },
+      {
+        "name": 100,
+        "x": 371,
+        "y": 94,
+        "length": 260,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 107,
+        "y": 71,
+        "length": 50,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 54,
+        "x": 370,
+        "y": 390
+      },
+      {
+        "name": 54,
+        "x": 210,
+        "y": 343
+      },
+      {
+        "name": 54,
+        "x": 109,
+        "y": 209
+      },
+      {
+        "name": 3,
+        "x": 108,
+        "y": 210,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 209,
+        "y": 343,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 371,
+        "y": 390,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 149,
+        "y": 284
+      },
+      {
+        "name": 54,
+        "x": 288,
+        "y": 381
+      }
+    ],
+    "mirroredFrom": "07-16.json"
+  },
+  "reference_solution": "cut_rope 1 when candy_x < 500\npop_bubble 2 when candy_x < 480\npop_bubble 3 when candy_y > 610\npop_bubble 1 when candy_y > 740\npop_bubble 4 when candy_y > 830\npop_bubble 0 when candy_x > 1060",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:50:48.146Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 when candy_x < 500\npop_bubble 2 when candy_x < 480\npop_bubble 3 when candy_y > 610\npop_bubble 1 when candy_y > 740\npop_bubble 4 when candy_y > 830\npop_bubble 0 when candy_x > 1060",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 when candy_x < 500\npop_bubble 2 when candy_x < 480\npop_bubble 3 when candy_y > 610\npop_bubble 1 when candy_y > 740\npop_bubble 4 when candy_y > 830\npop_bubble 0 when candy_x > 1060",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-143\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - bubble: 5\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Bubbles: candy entering a bubble usually floats upward. `pop_bubble N` pops bubble N. In split-candy levels, `pop_bubble_left` / `pop_bubble_right` pop the bubble holding the corresponding half.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-144.json

@@ -0,0 +1,120 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-144",
+  "level_file": "data/task/rope-144.json",
+  "video": {
+    "path": "data/video/rope-144.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 2,
+      "gravity_button": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "uk": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 153,
+        "y": 341
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 152,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 157,
+        "y": 195,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 97,
+        "timeout": -1
+      },
+      {
+        "name": 53,
+        "x": 52,
+        "y": 413
+      },
+      {
+        "name": 100,
+        "x": 22,
+        "y": 258,
+        "length": 75,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 286,
+        "y": 255,
+        "length": 75,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 52,
+        "x": 159,
+        "y": 256
+      }
+    ],
+    "levelId": "08-21",
+    "levelEditedAt": "2026-05-05T18:16:58.220Z",
+    "mirroredFrom": "08-21.json"
+  },
+  "reference_solution": "cut_rope 0,1\ntoggle_gravity 0\ntoggle_gravity when candy_y < 300",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:17:50.592Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1\ntoggle_gravity 0\ntoggle_gravity when candy_y < 300",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1\ntoggle_gravity 0\ntoggle_gravity when candy_y < 300",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-144\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 2\n  - gravity switch: 1\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n- Gravity switch: `toggle_gravity` reverses or rotates gravity. Often triggered after the candy crosses a coordinate threshold.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-145.json

@@ -0,0 +1,218 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-145",
+  "level_file": "data/task/rope-145.json",
+  "video": {
+    "path": "data/video/rope-145.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 4,
+      "object_13": 1,
+      "object_4": 1,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "special": 1,
+        "twoParts": false
+      }
+    ],
+    "objects": [
+      {
+        "name": 52,
+        "x": 243,
+        "y": 313
+      },
+      {
+        "name": 2,
+        "x": 230,
+        "y": 436
+      },
+      {
+        "name": 3,
+        "x": 172,
+        "y": 139,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 63,
+        "y": 351,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 198,
+        "y": 346,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 196,
+        "y": 191,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 73,
+        "y": 92,
+        "length": 150,
+        "wheel": false,
+        "gun": false,
+        "radius": 55,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 132,
+        "y": 289,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 65,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 271,
+        "y": 381,
+        "length": 25,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 4,
+        "x": 300,
+        "y": 98,
+        "locale": "en",
+        "text": "Stretched ropes turn red",
+        "width": 100
+      },
+      {
+        "name": 13,
+        "x": 253,
+        "y": 85,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "ru": [
+      {
+        "name": 13,
+        "x": 68,
+        "y": 64,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 21,
+        "y": 77,
+        "locale": "ru",
+        "text": "Натянутые веревки становятся красными",
+        "width": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 13,
+        "x": 67,
+        "y": 61,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 20,
+        "y": 74,
+        "locale": "fr",
+        "text": "Les cordes étirées deviennent rouges",
+        "width": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 57,
+        "y": 33,
+        "locale": "de",
+        "text": "Gespannte Seile werden rot",
+        "width": 140
+      },
+      {
+        "name": 13,
+        "x": 39,
+        "y": 68,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "levelId": "01-12",
+    "levelEditedAt": "2026-05-05T13:10:17.063Z",
+    "mirroredFrom": "01-12.json"
+  },
+  "reference_solution": "cut_rope 3 \ncut_rope 0,1 when candy_still for 0.3 \ncut_rope 2 when candy_near 1070,800,60",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T13:10:29.123Z"
+  },
+  "answer": {
+    "commands": "cut_rope 3 \ncut_rope 0,1 when candy_still for 0.3 \ncut_rope 2 when candy_near 1070,800,60",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 3 \ncut_rope 0,1 when candy_still for 0.3 \ncut_rope 2 when candy_near 1070,800,60",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-145\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 4\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}

cutrope/data/rope-146.json

@@ -0,0 +1,161 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-146",
+  "level_file": "data/task/rope-146.json",
+  "video": {
+    "path": "data/video/rope-146.mp4",
+    "mime_type": "video/mp4",
+    "duration_seconds": 3,
+    "fps": 30
+  },
+  "prompt_level": {
+    "canvas_width": 1920,
+    "canvas_height": 1080,
+    "map_width": 320,
+    "map_height": 480,
+    "two_parts": false,
+    "special": 1,
+    "object_counts": {
+      "candy": 1,
+      "grab_or_rope_anchor": 6,
+      "star": 3,
+      "target": 1
+    }
+  },
+  "level_json_without_solution": {
+    "settings": [
+      {
+        "name": 0,
+        "gridSize": 32,
+        "width": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "special": 1,
+        "ropePhysicsSpeed": 1
+      }
+    ],
+    "objects": [
+      {
+        "name": 2,
+        "x": 22,
+        "y": 409
+      },
+      {
+        "name": 52,
+        "x": 160,
+        "y": 149
+      },
+      {
+        "name": 3,
+        "x": 62,
+        "y": 356,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 254,
+        "y": 358,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 53,
+        "y": 196,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 156,
+        "y": 42,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 54,
+        "y": 84,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 252,
+        "y": 85,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 252,
+        "y": 232,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 156,
+        "y": 287,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 51,
+        "y": 229,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      }
+    ],
+    "levelId": "01-07",
+    "levelEditedAt": "2026-05-05T11:29:53.852Z",
+    "mirroredFrom": "01-07.json"
+  },
+  "reference_solution": "cut_rope 0,2,3 \ncut_rope 1 when candy_near 760,520,40\ncut_rope 5 when candy_y > 820\ncut_rope 4 when candy_near 820,900,40 times 2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T11:30:02.639Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,2,3 \ncut_rope 1 when candy_near 760,520,40\ncut_rope 5 when candy_y > 820\ncut_rope 4 when candy_near 820,900,40 times 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,2,3 \ncut_rope 1 when candy_near 760,520,40\ncut_rope 5 when candy_y > 820\ncut_rope 4 when candy_near 820,900,40 times 2",
+  "prompt": {
+    "system": "You are a Cut the Rope replayable-script solver. Given a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and a structured description of the level objects, you produce a deterministic command script that, replayed from the same initial state, wins the level.\n\nYou will receive: (a) one short gameplay video of the level, (b) a structured text body listing the canvas size, the count of each gameplay object (candy, target, ropes, bubbles, pumps, gravity switches, stars), and the level metadata.\n\nYou may reason step by step before the final answer. Place your final answer on the very last line of your reply, in the form: FINAL_JSON: <one-line JSON>\n\nThe JSON object must follow the commands schema described in section 8 of the user prompt: keys `commands`, `reason`, `confidence`. Multiple winning scripts may be valid; output any one. Do NOT wrap FINAL_JSON in Markdown code fences. Do NOT write anything after the FINAL_JSON line. Emit exactly one FINAL_JSON line.\n\nIf you cannot reason out a confident solution, still emit a best-effort FINAL_JSON line with at least one command and a confidence value below 1.0. Never refuse, never return prose only.",
+    "user": "## 1. TASK\nProduce a replayable command script that wins the Cut the Rope level shown in the gameplay video. The script must be deterministic so that replaying it from the same initial state in the simulator produces a win.\nThe level is solved when the candy reaches the target monster's mouth. Three-star completion is preferred but stable winning is the first priority.\n\n## 2. WORLD MODEL\n- Candy: a physics object affected by gravity, ropes, bubbles, pumps, and other active objects. Must be delivered to the target.\n- Target / monster: the level is won when the candy enters the target's mouth.\n- Rope: a constraint between the candy (or a split-candy half) and an anchor. Cutting a rope releases the candy from that constraint. Ropes are zero-indexed.\n- Grab: an anchor for a rope that can be repositioned during play. Zero-indexed.\n- Bubble: when the candy enters a bubble, it usually floats upward. Pop the bubble to release the candy.\n- Pump: a gust source that pushes objects. Each activation produces one impulse.\n- Gravity switch: toggles gravity direction or magnitude.\n- Star: collected by candy passing through its position. Up to 3 stars per level.\n- Split candy: some levels start with two candy halves (`left_candy`, `right_candy`). Use the half-specific conditions before the halves merge; switch to `candy_*` conditions after they merge.\n\n## 3. VISUAL LEGEND\n- The video is a short clip showing the full level layout and a coordinate grid overlay.\n- Coordinates: origin at the top-left corner. x increases to the right; y increases downward.\n- Indices in the video are zero-based for ropes, bubbles, pumps, and grabs. Infer indices from object positions and the order they appear in the clip.\n\n## 4. INPUT FIELDS\n- level_id: rope-146\n- canvas_size: 1920 x 1080\n- two_parts (split candy): False\n- object_counts:\n  - target: 1\n  - star: 3\n  - candy: 1\n  - rope/grab: 6\n- gameplay objects in this level:\n- Target (monster): the level is won when the candy reaches the monster's mouth. The target is usually fixed in place.\n- Candy: the candy is affected by gravity, ropes, bubbles, pumps, and other active objects. It must be delivered to the target.\n- Ropes / grabs: ropes constrain the candy or split-candy movement. `cut_rope N` cuts rope N; `cut_rope N,M,K` cuts multiple at once. If a grab can move, `move_grab N X` or `move_grab N X Y` repositions it.\n- Stars: candy or split-candy collects a star by passing through it. Prefer paths that collect 3 stars, but stable completion takes priority.\n\n## 5. ACTION VOCABULARY\nA complete answer is one ordered command script (one command per line). Available actions:\n- `cut_rope N`               — cut rope N\n- `cut_rope N,M,K`           — cut several ropes simultaneously\n- `pop_bubble N`             — pop bubble N\n- `pop_bubble_left`          — pop the bubble holding the left split-candy half\n- `pop_bubble_right`         — pop the bubble holding the right split-candy half\n- `activate_pump N`          — fire pump N once\n- `activate_pump N times C`  — fire C times back-to-back\n- `activate_pump N times C every S` — fire C times spaced S seconds apart\n- `activate_pump N every S until <CONDITION>` — fire repeatedly until CONDITION holds\n- `move_grab N X`            — move grab N to x-coordinate X (y unchanged)\n- `move_grab N X Y`          — move grab N to (X, Y)\n- `kick_rope N`              — apply a one-shot impulse to rope N\n- `toggle_gravity`           — toggle the gravity setting\n\nEach command may be guarded by an optional `when <CONDITION>` clause. Available condition primitives:\n- `candy_x > N`, `candy_x < N`, `candy_y > N`, `candy_y < N`\n- `candy_near X,Y,R` — candy is within radius R of (X,Y); optional `for S` requires the predicate to hold for S seconds\n- `candy_still for S` — candy speed is below threshold for S seconds (useful before cutting after a swing settles)\n- `left_candy_*` and `right_candy_*` mirror the candy_* primitives for the corresponding split half\n- `grab_x I > N`, `grab_y I < N`, `grab_near I,X,Y,R`\n- `rope_cut N` — rope N has been cut\n- `no_rope`, `candy_in_bubble`\n- Conditions can be combined with `and`, `or`, and parentheses.\n\n## 6. CONSTRAINTS\n- One command per line in the `commands` string.\n- Indices must be zero-based and consistent with the level video.\n- `wait_frames` is NOT allowed in this benchmark version. Use condition-based waits.\n- The script must be deterministic; do not output stochastic or \"either-or\" entries.\n- The script must terminate without manual intervention; the simulator runs it from the initial state and decides win/loss.\n\n## 7. SOLVING ADVICE\n- Identify which ropes mainly control the candy first; choose a cut order that releases the candy along the shortest stable path to the target.\n- When waiting for the candy to reach a place, prefer `candy_near` / `candy_y` / `candy_x` over fixed timing.\n- For repeated pumping, use `activate_pump N times C every S` or `activate_pump N every S until CONDITION`.\n- In split-candy levels, do NOT use `candy_still` / `candy_near` for the merged candy before the halves merge; use the `left_candy_*` / `right_candy_*` primitives instead.\n\n## 8. OUTPUT SCHEMA\nFINAL_JSON: {\"commands\":\"<command-script with \\n separators>\",\"reason\":\"<one-sentence intent>\",\"confidence\":<0..1>}\n- commands: string. One command per line; lines separated by `\\n`.\n- reason: short one-sentence explanation of the intended sequence.\n- confidence: float in `[0, 1]`.\n\n## 9. FINAL INSTRUCTION\nYou may include reasoning above, but the very last line of your reply must start with FINAL_JSON: followed by exactly one valid JSON object.\nDo not wrap FINAL_JSON in code fences and do not write anything after it."
+  }
+}