Add files using upload-large-folder tool

cutrope/data/manifest.json

@@ -0,0 +1,1638 @@
+{
+  "schema_version": 1,
+  "data_dir": "eval/data",
+  "levels": [
+    {
+      "level_id": "rope-000",
+      "data_file": "eval/data/rope-000.json",
+      "level_file": "data/task/rope-000.json",
+      "video_file": "data/video/rope-000.mp4"
+    },
+    {
+      "level_id": "rope-001",
+      "data_file": "eval/data/rope-001.json",
+      "level_file": "data/task/rope-001.json",
+      "video_file": "data/video/rope-001.mp4"
+    },
+    {
+      "level_id": "rope-002",
+      "data_file": "eval/data/rope-002.json",
+      "level_file": "data/task/rope-002.json",
+      "video_file": "data/video/rope-002.mp4"
+    },
+    {
+      "level_id": "rope-003",
+      "data_file": "eval/data/rope-003.json",
+      "level_file": "data/task/rope-003.json",
+      "video_file": "data/video/rope-003.mp4"
+    },
+    {
+      "level_id": "rope-004",
+      "data_file": "eval/data/rope-004.json",
+      "level_file": "data/task/rope-004.json",
+      "video_file": "data/video/rope-004.mp4"
+    },
+    {
+      "level_id": "rope-005",
+      "data_file": "eval/data/rope-005.json",
+      "level_file": "data/task/rope-005.json",
+      "video_file": "data/video/rope-005.mp4"
+    },
+    {
+      "level_id": "rope-006",
+      "data_file": "eval/data/rope-006.json",
+      "level_file": "data/task/rope-006.json",
+      "video_file": "data/video/rope-006.mp4"
+    },
+    {
+      "level_id": "rope-007",
+      "data_file": "eval/data/rope-007.json",
+      "level_file": "data/task/rope-007.json",
+      "video_file": "data/video/rope-007.mp4"
+    },
+    {
+      "level_id": "rope-008",
+      "data_file": "eval/data/rope-008.json",
+      "level_file": "data/task/rope-008.json",
+      "video_file": "data/video/rope-008.mp4"
+    },
+    {
+      "level_id": "rope-009",
+      "data_file": "eval/data/rope-009.json",
+      "level_file": "data/task/rope-009.json",
+      "video_file": "data/video/rope-009.mp4"
+    },
+    {
+      "level_id": "rope-010",
+      "data_file": "eval/data/rope-010.json",
+      "level_file": "data/task/rope-010.json",
+      "video_file": "data/video/rope-010.mp4"
+    },
+    {
+      "level_id": "rope-011",
+      "data_file": "eval/data/rope-011.json",
+      "level_file": "data/task/rope-011.json",
+      "video_file": "data/video/rope-011.mp4"
+    },
+    {
+      "level_id": "rope-012",
+      "data_file": "eval/data/rope-012.json",
+      "level_file": "data/task/rope-012.json",
+      "video_file": "data/video/rope-012.mp4"
+    },
+    {
+      "level_id": "rope-013",
+      "data_file": "eval/data/rope-013.json",
+      "level_file": "data/task/rope-013.json",
+      "video_file": "data/video/rope-013.mp4"
+    },
+    {
+      "level_id": "rope-014",
+      "data_file": "eval/data/rope-014.json",
+      "level_file": "data/task/rope-014.json",
+      "video_file": "data/video/rope-014.mp4"
+    },
+    {
+      "level_id": "rope-015",
+      "data_file": "eval/data/rope-015.json",
+      "level_file": "data/task/rope-015.json",
+      "video_file": "data/video/rope-015.mp4"
+    },
+    {
+      "level_id": "rope-016",
+      "data_file": "eval/data/rope-016.json",
+      "level_file": "data/task/rope-016.json",
+      "video_file": "data/video/rope-016.mp4"
+    },
+    {
+      "level_id": "rope-017",
+      "data_file": "eval/data/rope-017.json",
+      "level_file": "data/task/rope-017.json",
+      "video_file": "data/video/rope-017.mp4"
+    },
+    {
+      "level_id": "rope-018",
+      "data_file": "eval/data/rope-018.json",
+      "level_file": "data/task/rope-018.json",
+      "video_file": "data/video/rope-018.mp4"
+    },
+    {
+      "level_id": "rope-019",
+      "data_file": "eval/data/rope-019.json",
+      "level_file": "data/task/rope-019.json",
+      "video_file": "data/video/rope-019.mp4"
+    },
+    {
+      "level_id": "rope-020",
+      "data_file": "eval/data/rope-020.json",
+      "level_file": "data/task/rope-020.json",
+      "video_file": "data/video/rope-020.mp4"
+    },
+    {
+      "level_id": "rope-021",
+      "data_file": "eval/data/rope-021.json",
+      "level_file": "data/task/rope-021.json",
+      "video_file": "data/video/rope-021.mp4"
+    },
+    {
+      "level_id": "rope-022",
+      "data_file": "eval/data/rope-022.json",
+      "level_file": "data/task/rope-022.json",
+      "video_file": "data/video/rope-022.mp4"
+    },
+    {
+      "level_id": "rope-023",
+      "data_file": "eval/data/rope-023.json",
+      "level_file": "data/task/rope-023.json",
+      "video_file": "data/video/rope-023.mp4"
+    },
+    {
+      "level_id": "rope-024",
+      "data_file": "eval/data/rope-024.json",
+      "level_file": "data/task/rope-024.json",
+      "video_file": "data/video/rope-024.mp4"
+    },
+    {
+      "level_id": "rope-025",
+      "data_file": "eval/data/rope-025.json",
+      "level_file": "data/task/rope-025.json",
+      "video_file": "data/video/rope-025.mp4"
+    },
+    {
+      "level_id": "rope-026",
+      "data_file": "eval/data/rope-026.json",
+      "level_file": "data/task/rope-026.json",
+      "video_file": "data/video/rope-026.mp4"
+    },
+    {
+      "level_id": "rope-027",
+      "data_file": "eval/data/rope-027.json",
+      "level_file": "data/task/rope-027.json",
+      "video_file": "data/video/rope-027.mp4"
+    },
+    {
+      "level_id": "rope-028",
+      "data_file": "eval/data/rope-028.json",
+      "level_file": "data/task/rope-028.json",
+      "video_file": "data/video/rope-028.mp4"
+    },
+    {
+      "level_id": "rope-029",
+      "data_file": "eval/data/rope-029.json",
+      "level_file": "data/task/rope-029.json",
+      "video_file": "data/video/rope-029.mp4"
+    },
+    {
+      "level_id": "rope-030",
+      "data_file": "eval/data/rope-030.json",
+      "level_file": "data/task/rope-030.json",
+      "video_file": "data/video/rope-030.mp4"
+    },
+    {
+      "level_id": "rope-031",
+      "data_file": "eval/data/rope-031.json",
+      "level_file": "data/task/rope-031.json",
+      "video_file": "data/video/rope-031.mp4"
+    },
+    {
+      "level_id": "rope-032",
+      "data_file": "eval/data/rope-032.json",
+      "level_file": "data/task/rope-032.json",
+      "video_file": "data/video/rope-032.mp4"
+    },
+    {
+      "level_id": "rope-033",
+      "data_file": "eval/data/rope-033.json",
+      "level_file": "data/task/rope-033.json",
+      "video_file": "data/video/rope-033.mp4"
+    },
+    {
+      "level_id": "rope-034",
+      "data_file": "eval/data/rope-034.json",
+      "level_file": "data/task/rope-034.json",
+      "video_file": "data/video/rope-034.mp4"
+    },
+    {
+      "level_id": "rope-035",
+      "data_file": "eval/data/rope-035.json",
+      "level_file": "data/task/rope-035.json",
+      "video_file": "data/video/rope-035.mp4"
+    },
+    {
+      "level_id": "rope-036",
+      "data_file": "eval/data/rope-036.json",
+      "level_file": "data/task/rope-036.json",
+      "video_file": "data/video/rope-036.mp4"
+    },
+    {
+      "level_id": "rope-037",
+      "data_file": "eval/data/rope-037.json",
+      "level_file": "data/task/rope-037.json",
+      "video_file": "data/video/rope-037.mp4"
+    },
+    {
+      "level_id": "rope-038",
+      "data_file": "eval/data/rope-038.json",
+      "level_file": "data/task/rope-038.json",
+      "video_file": "data/video/rope-038.mp4"
+    },
+    {
+      "level_id": "rope-039",
+      "data_file": "eval/data/rope-039.json",
+      "level_file": "data/task/rope-039.json",
+      "video_file": "data/video/rope-039.mp4"
+    },
+    {
+      "level_id": "rope-040",
+      "data_file": "eval/data/rope-040.json",
+      "level_file": "data/task/rope-040.json",
+      "video_file": "data/video/rope-040.mp4"
+    },
+    {
+      "level_id": "rope-041",
+      "data_file": "eval/data/rope-041.json",
+      "level_file": "data/task/rope-041.json",
+      "video_file": "data/video/rope-041.mp4"
+    },
+    {
+      "level_id": "rope-042",
+      "data_file": "eval/data/rope-042.json",
+      "level_file": "data/task/rope-042.json",
+      "video_file": "data/video/rope-042.mp4"
+    },
+    {
+      "level_id": "rope-043",
+      "data_file": "eval/data/rope-043.json",
+      "level_file": "data/task/rope-043.json",
+      "video_file": "data/video/rope-043.mp4"
+    },
+    {
+      "level_id": "rope-044",
+      "data_file": "eval/data/rope-044.json",
+      "level_file": "data/task/rope-044.json",
+      "video_file": "data/video/rope-044.mp4"
+    },
+    {
+      "level_id": "rope-045",
+      "data_file": "eval/data/rope-045.json",
+      "level_file": "data/task/rope-045.json",
+      "video_file": "data/video/rope-045.mp4"
+    },
+    {
+      "level_id": "rope-046",
+      "data_file": "eval/data/rope-046.json",
+      "level_file": "data/task/rope-046.json",
+      "video_file": "data/video/rope-046.mp4"
+    },
+    {
+      "level_id": "rope-047",
+      "data_file": "eval/data/rope-047.json",
+      "level_file": "data/task/rope-047.json",
+      "video_file": "data/video/rope-047.mp4"
+    },
+    {
+      "level_id": "rope-048",
+      "data_file": "eval/data/rope-048.json",
+      "level_file": "data/task/rope-048.json",
+      "video_file": "data/video/rope-048.mp4"
+    },
+    {
+      "level_id": "rope-049",
+      "data_file": "eval/data/rope-049.json",
+      "level_file": "data/task/rope-049.json",
+      "video_file": "data/video/rope-049.mp4"
+    },
+    {
+      "level_id": "rope-050",
+      "data_file": "eval/data/rope-050.json",
+      "level_file": "data/task/rope-050.json",
+      "video_file": "data/video/rope-050.mp4"
+    },
+    {
+      "level_id": "rope-051",
+      "data_file": "eval/data/rope-051.json",
+      "level_file": "data/task/rope-051.json",
+      "video_file": "data/video/rope-051.mp4"
+    },
+    {
+      "level_id": "rope-052",
+      "data_file": "eval/data/rope-052.json",
+      "level_file": "data/task/rope-052.json",
+      "video_file": "data/video/rope-052.mp4"
+    },
+    {
+      "level_id": "rope-053",
+      "data_file": "eval/data/rope-053.json",
+      "level_file": "data/task/rope-053.json",
+      "video_file": "data/video/rope-053.mp4"
+    },
+    {
+      "level_id": "rope-054",
+      "data_file": "eval/data/rope-054.json",
+      "level_file": "data/task/rope-054.json",
+      "video_file": "data/video/rope-054.mp4"
+    },
+    {
+      "level_id": "rope-055",
+      "data_file": "eval/data/rope-055.json",
+      "level_file": "data/task/rope-055.json",
+      "video_file": "data/video/rope-055.mp4"
+    },
+    {
+      "level_id": "rope-056",
+      "data_file": "eval/data/rope-056.json",
+      "level_file": "data/task/rope-056.json",
+      "video_file": "data/video/rope-056.mp4"
+    },
+    {
+      "level_id": "rope-057",
+      "data_file": "eval/data/rope-057.json",
+      "level_file": "data/task/rope-057.json",
+      "video_file": "data/video/rope-057.mp4"
+    },
+    {
+      "level_id": "rope-058",
+      "data_file": "eval/data/rope-058.json",
+      "level_file": "data/task/rope-058.json",
+      "video_file": "data/video/rope-058.mp4"
+    },
+    {
+      "level_id": "rope-059",
+      "data_file": "eval/data/rope-059.json",
+      "level_file": "data/task/rope-059.json",
+      "video_file": "data/video/rope-059.mp4"
+    },
+    {
+      "level_id": "rope-060",
+      "data_file": "eval/data/rope-060.json",
+      "level_file": "data/task/rope-060.json",
+      "video_file": "data/video/rope-060.mp4"
+    },
+    {
+      "level_id": "rope-061",
+      "data_file": "eval/data/rope-061.json",
+      "level_file": "data/task/rope-061.json",
+      "video_file": "data/video/rope-061.mp4"
+    },
+    {
+      "level_id": "rope-062",
+      "data_file": "eval/data/rope-062.json",
+      "level_file": "data/task/rope-062.json",
+      "video_file": "data/video/rope-062.mp4"
+    },
+    {
+      "level_id": "rope-063",
+      "data_file": "eval/data/rope-063.json",
+      "level_file": "data/task/rope-063.json",
+      "video_file": "data/video/rope-063.mp4"
+    },
+    {
+      "level_id": "rope-064",
+      "data_file": "eval/data/rope-064.json",
+      "level_file": "data/task/rope-064.json",
+      "video_file": "data/video/rope-064.mp4"
+    },
+    {
+      "level_id": "rope-065",
+      "data_file": "eval/data/rope-065.json",
+      "level_file": "data/task/rope-065.json",
+      "video_file": "data/video/rope-065.mp4"
+    },
+    {
+      "level_id": "rope-066",
+      "data_file": "eval/data/rope-066.json",
+      "level_file": "data/task/rope-066.json",
+      "video_file": "data/video/rope-066.mp4"
+    },
+    {
+      "level_id": "rope-067",
+      "data_file": "eval/data/rope-067.json",
+      "level_file": "data/task/rope-067.json",
+      "video_file": "data/video/rope-067.mp4"
+    },
+    {
+      "level_id": "rope-068",
+      "data_file": "eval/data/rope-068.json",
+      "level_file": "data/task/rope-068.json",
+      "video_file": "data/video/rope-068.mp4"
+    },
+    {
+      "level_id": "rope-069",
+      "data_file": "eval/data/rope-069.json",
+      "level_file": "data/task/rope-069.json",
+      "video_file": "data/video/rope-069.mp4"
+    },
+    {
+      "level_id": "rope-070",
+      "data_file": "eval/data/rope-070.json",
+      "level_file": "data/task/rope-070.json",
+      "video_file": "data/video/rope-070.mp4"
+    },
+    {
+      "level_id": "rope-071",
+      "data_file": "eval/data/rope-071.json",
+      "level_file": "data/task/rope-071.json",
+      "video_file": "data/video/rope-071.mp4"
+    },
+    {
+      "level_id": "rope-072",
+      "data_file": "eval/data/rope-072.json",
+      "level_file": "data/task/rope-072.json",
+      "video_file": "data/video/rope-072.mp4"
+    },
+    {
+      "level_id": "rope-073",
+      "data_file": "eval/data/rope-073.json",
+      "level_file": "data/task/rope-073.json",
+      "video_file": "data/video/rope-073.mp4"
+    },
+    {
+      "level_id": "rope-074",
+      "data_file": "eval/data/rope-074.json",
+      "level_file": "data/task/rope-074.json",
+      "video_file": "data/video/rope-074.mp4"
+    },
+    {
+      "level_id": "rope-075",
+      "data_file": "eval/data/rope-075.json",
+      "level_file": "data/task/rope-075.json",
+      "video_file": "data/video/rope-075.mp4"
+    },
+    {
+      "level_id": "rope-076",
+      "data_file": "eval/data/rope-076.json",
+      "level_file": "data/task/rope-076.json",
+      "video_file": "data/video/rope-076.mp4"
+    },
+    {
+      "level_id": "rope-077",
+      "data_file": "eval/data/rope-077.json",
+      "level_file": "data/task/rope-077.json",
+      "video_file": "data/video/rope-077.mp4"
+    },
+    {
+      "level_id": "rope-078",
+      "data_file": "eval/data/rope-078.json",
+      "level_file": "data/task/rope-078.json",
+      "video_file": "data/video/rope-078.mp4"
+    },
+    {
+      "level_id": "rope-079",
+      "data_file": "eval/data/rope-079.json",
+      "level_file": "data/task/rope-079.json",
+      "video_file": "data/video/rope-079.mp4"
+    },
+    {
+      "level_id": "rope-080",
+      "data_file": "eval/data/rope-080.json",
+      "level_file": "data/task/rope-080.json",
+      "video_file": "data/video/rope-080.mp4"
+    },
+    {
+      "level_id": "rope-081",
+      "data_file": "eval/data/rope-081.json",
+      "level_file": "data/task/rope-081.json",
+      "video_file": "data/video/rope-081.mp4"
+    },
+    {
+      "level_id": "rope-082",
+      "data_file": "eval/data/rope-082.json",
+      "level_file": "data/task/rope-082.json",
+      "video_file": "data/video/rope-082.mp4"
+    },
+    {
+      "level_id": "rope-083",
+      "data_file": "eval/data/rope-083.json",
+      "level_file": "data/task/rope-083.json",
+      "video_file": "data/video/rope-083.mp4"
+    },
+    {
+      "level_id": "rope-084",
+      "data_file": "eval/data/rope-084.json",
+      "level_file": "data/task/rope-084.json",
+      "video_file": "data/video/rope-084.mp4"
+    },
+    {
+      "level_id": "rope-085",
+      "data_file": "eval/data/rope-085.json",
+      "level_file": "data/task/rope-085.json",
+      "video_file": "data/video/rope-085.mp4"
+    },
+    {
+      "level_id": "rope-086",
+      "data_file": "eval/data/rope-086.json",
+      "level_file": "data/task/rope-086.json",
+      "video_file": "data/video/rope-086.mp4"
+    },
+    {
+      "level_id": "rope-087",
+      "data_file": "eval/data/rope-087.json",
+      "level_file": "data/task/rope-087.json",
+      "video_file": "data/video/rope-087.mp4"
+    },
+    {
+      "level_id": "rope-088",
+      "data_file": "eval/data/rope-088.json",
+      "level_file": "data/task/rope-088.json",
+      "video_file": "data/video/rope-088.mp4"
+    },
+    {
+      "level_id": "rope-089",
+      "data_file": "eval/data/rope-089.json",
+      "level_file": "data/task/rope-089.json",
+      "video_file": "data/video/rope-089.mp4"
+    },
+    {
+      "level_id": "rope-090",
+      "data_file": "eval/data/rope-090.json",
+      "level_file": "data/task/rope-090.json",
+      "video_file": "data/video/rope-090.mp4"
+    },
+    {
+      "level_id": "rope-091",
+      "data_file": "eval/data/rope-091.json",
+      "level_file": "data/task/rope-091.json",
+      "video_file": "data/video/rope-091.mp4"
+    },
+    {
+      "level_id": "rope-092",
+      "data_file": "eval/data/rope-092.json",
+      "level_file": "data/task/rope-092.json",
+      "video_file": "data/video/rope-092.mp4"
+    },
+    {
+      "level_id": "rope-093",
+      "data_file": "eval/data/rope-093.json",
+      "level_file": "data/task/rope-093.json",
+      "video_file": "data/video/rope-093.mp4"
+    },
+    {
+      "level_id": "rope-094",
+      "data_file": "eval/data/rope-094.json",
+      "level_file": "data/task/rope-094.json",
+      "video_file": "data/video/rope-094.mp4"
+    },
+    {
+      "level_id": "rope-095",
+      "data_file": "eval/data/rope-095.json",
+      "level_file": "data/task/rope-095.json",
+      "video_file": "data/video/rope-095.mp4"
+    },
+    {
+      "level_id": "rope-096",
+      "data_file": "eval/data/rope-096.json",
+      "level_file": "data/task/rope-096.json",
+      "video_file": "data/video/rope-096.mp4"
+    },
+    {
+      "level_id": "rope-097",
+      "data_file": "eval/data/rope-097.json",
+      "level_file": "data/task/rope-097.json",
+      "video_file": "data/video/rope-097.mp4"
+    },
+    {
+      "level_id": "rope-098",
+      "data_file": "eval/data/rope-098.json",
+      "level_file": "data/task/rope-098.json",
+      "video_file": "data/video/rope-098.mp4"
+    },
+    {
+      "level_id": "rope-099",
+      "data_file": "eval/data/rope-099.json",
+      "level_file": "data/task/rope-099.json",
+      "video_file": "data/video/rope-099.mp4"
+    },
+    {
+      "level_id": "rope-100",
+      "data_file": "eval/data/rope-100.json",
+      "level_file": "data/task/rope-100.json",
+      "video_file": "data/video/rope-100.mp4"
+    },
+    {
+      "level_id": "rope-101",
+      "data_file": "eval/data/rope-101.json",
+      "level_file": "data/task/rope-101.json",
+      "video_file": "data/video/rope-101.mp4"
+    },
+    {
+      "level_id": "rope-102",
+      "data_file": "eval/data/rope-102.json",
+      "level_file": "data/task/rope-102.json",
+      "video_file": "data/video/rope-102.mp4"
+    },
+    {
+      "level_id": "rope-103",
+      "data_file": "eval/data/rope-103.json",
+      "level_file": "data/task/rope-103.json",
+      "video_file": "data/video/rope-103.mp4"
+    },
+    {
+      "level_id": "rope-104",
+      "data_file": "eval/data/rope-104.json",
+      "level_file": "data/task/rope-104.json",
+      "video_file": "data/video/rope-104.mp4"
+    },
+    {
+      "level_id": "rope-105",
+      "data_file": "eval/data/rope-105.json",
+      "level_file": "data/task/rope-105.json",
+      "video_file": "data/video/rope-105.mp4"
+    },
+    {
+      "level_id": "rope-106",
+      "data_file": "eval/data/rope-106.json",
+      "level_file": "data/task/rope-106.json",
+      "video_file": "data/video/rope-106.mp4"
+    },
+    {
+      "level_id": "rope-107",
+      "data_file": "eval/data/rope-107.json",
+      "level_file": "data/task/rope-107.json",
+      "video_file": "data/video/rope-107.mp4"
+    },
+    {
+      "level_id": "rope-108",
+      "data_file": "eval/data/rope-108.json",
+      "level_file": "data/task/rope-108.json",
+      "video_file": "data/video/rope-108.mp4"
+    },
+    {
+      "level_id": "rope-109",
+      "data_file": "eval/data/rope-109.json",
+      "level_file": "data/task/rope-109.json",
+      "video_file": "data/video/rope-109.mp4"
+    },
+    {
+      "level_id": "rope-110",
+      "data_file": "eval/data/rope-110.json",
+      "level_file": "data/task/rope-110.json",
+      "video_file": "data/video/rope-110.mp4"
+    },
+    {
+      "level_id": "rope-111",
+      "data_file": "eval/data/rope-111.json",
+      "level_file": "data/task/rope-111.json",
+      "video_file": "data/video/rope-111.mp4"
+    },
+    {
+      "level_id": "rope-112",
+      "data_file": "eval/data/rope-112.json",
+      "level_file": "data/task/rope-112.json",
+      "video_file": "data/video/rope-112.mp4"
+    },
+    {
+      "level_id": "rope-113",
+      "data_file": "eval/data/rope-113.json",
+      "level_file": "data/task/rope-113.json",
+      "video_file": "data/video/rope-113.mp4"
+    },
+    {
+      "level_id": "rope-114",
+      "data_file": "eval/data/rope-114.json",
+      "level_file": "data/task/rope-114.json",
+      "video_file": "data/video/rope-114.mp4"
+    },
+    {
+      "level_id": "rope-115",
+      "data_file": "eval/data/rope-115.json",
+      "level_file": "data/task/rope-115.json",
+      "video_file": "data/video/rope-115.mp4"
+    },
+    {
+      "level_id": "rope-116",
+      "data_file": "eval/data/rope-116.json",
+      "level_file": "data/task/rope-116.json",
+      "video_file": "data/video/rope-116.mp4"
+    },
+    {
+      "level_id": "rope-117",
+      "data_file": "eval/data/rope-117.json",
+      "level_file": "data/task/rope-117.json",
+      "video_file": "data/video/rope-117.mp4"
+    },
+    {
+      "level_id": "rope-118",
+      "data_file": "eval/data/rope-118.json",
+      "level_file": "data/task/rope-118.json",
+      "video_file": "data/video/rope-118.mp4"
+    },
+    {
+      "level_id": "rope-119",
+      "data_file": "eval/data/rope-119.json",
+      "level_file": "data/task/rope-119.json",
+      "video_file": "data/video/rope-119.mp4"
+    },
+    {
+      "level_id": "rope-120",
+      "data_file": "eval/data/rope-120.json",
+      "level_file": "data/task/rope-120.json",
+      "video_file": "data/video/rope-120.mp4"
+    },
+    {
+      "level_id": "rope-121",
+      "data_file": "eval/data/rope-121.json",
+      "level_file": "data/task/rope-121.json",
+      "video_file": "data/video/rope-121.mp4"
+    },
+    {
+      "level_id": "rope-122",
+      "data_file": "eval/data/rope-122.json",
+      "level_file": "data/task/rope-122.json",
+      "video_file": "data/video/rope-122.mp4"
+    },
+    {
+      "level_id": "rope-123",
+      "data_file": "eval/data/rope-123.json",
+      "level_file": "data/task/rope-123.json",
+      "video_file": "data/video/rope-123.mp4"
+    },
+    {
+      "level_id": "rope-124",
+      "data_file": "eval/data/rope-124.json",
+      "level_file": "data/task/rope-124.json",
+      "video_file": "data/video/rope-124.mp4"
+    },
+    {
+      "level_id": "rope-125",
+      "data_file": "eval/data/rope-125.json",
+      "level_file": "data/task/rope-125.json",
+      "video_file": "data/video/rope-125.mp4"
+    },
+    {
+      "level_id": "rope-126",
+      "data_file": "eval/data/rope-126.json",
+      "level_file": "data/task/rope-126.json",
+      "video_file": "data/video/rope-126.mp4"
+    },
+    {
+      "level_id": "rope-127",
+      "data_file": "eval/data/rope-127.json",
+      "level_file": "data/task/rope-127.json",
+      "video_file": "data/video/rope-127.mp4"
+    },
+    {
+      "level_id": "rope-128",
+      "data_file": "eval/data/rope-128.json",
+      "level_file": "data/task/rope-128.json",
+      "video_file": "data/video/rope-128.mp4"
+    },
+    {
+      "level_id": "rope-129",
+      "data_file": "eval/data/rope-129.json",
+      "level_file": "data/task/rope-129.json",
+      "video_file": "data/video/rope-129.mp4"
+    },
+    {
+      "level_id": "rope-130",
+      "data_file": "eval/data/rope-130.json",
+      "level_file": "data/task/rope-130.json",
+      "video_file": "data/video/rope-130.mp4"
+    },
+    {
+      "level_id": "rope-131",
+      "data_file": "eval/data/rope-131.json",
+      "level_file": "data/task/rope-131.json",
+      "video_file": "data/video/rope-131.mp4"
+    },
+    {
+      "level_id": "rope-132",
+      "data_file": "eval/data/rope-132.json",
+      "level_file": "data/task/rope-132.json",
+      "video_file": "data/video/rope-132.mp4"
+    },
+    {
+      "level_id": "rope-133",
+      "data_file": "eval/data/rope-133.json",
+      "level_file": "data/task/rope-133.json",
+      "video_file": "data/video/rope-133.mp4"
+    },
+    {
+      "level_id": "rope-134",
+      "data_file": "eval/data/rope-134.json",
+      "level_file": "data/task/rope-134.json",
+      "video_file": "data/video/rope-134.mp4"
+    },
+    {
+      "level_id": "rope-135",
+      "data_file": "eval/data/rope-135.json",
+      "level_file": "data/task/rope-135.json",
+      "video_file": "data/video/rope-135.mp4"
+    },
+    {
+      "level_id": "rope-136",
+      "data_file": "eval/data/rope-136.json",
+      "level_file": "data/task/rope-136.json",
+      "video_file": "data/video/rope-136.mp4"
+    },
+    {
+      "level_id": "rope-137",
+      "data_file": "eval/data/rope-137.json",
+      "level_file": "data/task/rope-137.json",
+      "video_file": "data/video/rope-137.mp4"
+    },
+    {
+      "level_id": "rope-138",
+      "data_file": "eval/data/rope-138.json",
+      "level_file": "data/task/rope-138.json",
+      "video_file": "data/video/rope-138.mp4"
+    },
+    {
+      "level_id": "rope-139",
+      "data_file": "eval/data/rope-139.json",
+      "level_file": "data/task/rope-139.json",
+      "video_file": "data/video/rope-139.mp4"
+    },
+    {
+      "level_id": "rope-140",
+      "data_file": "eval/data/rope-140.json",
+      "level_file": "data/task/rope-140.json",
+      "video_file": "data/video/rope-140.mp4"
+    },
+    {
+      "level_id": "rope-141",
+      "data_file": "eval/data/rope-141.json",
+      "level_file": "data/task/rope-141.json",
+      "video_file": "data/video/rope-141.mp4"
+    },
+    {
+      "level_id": "rope-142",
+      "data_file": "eval/data/rope-142.json",
+      "level_file": "data/task/rope-142.json",
+      "video_file": "data/video/rope-142.mp4"
+    },
+    {
+      "level_id": "rope-143",
+      "data_file": "eval/data/rope-143.json",
+      "level_file": "data/task/rope-143.json",
+      "video_file": "data/video/rope-143.mp4"
+    },
+    {
+      "level_id": "rope-144",
+      "data_file": "eval/data/rope-144.json",
+      "level_file": "data/task/rope-144.json",
+      "video_file": "data/video/rope-144.mp4"
+    },
+    {
+      "level_id": "rope-145",
+      "data_file": "eval/data/rope-145.json",
+      "level_file": "data/task/rope-145.json",
+      "video_file": "data/video/rope-145.mp4"
+    },
+    {
+      "level_id": "rope-146",
+      "data_file": "eval/data/rope-146.json",
+      "level_file": "data/task/rope-146.json",
+      "video_file": "data/video/rope-146.mp4"
+    },
+    {
+      "level_id": "rope-147",
+      "data_file": "eval/data/rope-147.json",
+      "level_file": "data/task/rope-147.json",
+      "video_file": "data/video/rope-147.mp4"
+    },
+    {
+      "level_id": "rope-148",
+      "data_file": "eval/data/rope-148.json",
+      "level_file": "data/task/rope-148.json",
+      "video_file": "data/video/rope-148.mp4"
+    },
+    {
+      "level_id": "rope-149",
+      "data_file": "eval/data/rope-149.json",
+      "level_file": "data/task/rope-149.json",
+      "video_file": "data/video/rope-149.mp4"
+    },
+    {
+      "level_id": "rope-150",
+      "data_file": "eval/data/rope-150.json",
+      "level_file": "data/task/rope-150.json",
+      "video_file": "data/video/rope-150.mp4"
+    },
+    {
+      "level_id": "rope-151",
+      "data_file": "eval/data/rope-151.json",
+      "level_file": "data/task/rope-151.json",
+      "video_file": "data/video/rope-151.mp4"
+    },
+    {
+      "level_id": "rope-152",
+      "data_file": "eval/data/rope-152.json",
+      "level_file": "data/task/rope-152.json",
+      "video_file": "data/video/rope-152.mp4"
+    },
+    {
+      "level_id": "rope-153",
+      "data_file": "eval/data/rope-153.json",
+      "level_file": "data/task/rope-153.json",
+      "video_file": "data/video/rope-153.mp4"
+    },
+    {
+      "level_id": "rope-154",
+      "data_file": "eval/data/rope-154.json",
+      "level_file": "data/task/rope-154.json",
+      "video_file": "data/video/rope-154.mp4"
+    },
+    {
+      "level_id": "rope-155",
+      "data_file": "eval/data/rope-155.json",
+      "level_file": "data/task/rope-155.json",
+      "video_file": "data/video/rope-155.mp4"
+    },
+    {
+      "level_id": "rope-156",
+      "data_file": "eval/data/rope-156.json",
+      "level_file": "data/task/rope-156.json",
+      "video_file": "data/video/rope-156.mp4"
+    },
+    {
+      "level_id": "rope-157",
+      "data_file": "eval/data/rope-157.json",
+      "level_file": "data/task/rope-157.json",
+      "video_file": "data/video/rope-157.mp4"
+    },
+    {
+      "level_id": "rope-158",
+      "data_file": "eval/data/rope-158.json",
+      "level_file": "data/task/rope-158.json",
+      "video_file": "data/video/rope-158.mp4"
+    },
+    {
+      "level_id": "rope-159",
+      "data_file": "eval/data/rope-159.json",
+      "level_file": "data/task/rope-159.json",
+      "video_file": "data/video/rope-159.mp4"
+    },
+    {
+      "level_id": "rope-160",
+      "data_file": "eval/data/rope-160.json",
+      "level_file": "data/task/rope-160.json",
+      "video_file": "data/video/rope-160.mp4"
+    },
+    {
+      "level_id": "rope-161",
+      "data_file": "eval/data/rope-161.json",
+      "level_file": "data/task/rope-161.json",
+      "video_file": "data/video/rope-161.mp4"
+    },
+    {
+      "level_id": "rope-162",
+      "data_file": "eval/data/rope-162.json",
+      "level_file": "data/task/rope-162.json",
+      "video_file": "data/video/rope-162.mp4"
+    },
+    {
+      "level_id": "rope-163",
+      "data_file": "eval/data/rope-163.json",
+      "level_file": "data/task/rope-163.json",
+      "video_file": "data/video/rope-163.mp4"
+    },
+    {
+      "level_id": "rope-164",
+      "data_file": "eval/data/rope-164.json",
+      "level_file": "data/task/rope-164.json",
+      "video_file": "data/video/rope-164.mp4"
+    },
+    {
+      "level_id": "rope-165",
+      "data_file": "eval/data/rope-165.json",
+      "level_file": "data/task/rope-165.json",
+      "video_file": "data/video/rope-165.mp4"
+    },
+    {
+      "level_id": "rope-166",
+      "data_file": "eval/data/rope-166.json",
+      "level_file": "data/task/rope-166.json",
+      "video_file": "data/video/rope-166.mp4"
+    },
+    {
+      "level_id": "rope-167",
+      "data_file": "eval/data/rope-167.json",
+      "level_file": "data/task/rope-167.json",
+      "video_file": "data/video/rope-167.mp4"
+    },
+    {
+      "level_id": "rope-168",
+      "data_file": "eval/data/rope-168.json",
+      "level_file": "data/task/rope-168.json",
+      "video_file": "data/video/rope-168.mp4"
+    },
+    {
+      "level_id": "rope-169",
+      "data_file": "eval/data/rope-169.json",
+      "level_file": "data/task/rope-169.json",
+      "video_file": "data/video/rope-169.mp4"
+    },
+    {
+      "level_id": "rope-170",
+      "data_file": "eval/data/rope-170.json",
+      "level_file": "data/task/rope-170.json",
+      "video_file": "data/video/rope-170.mp4"
+    },
+    {
+      "level_id": "rope-171",
+      "data_file": "eval/data/rope-171.json",
+      "level_file": "data/task/rope-171.json",
+      "video_file": "data/video/rope-171.mp4"
+    },
+    {
+      "level_id": "rope-172",
+      "data_file": "eval/data/rope-172.json",
+      "level_file": "data/task/rope-172.json",
+      "video_file": "data/video/rope-172.mp4"
+    },
+    {
+      "level_id": "rope-173",
+      "data_file": "eval/data/rope-173.json",
+      "level_file": "data/task/rope-173.json",
+      "video_file": "data/video/rope-173.mp4"
+    },
+    {
+      "level_id": "rope-174",
+      "data_file": "eval/data/rope-174.json",
+      "level_file": "data/task/rope-174.json",
+      "video_file": "data/video/rope-174.mp4"
+    },
+    {
+      "level_id": "rope-175",
+      "data_file": "eval/data/rope-175.json",
+      "level_file": "data/task/rope-175.json",
+      "video_file": "data/video/rope-175.mp4"
+    },
+    {
+      "level_id": "rope-176",
+      "data_file": "eval/data/rope-176.json",
+      "level_file": "data/task/rope-176.json",
+      "video_file": "data/video/rope-176.mp4"
+    },
+    {
+      "level_id": "rope-177",
+      "data_file": "eval/data/rope-177.json",
+      "level_file": "data/task/rope-177.json",
+      "video_file": "data/video/rope-177.mp4"
+    },
+    {
+      "level_id": "rope-178",
+      "data_file": "eval/data/rope-178.json",
+      "level_file": "data/task/rope-178.json",
+      "video_file": "data/video/rope-178.mp4"
+    },
+    {
+      "level_id": "rope-179",
+      "data_file": "eval/data/rope-179.json",
+      "level_file": "data/task/rope-179.json",
+      "video_file": "data/video/rope-179.mp4"
+    },
+    {
+      "level_id": "rope-180",
+      "data_file": "eval/data/rope-180.json",
+      "level_file": "data/task/rope-180.json",
+      "video_file": "data/video/rope-180.mp4"
+    },
+    {
+      "level_id": "rope-181",
+      "data_file": "eval/data/rope-181.json",
+      "level_file": "data/task/rope-181.json",
+      "video_file": "data/video/rope-181.mp4"
+    },
+    {
+      "level_id": "rope-182",
+      "data_file": "eval/data/rope-182.json",
+      "level_file": "data/task/rope-182.json",
+      "video_file": "data/video/rope-182.mp4"
+    },
+    {
+      "level_id": "rope-183",
+      "data_file": "eval/data/rope-183.json",
+      "level_file": "data/task/rope-183.json",
+      "video_file": "data/video/rope-183.mp4"
+    },
+    {
+      "level_id": "rope-184",
+      "data_file": "eval/data/rope-184.json",
+      "level_file": "data/task/rope-184.json",
+      "video_file": "data/video/rope-184.mp4"
+    },
+    {
+      "level_id": "rope-185",
+      "data_file": "eval/data/rope-185.json",
+      "level_file": "data/task/rope-185.json",
+      "video_file": "data/video/rope-185.mp4"
+    },
+    {
+      "level_id": "rope-186",
+      "data_file": "eval/data/rope-186.json",
+      "level_file": "data/task/rope-186.json",
+      "video_file": "data/video/rope-186.mp4"
+    },
+    {
+      "level_id": "rope-187",
+      "data_file": "eval/data/rope-187.json",
+      "level_file": "data/task/rope-187.json",
+      "video_file": "data/video/rope-187.mp4"
+    },
+    {
+      "level_id": "rope-188",
+      "data_file": "eval/data/rope-188.json",
+      "level_file": "data/task/rope-188.json",
+      "video_file": "data/video/rope-188.mp4"
+    },
+    {
+      "level_id": "rope-189",
+      "data_file": "eval/data/rope-189.json",
+      "level_file": "data/task/rope-189.json",
+      "video_file": "data/video/rope-189.mp4"
+    },
+    {
+      "level_id": "rope-190",
+      "data_file": "eval/data/rope-190.json",
+      "level_file": "data/task/rope-190.json",
+      "video_file": "data/video/rope-190.mp4"
+    },
+    {
+      "level_id": "rope-191",
+      "data_file": "eval/data/rope-191.json",
+      "level_file": "data/task/rope-191.json",
+      "video_file": "data/video/rope-191.mp4"
+    },
+    {
+      "level_id": "rope-192",
+      "data_file": "eval/data/rope-192.json",
+      "level_file": "data/task/rope-192.json",
+      "video_file": "data/video/rope-192.mp4"
+    },
+    {
+      "level_id": "rope-193",
+      "data_file": "eval/data/rope-193.json",
+      "level_file": "data/task/rope-193.json",
+      "video_file": "data/video/rope-193.mp4"
+    },
+    {
+      "level_id": "rope-194",
+      "data_file": "eval/data/rope-194.json",
+      "level_file": "data/task/rope-194.json",
+      "video_file": "data/video/rope-194.mp4"
+    },
+    {
+      "level_id": "rope-195",
+      "data_file": "eval/data/rope-195.json",
+      "level_file": "data/task/rope-195.json",
+      "video_file": "data/video/rope-195.mp4"
+    },
+    {
+      "level_id": "rope-196",
+      "data_file": "eval/data/rope-196.json",
+      "level_file": "data/task/rope-196.json",
+      "video_file": "data/video/rope-196.mp4"
+    },
+    {
+      "level_id": "rope-197",
+      "data_file": "eval/data/rope-197.json",
+      "level_file": "data/task/rope-197.json",
+      "video_file": "data/video/rope-197.mp4"
+    },
+    {
+      "level_id": "rope-198",
+      "data_file": "eval/data/rope-198.json",
+      "level_file": "data/task/rope-198.json",
+      "video_file": "data/video/rope-198.mp4"
+    },
+    {
+      "level_id": "rope-199",
+      "data_file": "eval/data/rope-199.json",
+      "level_file": "data/task/rope-199.json",
+      "video_file": "data/video/rope-199.mp4"
+    },
+    {
+      "level_id": "rope-200",
+      "data_file": "eval/data/rope-200.json",
+      "level_file": "data/task/rope-200.json",
+      "video_file": "data/video/rope-200.mp4"
+    },
+    {
+      "level_id": "rope-201",
+      "data_file": "eval/data/rope-201.json",
+      "level_file": "data/task/rope-201.json",
+      "video_file": "data/video/rope-201.mp4"
+    },
+    {
+      "level_id": "rope-202",
+      "data_file": "eval/data/rope-202.json",
+      "level_file": "data/task/rope-202.json",
+      "video_file": "data/video/rope-202.mp4"
+    },
+    {
+      "level_id": "rope-203",
+      "data_file": "eval/data/rope-203.json",
+      "level_file": "data/task/rope-203.json",
+      "video_file": "data/video/rope-203.mp4"
+    },
+    {
+      "level_id": "rope-204",
+      "data_file": "eval/data/rope-204.json",
+      "level_file": "data/task/rope-204.json",
+      "video_file": "data/video/rope-204.mp4"
+    },
+    {
+      "level_id": "rope-205",
+      "data_file": "eval/data/rope-205.json",
+      "level_file": "data/task/rope-205.json",
+      "video_file": "data/video/rope-205.mp4"
+    },
+    {
+      "level_id": "rope-206",
+      "data_file": "eval/data/rope-206.json",
+      "level_file": "data/task/rope-206.json",
+      "video_file": "data/video/rope-206.mp4"
+    },
+    {
+      "level_id": "rope-207",
+      "data_file": "eval/data/rope-207.json",
+      "level_file": "data/task/rope-207.json",
+      "video_file": "data/video/rope-207.mp4"
+    },
+    {
+      "level_id": "rope-208",
+      "data_file": "eval/data/rope-208.json",
+      "level_file": "data/task/rope-208.json",
+      "video_file": "data/video/rope-208.mp4"
+    },
+    {
+      "level_id": "rope-209",
+      "data_file": "eval/data/rope-209.json",
+      "level_file": "data/task/rope-209.json",
+      "video_file": "data/video/rope-209.mp4"
+    },
+    {
+      "level_id": "rope-210",
+      "data_file": "eval/data/rope-210.json",
+      "level_file": "data/task/rope-210.json",
+      "video_file": "data/video/rope-210.mp4"
+    },
+    {
+      "level_id": "rope-211",
+      "data_file": "eval/data/rope-211.json",
+      "level_file": "data/task/rope-211.json",
+      "video_file": "data/video/rope-211.mp4"
+    },
+    {
+      "level_id": "rope-212",
+      "data_file": "eval/data/rope-212.json",
+      "level_file": "data/task/rope-212.json",
+      "video_file": "data/video/rope-212.mp4"
+    },
+    {
+      "level_id": "rope-213",
+      "data_file": "eval/data/rope-213.json",
+      "level_file": "data/task/rope-213.json",
+      "video_file": "data/video/rope-213.mp4"
+    },
+    {
+      "level_id": "rope-214",
+      "data_file": "eval/data/rope-214.json",
+      "level_file": "data/task/rope-214.json",
+      "video_file": "data/video/rope-214.mp4"
+    },
+    {
+      "level_id": "rope-215",
+      "data_file": "eval/data/rope-215.json",
+      "level_file": "data/task/rope-215.json",
+      "video_file": "data/video/rope-215.mp4"
+    },
+    {
+      "level_id": "rope-216",
+      "data_file": "eval/data/rope-216.json",
+      "level_file": "data/task/rope-216.json",
+      "video_file": "data/video/rope-216.mp4"
+    },
+    {
+      "level_id": "rope-217",
+      "data_file": "eval/data/rope-217.json",
+      "level_file": "data/task/rope-217.json",
+      "video_file": "data/video/rope-217.mp4"
+    },
+    {
+      "level_id": "rope-218",
+      "data_file": "eval/data/rope-218.json",
+      "level_file": "data/task/rope-218.json",
+      "video_file": "data/video/rope-218.mp4"
+    },
+    {
+      "level_id": "rope-219",
+      "data_file": "eval/data/rope-219.json",
+      "level_file": "data/task/rope-219.json",
+      "video_file": "data/video/rope-219.mp4"
+    },
+    {
+      "level_id": "rope-220",
+      "data_file": "eval/data/rope-220.json",
+      "level_file": "data/task/rope-220.json",
+      "video_file": "data/video/rope-220.mp4"
+    },
+    {
+      "level_id": "rope-221",
+      "data_file": "eval/data/rope-221.json",
+      "level_file": "data/task/rope-221.json",
+      "video_file": "data/video/rope-221.mp4"
+    },
+    {
+      "level_id": "rope-222",
+      "data_file": "eval/data/rope-222.json",
+      "level_file": "data/task/rope-222.json",
+      "video_file": "data/video/rope-222.mp4"
+    },
+    {
+      "level_id": "rope-223",
+      "data_file": "eval/data/rope-223.json",
+      "level_file": "data/task/rope-223.json",
+      "video_file": "data/video/rope-223.mp4"
+    },
+    {
+      "level_id": "rope-224",
+      "data_file": "eval/data/rope-224.json",
+      "level_file": "data/task/rope-224.json",
+      "video_file": "data/video/rope-224.mp4"
+    },
+    {
+      "level_id": "rope-225",
+      "data_file": "eval/data/rope-225.json",
+      "level_file": "data/task/rope-225.json",
+      "video_file": "data/video/rope-225.mp4"
+    },
+    {
+      "level_id": "rope-226",
+      "data_file": "eval/data/rope-226.json",
+      "level_file": "data/task/rope-226.json",
+      "video_file": "data/video/rope-226.mp4"
+    },
+    {
+      "level_id": "rope-227",
+      "data_file": "eval/data/rope-227.json",
+      "level_file": "data/task/rope-227.json",
+      "video_file": "data/video/rope-227.mp4"
+    },
+    {
+      "level_id": "rope-228",
+      "data_file": "eval/data/rope-228.json",
+      "level_file": "data/task/rope-228.json",
+      "video_file": "data/video/rope-228.mp4"
+    },
+    {
+      "level_id": "rope-229",
+      "data_file": "eval/data/rope-229.json",
+      "level_file": "data/task/rope-229.json",
+      "video_file": "data/video/rope-229.mp4"
+    },
+    {
+      "level_id": "rope-230",
+      "data_file": "eval/data/rope-230.json",
+      "level_file": "data/task/rope-230.json",
+      "video_file": "data/video/rope-230.mp4"
+    },
+    {
+      "level_id": "rope-231",
+      "data_file": "eval/data/rope-231.json",
+      "level_file": "data/task/rope-231.json",
+      "video_file": "data/video/rope-231.mp4"
+    },
+    {
+      "level_id": "rope-232",
+      "data_file": "eval/data/rope-232.json",
+      "level_file": "data/task/rope-232.json",
+      "video_file": "data/video/rope-232.mp4"
+    },
+    {
+      "level_id": "rope-233",
+      "data_file": "eval/data/rope-233.json",
+      "level_file": "data/task/rope-233.json",
+      "video_file": "data/video/rope-233.mp4"
+    },
+    {
+      "level_id": "rope-234",
+      "data_file": "eval/data/rope-234.json",
+      "level_file": "data/task/rope-234.json",
+      "video_file": "data/video/rope-234.mp4"
+    },
+    {
+      "level_id": "rope-235",
+      "data_file": "eval/data/rope-235.json",
+      "level_file": "data/task/rope-235.json",
+      "video_file": "data/video/rope-235.mp4"
+    },
+    {
+      "level_id": "rope-236",
+      "data_file": "eval/data/rope-236.json",
+      "level_file": "data/task/rope-236.json",
+      "video_file": "data/video/rope-236.mp4"
+    },
+    {
+      "level_id": "rope-237",
+      "data_file": "eval/data/rope-237.json",
+      "level_file": "data/task/rope-237.json",
+      "video_file": "data/video/rope-237.mp4"
+    },
+    {
+      "level_id": "rope-238",
+      "data_file": "eval/data/rope-238.json",
+      "level_file": "data/task/rope-238.json",
+      "video_file": "data/video/rope-238.mp4"
+    },
+    {
+      "level_id": "rope-239",
+      "data_file": "eval/data/rope-239.json",
+      "level_file": "data/task/rope-239.json",
+      "video_file": "data/video/rope-239.mp4"
+    },
+    {
+      "level_id": "rope-240",
+      "data_file": "eval/data/rope-240.json",
+      "level_file": "data/task/rope-240.json",
+      "video_file": "data/video/rope-240.mp4"
+    },
+    {
+      "level_id": "rope-241",
+      "data_file": "eval/data/rope-241.json",
+      "level_file": "data/task/rope-241.json",
+      "video_file": "data/video/rope-241.mp4"
+    },
+    {
+      "level_id": "rope-242",
+      "data_file": "eval/data/rope-242.json",
+      "level_file": "data/task/rope-242.json",
+      "video_file": "data/video/rope-242.mp4"
+    },
+    {
+      "level_id": "rope-243",
+      "data_file": "eval/data/rope-243.json",
+      "level_file": "data/task/rope-243.json",
+      "video_file": "data/video/rope-243.mp4"
+    },
+    {
+      "level_id": "rope-244",
+      "data_file": "eval/data/rope-244.json",
+      "level_file": "data/task/rope-244.json",
+      "video_file": "data/video/rope-244.mp4"
+    },
+    {
+      "level_id": "rope-245",
+      "data_file": "eval/data/rope-245.json",
+      "level_file": "data/task/rope-245.json",
+      "video_file": "data/video/rope-245.mp4"
+    },
+    {
+      "level_id": "rope-246",
+      "data_file": "eval/data/rope-246.json",
+      "level_file": "data/task/rope-246.json",
+      "video_file": "data/video/rope-246.mp4"
+    },
+    {
+      "level_id": "rope-247",
+      "data_file": "eval/data/rope-247.json",
+      "level_file": "data/task/rope-247.json",
+      "video_file": "data/video/rope-247.mp4"
+    },
+    {
+      "level_id": "rope-248",
+      "data_file": "eval/data/rope-248.json",
+      "level_file": "data/task/rope-248.json",
+      "video_file": "data/video/rope-248.mp4"
+    },
+    {
+      "level_id": "rope-249",
+      "data_file": "eval/data/rope-249.json",
+      "level_file": "data/task/rope-249.json",
+      "video_file": "data/video/rope-249.mp4"
+    },
+    {
+      "level_id": "rope-250",
+      "data_file": "eval/data/rope-250.json",
+      "level_file": "data/task/rope-250.json",
+      "video_file": "data/video/rope-250.mp4"
+    },
+    {
+      "level_id": "rope-251",
+      "data_file": "eval/data/rope-251.json",
+      "level_file": "data/task/rope-251.json",
+      "video_file": "data/video/rope-251.mp4"
+    },
+    {
+      "level_id": "rope-252",
+      "data_file": "eval/data/rope-252.json",
+      "level_file": "data/task/rope-252.json",
+      "video_file": "data/video/rope-252.mp4"
+    },
+    {
+      "level_id": "rope-253",
+      "data_file": "eval/data/rope-253.json",
+      "level_file": "data/task/rope-253.json",
+      "video_file": "data/video/rope-253.mp4"
+    },
+    {
+      "level_id": "rope-254",
+      "data_file": "eval/data/rope-254.json",
+      "level_file": "data/task/rope-254.json",
+      "video_file": "data/video/rope-254.mp4"
+    },
+    {
+      "level_id": "rope-255",
+      "data_file": "eval/data/rope-255.json",
+      "level_file": "data/task/rope-255.json",
+      "video_file": "data/video/rope-255.mp4"
+    },
+    {
+      "level_id": "rope-256",
+      "data_file": "eval/data/rope-256.json",
+      "level_file": "data/task/rope-256.json",
+      "video_file": "data/video/rope-256.mp4"
+    },
+    {
+      "level_id": "rope-257",
+      "data_file": "eval/data/rope-257.json",
+      "level_file": "data/task/rope-257.json",
+      "video_file": "data/video/rope-257.mp4"
+    },
+    {
+      "level_id": "rope-258",
+      "data_file": "eval/data/rope-258.json",
+      "level_file": "data/task/rope-258.json",
+      "video_file": "data/video/rope-258.mp4"
+    },
+    {
+      "level_id": "rope-259",
+      "data_file": "eval/data/rope-259.json",
+      "level_file": "data/task/rope-259.json",
+      "video_file": "data/video/rope-259.mp4"
+    },
+    {
+      "level_id": "rope-260",
+      "data_file": "eval/data/rope-260.json",
+      "level_file": "data/task/rope-260.json",
+      "video_file": "data/video/rope-260.mp4"
+    },
+    {
+      "level_id": "rope-261",
+      "data_file": "eval/data/rope-261.json",
+      "level_file": "data/task/rope-261.json",
+      "video_file": "data/video/rope-261.mp4"
+    },
+    {
+      "level_id": "rope-262",
+      "data_file": "eval/data/rope-262.json",
+      "level_file": "data/task/rope-262.json",
+      "video_file": "data/video/rope-262.mp4"
+    },
+    {
+      "level_id": "rope-263",
+      "data_file": "eval/data/rope-263.json",
+      "level_file": "data/task/rope-263.json",
+      "video_file": "data/video/rope-263.mp4"
+    },
+    {
+      "level_id": "rope-264",
+      "data_file": "eval/data/rope-264.json",
+      "level_file": "data/task/rope-264.json",
+      "video_file": "data/video/rope-264.mp4"
+    },
+    {
+      "level_id": "rope-265",
+      "data_file": "eval/data/rope-265.json",
+      "level_file": "data/task/rope-265.json",
+      "video_file": "data/video/rope-265.mp4"
+    },
+    {
+      "level_id": "rope-266",
+      "data_file": "eval/data/rope-266.json",
+      "level_file": "data/task/rope-266.json",
+      "video_file": "data/video/rope-266.mp4"
+    },
+    {
+      "level_id": "rope-267",
+      "data_file": "eval/data/rope-267.json",
+      "level_file": "data/task/rope-267.json",
+      "video_file": "data/video/rope-267.mp4"
+    },
+    {
+      "level_id": "rope-268",
+      "data_file": "eval/data/rope-268.json",
+      "level_file": "data/task/rope-268.json",
+      "video_file": "data/video/rope-268.mp4"
+    },
+    {
+      "level_id": "rope-269",
+      "data_file": "eval/data/rope-269.json",
+      "level_file": "data/task/rope-269.json",
+      "video_file": "data/video/rope-269.mp4"
+    },
+    {
+      "level_id": "rope-270",
+      "data_file": "eval/data/rope-270.json",
+      "level_file": "data/task/rope-270.json",
+      "video_file": "data/video/rope-270.mp4"
+    },
+    {
+      "level_id": "rope-271",
+      "data_file": "eval/data/rope-271.json",
+      "level_file": "data/task/rope-271.json",
+      "video_file": "data/video/rope-271.mp4"
+    }
+  ]
+}

cutrope/data/rope-000.json

@@ -0,0 +1,160 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-000",
+  "level_file": "data/task/rope-000.json",
+  "video": {
+    "path": "data/video/rope-000.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": 6,
+      "candy": 1,
+      "grab_or_rope_anchor": 3,
+      "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": 158,
+        "y": 276
+      },
+      {
+        "name": 54,
+        "x": 64,
+        "y": 276
+      },
+      {
+        "name": 54,
+        "x": 88,
+        "y": 211
+      },
+      {
+        "name": 54,
+        "x": 88,
+        "y": 339
+      },
+      {
+        "name": 54,
+        "x": 226,
+        "y": 214
+      },
+      {
+        "name": 54,
+        "x": 252,
+        "y": 274
+      },
+      {
+        "name": 54,
+        "x": 232,
+        "y": 341
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 280,
+        "length": 80,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 52,
+        "x": 156,
+        "y": 123
+      },
+      {
+        "name": 100,
+        "x": 258,
+        "y": 117,
+        "length": 60,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 65,
+        "y": 118,
+        "length": 60,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 254,
+        "y": 277,
+        "timeout": "-1"
+      },
+      {
+        "name": 3,
+        "x": 160,
+        "y": 370,
+        "timeout": "-1"
+      },
+      {
+        "name": 3,
+        "x": 65,
+        "y": 274,
+        "timeout": "-1"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 2 \npop_bubble 3 \ncut_rope 1 when candy_still for 0.3 \npop_bubble 4 when candy_y > 570\npop_bubble 5 when candy_y > 750\npop_bubble 1 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:26:28.025Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2 \npop_bubble 3 \ncut_rope 1 when candy_still for 0.3 \npop_bubble 4 when candy_y > 570\npop_bubble 5 when candy_y > 750\npop_bubble 1 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2 \npop_bubble 3 \ncut_rope 1 when candy_still for 0.3 \npop_bubble 4 when candy_y > 570\npop_bubble 5 when candy_y > 750\npop_bubble 1 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-000\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: 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- 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-001.json

@@ -0,0 +1,221 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-001",
+  "level_file": "data/task/rope-001.json",
+  "video": {
+    "path": "data/video/rope-001.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,
+      "object_13": 1,
+      "object_4": 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": 50,
+        "x": 101,
+        "y": 170
+      },
+      {
+        "name": 51,
+        "x": 232,
+        "y": 171
+      },
+      {
+        "name": 100,
+        "x": 164,
+        "y": 146,
+        "length": 150,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 2,
+        "x": 165,
+        "y": 428
+      },
+      {
+        "name": 3,
+        "x": 187,
+        "y": 327,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 227,
+        "y": 58,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 98,
+        "y": 59,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 164,
+        "y": 146,
+        "length": 150,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 4,
+        "x": 79,
+        "y": 219,
+        "locale": "en",
+        "text": "Combine two pieces of candy before feeding it to Om Nom",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 55,
+        "y": 267,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 3,
+        "x": 139,
+        "y": 328,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 163,
+        "y": 369,
+        "timeout": -1
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 79,
+        "y": 219,
+        "locale": "ru",
+        "text": "Соедините половинки леденца прежде чем скормить его АмНяму",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 55,
+        "y": 267,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 78,
+        "y": 213,
+        "locale": "fr",
+        "text": "Associe deux morceaux de bonbon avant de le donner à Om Nom pour qu'il le mange",
+        "width": 230
+      },
+      {
+        "name": 13,
+        "x": 55,
+        "y": 267,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 56,
+        "y": 219,
+        "locale": "de",
+        "text": "Bring zwei Bonbons zusammen, bevor Du sie an Om Nom verfütterst",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 39,
+        "y": 268,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 1,2\ncut_rope 0 when candy_near 970,750,60\ncut_rope 3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:51:24.671Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1,2\ncut_rope 0 when candy_near 970,750,60\ncut_rope 3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1,2\ncut_rope 0 when candy_near 970,750,60\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-001\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-002.json

@@ -0,0 +1,129 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-002",
+  "level_file": "data/task/rope-002.json",
+  "video": {
+    "path": "data/video/rope-002.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": 360,
+        "y": 221
+      },
+      {
+        "name": 100,
+        "x": 279,
+        "y": 281,
+        "length": 95,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 277,
+        "y": 120,
+        "length": 95,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 387,
+        "y": 387
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 318,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 277,
+        "y": 403,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 225,
+        "y": 223,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 532,
+        "y": 169,
+        "length": 140,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 2 \ncut_rope 1 when candy_near 720,500,40",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:52:21.093Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2 \ncut_rope 1 when candy_near 720,500,40",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2 \ncut_rope 1 when candy_near 720,500,40",
+  "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-002\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-003.json

@@ -0,0 +1,136 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-003",
+  "level_file": "data/task/rope-003.json",
+  "video": {
+    "path": "data/video/rope-003.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": 201,
+        "y": 373
+      },
+      {
+        "name": 52,
+        "x": 532,
+        "y": 150
+      },
+      {
+        "name": 100,
+        "x": 269,
+        "y": 94,
+        "length": 260,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 533,
+        "y": 71,
+        "length": 50,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 54,
+        "x": 270,
+        "y": 390
+      },
+      {
+        "name": 54,
+        "x": 430,
+        "y": 343
+      },
+      {
+        "name": 54,
+        "x": 531,
+        "y": 209
+      },
+      {
+        "name": 3,
+        "x": 532,
+        "y": 210,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 431,
+        "y": 343,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 269,
+        "y": 390,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 491,
+        "y": 284
+      },
+      {
+        "name": 54,
+        "x": 352,
+        "y": 381
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 1 when candy_x > 1420\npop_bubble 2 when candy_x > 1440\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 < 860",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:50:48.146Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 when candy_x > 1420\npop_bubble 2 when candy_x > 1440\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 < 860",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 when candy_x > 1420\npop_bubble 2 when candy_x > 1440\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 < 860",
+  "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-003\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-004.json

@@ -0,0 +1,121 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-004",
+  "level_file": "data/task/rope-004.json",
+  "video": {
+    "path": "data/video/rope-004.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": 403,
+        "y": 124
+      },
+      {
+        "name": 54,
+        "x": 403,
+        "y": 319
+      },
+      {
+        "name": 55,
+        "x": 478,
+        "y": 126,
+        "angle": 180
+      },
+      {
+        "name": 2,
+        "x": 257,
+        "y": 420
+      },
+      {
+        "name": 52,
+        "x": 404,
+        "y": 324
+      },
+      {
+        "name": 3,
+        "x": 253,
+        "y": 317,
+        "timeout": 8
+      },
+      {
+        "name": 3,
+        "x": 260,
+        "y": 135,
+        "timeout": 9
+      },
+      {
+        "name": 3,
+        "x": 404,
+        "y": 124,
+        "timeout": 7
+      },
+      {
+        "name": 100,
+        "x": 403,
+        "y": 434,
+        "length": 80,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ],
+    "levelId": "02-19",
+    "levelEditedAt": "2026-05-05T14:21:35.907Z"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 1150,720,60\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 1150,720,60\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 1150,720,60\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-004\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-005.json

@@ -0,0 +1,152 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-005",
+  "level_file": "data/task/rope-005.json",
+  "video": {
+    "path": "data/video/rope-005.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": 3,
+      "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": 423,
+        "y": 424
+      },
+      {
+        "name": 3,
+        "x": 226,
+        "y": 158,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 402,
+        "y": 160,
+        "timeout": -1
+      },
+      {
+        "name": 55,
+        "x": 142,
+        "y": 160,
+        "angle": 0
+      },
+      {
+        "name": 3,
+        "x": 529,
+        "y": 158,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 400,
+        "y": 347
+      },
+      {
+        "name": 54,
+        "x": 314,
+        "y": 346
+      },
+      {
+        "name": 52,
+        "x": 528,
+        "y": 225
+      },
+      {
+        "name": 100,
+        "x": 529,
+        "y": 87,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 306,
+        "y": 87,
+        "length": 250,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 529,
+        "y": 352,
+        "length": 90,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      }
+    ],
+    "levelId": "02-16",
+    "levelEditedAt": "2026-05-05T14:15:54.612Z"
+  },
+  "reference_solution": "cut_rope 0 \ncut_rope 2 when candy_still for 0.3 \npop_bubble 0 when candy_y < 390\nactivate_pump 0 times 2 when candy_near 720,370,60\npop_bubble 1 when candy_y < 150\ncut_rope 1 when candy_x > 1130",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:16:10.970Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 \ncut_rope 2 when candy_still for 0.3 \npop_bubble 0 when candy_y < 390\nactivate_pump 0 times 2 when candy_near 720,370,60\npop_bubble 1 when candy_y < 150\ncut_rope 1 when candy_x > 1130",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \ncut_rope 2 when candy_still for 0.3 \npop_bubble 0 when candy_y < 390\nactivate_pump 0 times 2 when candy_near 720,370,60\npop_bubble 1 when candy_y < 150\ncut_rope 1 when candy_x > 1130",
+  "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-005\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  - 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-006.json

@@ -0,0 +1,146 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-006",
+  "level_file": "data/task/rope-006.json",
+  "video": {
+    "path": "data/video/rope-006.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": 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": 100,
+        "x": 320,
+        "y": 303,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 194,
+        "y": 174,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 442,
+        "y": 176,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 52,
+        "x": 324,
+        "y": 224
+      },
+      {
+        "name": 2,
+        "x": 322,
+        "y": 69
+      },
+      {
+        "name": 54,
+        "x": 238,
+        "y": 395
+      },
+      {
+        "name": 3,
+        "x": 415,
+        "y": 299,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 223,
+        "y": 296,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 327,
+        "y": 374,
+        "timeout": -1
+      },
+      {
+        "name": 81,
+        "x": 444,
+        "y": 397,
+        "angle": 150,
+        "size": 1
+      }
+    ],
+    "levelId": "06-03",
+    "levelEditedAt": "2026-05-05T17:13:29.998Z",
+    "mirroredFrom": "06-03.json"
+  },
+  "reference_solution": "cut_rope 1 \ncut_rope 2 when candy_x > 1250\ncut_rope 0 when candy_still for 0.3 and candy_near 960,380,60",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:13:52.716Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 \ncut_rope 2 when candy_x > 1250\ncut_rope 0 when candy_still for 0.3 and candy_near 960,380,60",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 \ncut_rope 2 when candy_x > 1250\ncut_rope 0 when candy_still for 0.3 and candy_near 960,380,60",
+  "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-006\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: 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-007.json

@@ -0,0 +1,157 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-007",
+  "level_file": "data/task/rope-007.json",
+  "video": {
+    "path": "data/video/rope-007.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": 3,
+      "bubble": 3,
+      "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": 74,
+        "y": 103
+      },
+      {
+        "name": 100,
+        "x": 69,
+        "y": 195,
+        "length": 60,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 82,
+        "x": 294,
+        "y": 413,
+        "angle": 180,
+        "size": 2
+      },
+      {
+        "name": 3,
+        "x": 400,
+        "y": 297,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 291,
+        "y": 286,
+        "timeout": -1
+      },
+      {
+        "name": 82,
+        "x": 53,
+        "y": 382,
+        "angle": 217,
+        "size": 2
+      },
+      {
+        "name": 56,
+        "x": 530,
+        "y": 292,
+        "group": 0,
+        "angle": 180
+      },
+      {
+        "name": 56,
+        "x": 164,
+        "y": 420,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 2,
+        "x": 290,
+        "y": 129
+      },
+      {
+        "name": 54,
+        "x": 399,
+        "y": 286
+      },
+      {
+        "name": 3,
+        "x": 172,
+        "y": 294,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 73,
+        "y": 105
+      },
+      {
+        "name": 54,
+        "x": 172,
+        "y": 285
+      },
+      {
+        "name": 82,
+        "x": 409,
+        "y": 412,
+        "angle": 180,
+        "size": 2
+      }
+    ],
+    "levelId": "06-20",
+    "levelEditedAt": "2026-05-05T17:37:15.591Z",
+    "mirroredFrom": "06-20.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 400,220,60\npop_bubble 1 when candy_near 400,210,60\npop_bubble 2 when candy_x > 600",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:37:47.859Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 400,220,60\npop_bubble 1 when candy_near 400,210,60\npop_bubble 2 when candy_x > 600",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 400,220,60\npop_bubble 1 when candy_near 400,210,60\npop_bubble 2 when candy_x > 600",
+  "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-007\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: 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-008.json

@@ -0,0 +1,185 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-008",
+  "level_file": "data/task/rope-008.json",
+  "video": {
+    "path": "data/video/rope-008.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": 279,
+        "y": 333
+      },
+      {
+        "name": 52,
+        "x": 207,
+        "y": 149
+      },
+      {
+        "name": 100,
+        "x": 126,
+        "y": 55,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 149,
+        "y": 234,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 214,
+        "y": 286,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 56,
+        "y": 225,
+        "timeout": -1
+      },
+      {
+        "name": 81,
+        "x": 38,
+        "y": 261,
+        "angle": 40,
+        "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
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0 when candy_x < 840",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:10:40.704Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_x < 840",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_x < 840",
+  "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-008\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-009.json

@@ -0,0 +1,145 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-009",
+  "level_file": "data/task/rope-009.json",
+  "video": {
+    "path": "data/video/rope-009.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,
+      "pump": 1,
+      "sock": 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": 2,
+        "x": 325,
+        "y": 365
+      },
+      {
+        "name": 56,
+        "x": 227,
+        "y": 62,
+        "group": 1,
+        "angle": 0
+      },
+      {
+        "name": 52,
+        "x": 220,
+        "y": 543
+      },
+      {
+        "name": 56,
+        "x": 227,
+        "y": 399,
+        "group": 1,
+        "angle": 270
+      },
+      {
+        "name": 55,
+        "x": 217,
+        "y": 188,
+        "angle": 90
+      },
+      {
+        "name": 56,
+        "x": 413,
+        "y": 399,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 418,
+        "y": 56,
+        "group": 0,
+        "angle": 180
+      },
+      {
+        "name": 54,
+        "x": 220,
+        "y": 541
+      },
+      {
+        "name": 3,
+        "x": 412,
+        "y": 282,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 223,
+        "y": 341,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 315,
+        "y": 57,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 328,
+        "y": 250,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 75,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ],
+    "levelId": "04-17",
+    "levelEditedAt": "2026-05-05T15:45:54.272Z"
+  },
+  "reference_solution": "pop_bubble 0 when candy_y < 810",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:46:31.627Z"
+  },
+  "answer": {
+    "commands": "pop_bubble 0 when candy_y < 810",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "pop_bubble 0 when candy_y < 810",
+  "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-009\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: 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-010.json

@@ -0,0 +1,158 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-010",
+  "level_file": "data/task/rope-010.json",
+  "video": {
+    "path": "data/video/rope-010.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": 5,
+      "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": 290,
+        "y": 180
+      },
+      {
+        "name": 100,
+        "x": 293,
+        "y": 61,
+        "length": 85,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 290,
+        "y": 304,
+        "length": 85,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 2,
+        "x": 166,
+        "y": 424
+      },
+      {
+        "name": 50,
+        "x": 68,
+        "y": 142
+      },
+      {
+        "name": 3,
+        "x": 36,
+        "y": 242,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 227,
+        "y": 179,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 163,
+        "y": 338,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 164,
+        "y": 63,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 163,
+        "y": 189,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 35,
+        "y": 61,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 4\ncut_rope 2 when left_candy_near 800,400,40 times 2\ncut_rope 0 when left_candy_near 1200,600,50\ncut_rope 1,3 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T16:22:49.603Z"
+  },
+  "answer": {
+    "commands": "cut_rope 4\ncut_rope 2 when left_candy_near 800,400,40 times 2\ncut_rope 0 when left_candy_near 1200,600,50\ncut_rope 1,3 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 4\ncut_rope 2 when left_candy_near 800,400,40 times 2\ncut_rope 0 when left_candy_near 1200,600,50\ncut_rope 1,3 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-010\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: 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- 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-011.json

@@ -0,0 +1,155 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-011",
+  "level_file": "data/task/rope-011.json",
+  "video": {
+    "path": "data/video/rope-011.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": 3,
+      "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": 100,
+        "x": 342,
+        "y": 131,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 260,
+        "y": 129,
+        "length": 150,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 187,
+        "y": 129,
+        "length": 250,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 2,
+        "x": 189,
+        "y": 410
+      },
+      {
+        "name": 3,
+        "x": 184,
+        "y": 299,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 196,
+        "y": 334,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 193,
+        "y": 266,
+        "timeout": -1
+      },
+      {
+        "name": 82,
+        "x": 238,
+        "y": 300,
+        "angle": 90,
+        "size": 2
+      },
+      {
+        "name": 52,
+        "x": 336,
+        "y": 269
+      },
+      {
+        "name": 54,
+        "x": 412,
+        "y": 310
+      },
+      {
+        "name": 55,
+        "x": 461,
+        "y": 94,
+        "angle": 180
+      }
+    ],
+    "levelId": "06-04",
+    "levelEditedAt": "2026-05-05T17:15:59.668Z"
+  },
+  "reference_solution": "cut_rope 0 \ncut_rope 1 when candy_x > 1050\nactivate_pump 0 when candy_y < 270\npop_bubble 0 when candy_y < 160",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:16:09.580Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 \ncut_rope 1 when candy_x > 1050\nactivate_pump 0 when candy_y < 270\npop_bubble 0 when candy_y < 160",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \ncut_rope 1 when candy_x > 1050\nactivate_pump 0 when candy_y < 270\npop_bubble 0 when candy_y < 160",
+  "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-011\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: 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-012.json

@@ -0,0 +1,205 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-012",
+  "level_file": "data/task/rope-012.json",
+  "video": {
+    "path": "data/video/rope-012.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": 8,
+      "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": 262,
+        "y": 215
+      },
+      {
+        "name": 100,
+        "x": 163,
+        "y": 43,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 51,
+        "x": 153,
+        "y": 140
+      },
+      {
+        "name": 100,
+        "x": 163,
+        "y": 261,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 50,
+        "x": 171,
+        "y": 358
+      },
+      {
+        "name": 100,
+        "x": 163,
+        "y": 450,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 163,
+        "y": 232,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 3,
+        "x": 63,
+        "y": 433,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 258,
+        "y": 76,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 164,
+        "y": 185,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 62,
+        "y": 433
+      },
+      {
+        "name": 100,
+        "x": 257,
+        "y": 141,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 66,
+        "y": 141,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 257,
+        "y": 355,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 66,
+        "y": 355,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ],
+    "levelId": "05-15",
+    "levelEditedAt": "2026-05-05T16:40:05.829Z"
+  },
+  "reference_solution": "cut_rope 1,6 when left_candy_still for 0.3\ncut_rope 2,7 when left_candy_still for 0.3 \ncut_rope 3,5 when candy_still for 0.3 \npop_bubble 0 when candy_still for 0.3 and candy_near 1170,100,60\ncut_rope 0 when candy_y > 310",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T16:42:37.450Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1,6 when left_candy_still for 0.3\ncut_rope 2,7 when left_candy_still for 0.3 \ncut_rope 3,5 when candy_still for 0.3 \npop_bubble 0 when candy_still for 0.3 and candy_near 1170,100,60\ncut_rope 0 when candy_y > 310",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1,6 when left_candy_still for 0.3\ncut_rope 2,7 when left_candy_still for 0.3 \ncut_rope 3,5 when candy_still for 0.3 \npop_bubble 0 when candy_still for 0.3 and candy_near 1170,100,60\ncut_rope 0 when candy_y > 310",
+  "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-012\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: 8\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-013.json

@@ -0,0 +1,126 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-013",
+  "level_file": "data/task/rope-013.json",
+  "video": {
+    "path": "data/video/rope-013.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": 2,
+      "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": 52,
+        "x": 74,
+        "y": 361
+      },
+      {
+        "name": 100,
+        "x": 77,
+        "y": 445,
+        "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": 3,
+        "x": 107,
+        "y": 209,
+        "timeout": -1
+      },
+      {
+        "name": 2,
+        "x": 257,
+        "y": 156
+      },
+      {
+        "name": 54,
+        "x": 74,
+        "y": 361
+      },
+      {
+        "name": 3,
+        "x": 132,
+        "y": 231,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 80,
+        "y": 187,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 156,
+        "y": 93,
+        "length": 60,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": 70,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 770,810,60\npop_bubble 0 when candy_y < 120",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:40:17.555Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 770,810,60\npop_bubble 0 when candy_y < 120",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 770,810,60\npop_bubble 0 when candy_y < 120",
+  "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-013\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-014.json

@@ -0,0 +1,171 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-014",
+  "level_file": "data/task/rope-014.json",
+  "video": {
+    "path": "data/video/rope-014.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": 158,
+        "y": 430
+      },
+      {
+        "name": 100,
+        "x": 219,
+        "y": 271,
+        "length": 100,
+        "wheel": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 284,
+        "y": 148,
+        "length": 100,
+        "wheel": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 29,
+        "y": 149,
+        "length": 100,
+        "wheel": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 93,
+        "y": 273,
+        "length": 100,
+        "wheel": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 52,
+        "x": 161,
+        "y": 159
+      },
+      {
+        "name": 3,
+        "x": 224,
+        "y": 379,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 291,
+        "y": 246,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 95,
+        "y": 377,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 222,
+        "y": 79,
+        "length": 90,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 94,
+        "y": 80,
+        "length": 90,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 158,
+        "y": 34,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      }
+    ],
+    "mirroredFrom": "01-11.json"
+  },
+  "reference_solution": "cut_rope 6,5 \ncut_rope 4 when candy_x > 1220\ncut_rope 1 when candy_still for 0.3\ncut_rope 0 when candy_still for 0.3 \ncut_rope 3 when candy_near 850,870,60 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 > 1220\ncut_rope 1 when candy_still for 0.3\ncut_rope 0 when candy_still for 0.3 \ncut_rope 3 when candy_near 850,870,60 times 2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 6,5 \ncut_rope 4 when candy_x > 1220\ncut_rope 1 when candy_still for 0.3\ncut_rope 0 when candy_still for 0.3 \ncut_rope 3 when candy_near 850,870,60 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-014\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-015.json

@@ -0,0 +1,136 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-015",
+  "level_file": "data/task/rope-015.json",
+  "video": {
+    "path": "data/video/rope-015.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": 50,
+        "x": 302,
+        "y": 324
+      },
+      {
+        "name": 51,
+        "x": 237,
+        "y": 210
+      },
+      {
+        "name": 2,
+        "x": 321,
+        "y": 397
+      },
+      {
+        "name": 100,
+        "x": 235,
+        "y": 107,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 296,
+        "y": 227,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 54,
+        "x": 239,
+        "y": 390
+      },
+      {
+        "name": 55,
+        "x": 408,
+        "y": 323,
+        "angle": 180
+      },
+      {
+        "name": 3,
+        "x": 307,
+        "y": 153,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 317,
+        "y": 110,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 240,
+        "y": 391,
+        "timeout": -1
+      }
+    ],
+    "levelId": "05-23",
+    "levelEditedAt": "2026-05-05T16:49:48.000Z"
+  },
+  "reference_solution": "cut_rope 0 \nactivate_pump 0 when right_candy_near 780,1000,50\ncut_rope 0 when candy_still for 0.3\npop_bubble 0 when candy_still for 0.3 and candy_near 910,270,60\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 right_candy_near 780,1000,50\ncut_rope 0 when candy_still for 0.3\npop_bubble 0 when candy_still for 0.3 and candy_near 910,270,60\ncut_rope 1 when candy_y > 290",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \nactivate_pump 0 when right_candy_near 780,1000,50\ncut_rope 0 when candy_still for 0.3\npop_bubble 0 when candy_still for 0.3 and candy_near 910,270,60\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-015\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-016.json

@@ -0,0 +1,218 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-016",
+  "level_file": "data/task/rope-016.json",
+  "video": {
+    "path": "data/video/rope-016.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": 3,
+      "object_12": 1,
+      "object_4": 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": 49,
+        "y": 152
+      },
+      {
+        "name": 100,
+        "x": 165,
+        "y": 250,
+        "length": 95,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 167,
+        "y": 89,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 2,
+        "x": 165,
+        "y": 430
+      },
+      {
+        "name": 3,
+        "x": 269,
+        "y": 214,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 168,
+        "y": 41,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 44,
+        "y": 254,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 68,
+        "y": 367
+      },
+      {
+        "name": 4,
+        "x": 303,
+        "y": 301,
+        "locale": "en",
+        "text": "You can restart the level by pressing the",
+        "width": 200
+      },
+      {
+        "name": 12,
+        "x": 235,
+        "y": 373,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 235,
+        "y": 348,
+        "locale": "en",
+        "text": "button",
+        "width": 100
+      },
+      {
+        "name": 100,
+        "x": 0,
+        "y": 117,
+        "length": 10,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 17,
+        "y": 304,
+        "locale": "ru",
+        "text": "Вы можете перезапустить уровень нажав кнопку",
+        "width": 200
+      },
+      {
+        "name": 12,
+        "x": 188,
+        "y": 376,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 27,
+        "y": 285,
+        "locale": "fr",
+        "text": "Tu peux recommencer le niveau en touchant le bouton",
+        "width": 180
+      },
+      {
+        "name": 12,
+        "x": 171,
+        "y": 361,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 12,
+        "x": 212,
+        "y": 329,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 31,
+        "y": 282,
+        "locale": "de",
+        "text": "Lass den Bonbon nicht aus der Schachtel",
+        "width": 175
+      }
+    ],
+    "levelId": "01-08",
+    "levelEditedAt": "2026-05-05T11:35:35.231Z",
+    "mirroredFrom": "01-08.json"
+  },
+  "reference_solution": "cut_rope 2\ncut_rope 1 when candy_near 1220,470,60\ncut_rope 0 when candy_still for 0.3 \npop_bubble 0 when candy_y < 130",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T11:36:29.406Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2\ncut_rope 1 when candy_near 1220,470,60\ncut_rope 0 when candy_still for 0.3 \npop_bubble 0 when candy_y < 130",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2\ncut_rope 1 when candy_near 1220,470,60\ncut_rope 0 when candy_still for 0.3 \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-016\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: 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-017.json

@@ -0,0 +1,150 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-017",
+  "level_file": "data/task/rope-017.json",
+  "video": {
+    "path": "data/video/rope-017.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": 3,
+      "pump": 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": 2,
+        "x": 261,
+        "y": 386
+      },
+      {
+        "name": 3,
+        "x": 261,
+        "y": 321,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 509,
+        "y": 128,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 415,
+        "y": 225,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 137,
+        "y": 28,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 52,
+        "x": 215,
+        "y": 112
+      },
+      {
+        "name": 55,
+        "x": 319,
+        "y": 243,
+        "angle": 270
+      },
+      {
+        "name": 100,
+        "x": 307,
+        "y": 33,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 441,
+        "y": 33,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 80,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 54,
+        "x": 414,
+        "y": 227
+      },
+      {
+        "name": 55,
+        "x": 495,
+        "y": 236,
+        "angle": -180
+      }
+    ],
+    "levelId": "07-14",
+    "levelEditedAt": "2026-05-05T17:48:27.305Z"
+  },
+  "reference_solution": "cut_rope 0\nactivate_pump 0 when candy_x > 900\ncut_rope 1 when candy_x > 1120\ncut_rope 2 when candy_near 1290,390,50 times 2\nactivate_pump 1 when candy_y > 500\npop_bubble 0 when candy_x < 910",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:50:14.972Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0\nactivate_pump 0 when candy_x > 900\ncut_rope 1 when candy_x > 1120\ncut_rope 2 when candy_near 1290,390,50 times 2\nactivate_pump 1 when candy_y > 500\npop_bubble 0 when candy_x < 910",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0\nactivate_pump 0 when candy_x > 900\ncut_rope 1 when candy_x > 1120\ncut_rope 2 when candy_near 1290,390,50 times 2\nactivate_pump 1 when candy_y > 500\npop_bubble 0 when candy_x < 910",
+  "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-017\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: 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-018.json

@@ -0,0 +1,150 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-018",
+  "level_file": "data/task/rope-018.json",
+  "video": {
+    "path": "data/video/rope-018.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": 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": 2,
+        "x": 287,
+        "y": 42
+      },
+      {
+        "name": 100,
+        "x": 439,
+        "y": 135,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 409,
+        "y": 263,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 270,
+        "y": 320,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 351,
+        "y": 424
+      },
+      {
+        "name": 3,
+        "x": 192,
+        "y": 278,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 362,
+        "y": 212,
+        "timeout": 10
+      },
+      {
+        "name": 52,
+        "x": 291,
+        "y": 187
+      },
+      {
+        "name": 100,
+        "x": 220,
+        "y": 69,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 3,
+        "x": 350,
+        "y": 424,
+        "timeout": -1
+      }
+    ],
+    "mirroredFrom": "02-21.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 920,430,60\ncut_rope 1 when candy_x < 890\ncut_rope 3 when candy_x < 740\ncut_rope 2 when candy_still for 0.3 and candy_near 850,420,60",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:35:13.421Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 920,430,60\ncut_rope 1 when candy_x < 890\ncut_rope 3 when candy_x < 740\ncut_rope 2 when candy_still for 0.3 and candy_near 850,420,60",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 920,430,60\ncut_rope 1 when candy_x < 890\ncut_rope 3 when candy_x < 740\ncut_rope 2 when candy_still for 0.3 and candy_near 850,420,60",
+  "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-018\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  - 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-019.json

@@ -0,0 +1,145 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-019",
+  "level_file": "data/task/rope-019.json",
+  "video": {
+    "path": "data/video/rope-019.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": 2,
+      "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": 54,
+        "x": 214,
+        "y": 376
+      },
+      {
+        "name": 56,
+        "x": 216,
+        "y": 420,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 52,
+        "x": 213,
+        "y": 160
+      },
+      {
+        "name": 100,
+        "x": 213,
+        "y": 73,
+        "length": 50,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 56,
+        "x": 292,
+        "y": 189,
+        "group": 0,
+        "angle": 0
+      },
+      {
+        "name": 3,
+        "x": 412,
+        "y": 382,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 315,
+        "y": 285,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 64,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 3,
+        "x": 314,
+        "y": 380,
+        "timeout": -1
+      },
+      {
+        "name": 2,
+        "x": 457,
+        "y": 248
+      },
+      {
+        "name": 3,
+        "x": 214,
+        "y": 376,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 313,
+        "y": 378
+      }
+    ],
+    "levelId": "04-09",
+    "levelEditedAt": "2026-05-05T15:36:33.909Z"
+  },
+  "reference_solution": "cut_rope 0 \npop_bubble 0 when candy_near 960,420,40\npop_bubble 1 when candy_near 700,800,60\ncut_rope 1 when candy_near 700,800,60",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:37:19.403Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 \npop_bubble 0 when candy_near 960,420,40\npop_bubble 1 when candy_near 700,800,60\ncut_rope 1 when candy_near 700,800,60",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 \npop_bubble 0 when candy_near 960,420,40\npop_bubble 1 when candy_near 700,800,60\ncut_rope 1 when candy_near 700,800,60",
+  "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-019\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-020.json

@@ -0,0 +1,132 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-020",
+  "level_file": "data/task/rope-020.json",
+  "video": {
+    "path": "data/video/rope-020.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,
+      "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": 193,
+        "y": 340
+      },
+      {
+        "name": 3,
+        "x": 67,
+        "y": 331,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 193,
+        "y": 169,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 78,
+        "y": 108,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 192,
+        "y": 434
+      },
+      {
+        "name": 100,
+        "x": 191,
+        "y": 262,
+        "length": 50,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 55,
+        "x": 276,
+        "y": 176,
+        "angle": 180
+      },
+      {
+        "name": 55,
+        "x": 194,
+        "y": 66,
+        "angle": 90
+      },
+      {
+        "name": 100,
+        "x": 81,
+        "y": 161,
+        "length": 100,
+        "wheel": false,
+        "radius": 50,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 2,
+        "x": 70,
+        "y": 425
+      }
+    ],
+    "levelId": "03-15",
+    "levelEditedAt": "2026-05-05T14:46:08.427Z",
+    "mirroredFrom": "03-15.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 1030,780,60\nactivate_pump 0 when candy_y < 400\npop_bubble 0 when candy_still for 0.3 and candy_near 780,140,60\ncut_rope 1",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:47:17.677Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 1030,780,60\nactivate_pump 0 when candy_y < 400\npop_bubble 0 when candy_still for 0.3 and candy_near 780,140,60\ncut_rope 1",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 1030,780,60\nactivate_pump 0 when candy_y < 400\npop_bubble 0 when candy_still for 0.3 and candy_near 780,140,60\ncut_rope 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-020\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-021.json

@@ -0,0 +1,198 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-021",
+  "level_file": "data/task/rope-021.json",
+  "video": {
+    "path": "data/video/rope-021.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": 44,
+        "y": 315,
+        "angle": 185,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 126,
+        "y": 318,
+        "angle": 181,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 208,
+        "y": 319,
+        "angle": 181,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 288,
+        "y": 320,
+        "angle": 181,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 363,
+        "y": 321,
+        "angle": 181,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 441,
+        "y": 322,
+        "angle": 181,
+        "size": 1
+      },
+      {
+        "name": 2,
+        "x": 583,
+        "y": 289
+      },
+      {
+        "name": 52,
+        "x": 70,
+        "y": 200
+      },
+      {
+        "name": 100,
+        "x": 66,
+        "y": 77,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 3,
+        "x": 444,
+        "y": 261,
+        "timeout": 9
+      },
+      {
+        "name": 3,
+        "x": 287,
+        "y": 255,
+        "timeout": 7
+      },
+      {
+        "name": 3,
+        "x": 140,
+        "y": 257,
+        "timeout": -1
+      },
+      {
+        "name": 81,
+        "x": 515,
+        "y": 323,
+        "angle": 181,
+        "size": 1
+      },
+      {
+        "name": 100,
+        "x": 167,
+        "y": 344,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 324,
+        "y": 342,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 478,
+        "y": 342,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      }
+    ],
+    "mirroredFrom": "09-08.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 380,480,60\ncut_rope 1 when candy_x > 660\ncut_rope 2 when candy_x > 1030",
+  "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 380,480,60\ncut_rope 1 when candy_x > 660\ncut_rope 2 when candy_x > 1030",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 380,480,60\ncut_rope 1 when candy_x > 660\ncut_rope 2 when candy_x > 1030",
+  "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-021\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-022.json

@@ -0,0 +1,118 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-022",
+  "level_file": "data/task/rope-022.json",
+  "video": {
+    "path": "data/video/rope-022.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,
+      "pump": 2,
+      "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": 223,
+        "y": 82
+      },
+      {
+        "name": 3,
+        "x": 158,
+        "y": 427,
+        "timeout": -1
+      },
+      {
+        "name": 52,
+        "x": 160,
+        "y": 351
+      },
+      {
+        "name": 54,
+        "x": 158,
+        "y": 426
+      },
+      {
+        "name": 100,
+        "x": 158,
+        "y": 247,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 55,
+        "x": 289,
+        "y": 287,
+        "angle": 180
+      },
+      {
+        "name": 55,
+        "x": 31,
+        "y": 211,
+        "angle": 0
+      },
+      {
+        "name": 3,
+        "x": 97,
+        "y": 234,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 222,
+        "y": 153,
+        "timeout": -1
+      }
+    ],
+    "mirroredFrom": "02-02.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_still for 0.3 and candy_near 960,770,60\nactivate_pump 0 when candy_y < 660\nactivate_pump 1 when candy_y < 490",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T13:38:26.635Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_still for 0.3 and candy_near 960,770,60\nactivate_pump 0 when candy_y < 660\nactivate_pump 1 when candy_y < 490",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_still for 0.3 and candy_near 960,770,60\nactivate_pump 0 when candy_y < 660\nactivate_pump 1 when candy_y < 490",
+  "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-022\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: 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-023.json

@@ -0,0 +1,129 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-023",
+  "level_file": "data/task/rope-023.json",
+  "video": {
+    "path": "data/video/rope-023.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": 3,
+      "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": 159,
+        "y": 350
+      },
+      {
+        "name": 2,
+        "x": 157,
+        "y": 67
+      },
+      {
+        "name": 100,
+        "x": 77,
+        "y": 290,
+        "length": 120,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 100,
+        "x": 237,
+        "y": 401,
+        "length": 90,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 54,
+        "x": 155,
+        "y": 421
+      },
+      {
+        "name": 100,
+        "x": 155,
+        "y": 123,
+        "length": 200,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false
+      },
+      {
+        "name": 3,
+        "x": 223,
+        "y": 294,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 154,
+        "y": 422,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 223,
+        "y": 228,
+        "timeout": -1
+      }
+    ],
+    "mirroredFrom": "01-09.json"
+  },
+  "reference_solution": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 0 when candy_y < 370",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T11:41:55.569Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 0 when candy_y < 370",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2\ncut_rope 1 when candy_still for 0.3 \ncut_rope 0 when candy_y < 370",
+  "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-023\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: 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-024.json

@@ -0,0 +1,180 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-024",
+  "level_file": "data/task/rope-024.json",
+  "video": {
+    "path": "data/video/rope-024.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": 5,
+      "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": 185,
+        "y": 243
+      },
+      {
+        "name": 100,
+        "x": 220,
+        "y": 110,
+        "length": 78,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 223,
+        "y": 366,
+        "length": 78,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 30,
+        "y": 432,
+        "length": 210,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 3,
+        "x": 117,
+        "y": 153,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 225,
+        "y": 235,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 190,
+        "y": 113,
+        "timeout": -1
+      },
+      {
+        "name": 2,
+        "x": 26,
+        "y": 349
+      },
+      {
+        "name": 56,
+        "x": 242,
+        "y": 232,
+        "group": 0,
+        "angle": 180
+      },
+      {
+        "name": 56,
+        "x": 241,
+        "y": 62,
+        "group": 0,
+        "angle": 180
+      },
+      {
+        "name": 100,
+        "x": 94,
+        "y": 240,
+        "length": 68,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 30,
+        "y": 55,
+        "length": 210,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      }
+    ],
+    "levelId": "04-11",
+    "levelEditedAt": "2026-05-05T15:38:36.315Z",
+    "mirroredFrom": "04-11.json"
+  },
+  "reference_solution": "cut_rope 3,4,2",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:38:51.556Z"
+  },
+  "answer": {
+    "commands": "cut_rope 3,4,2",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 3,4,2",
+  "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-024\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: 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\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-025.json

@@ -0,0 +1,161 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-025",
+  "level_file": "data/task/rope-025.json",
+  "video": {
+    "path": "data/video/rope-025.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": 4,
+      "bubble": 1,
+      "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,
+        "twoParts": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 50,
+        "x": 166,
+        "y": 360
+      },
+      {
+        "name": 54,
+        "x": 163,
+        "y": 358
+      },
+      {
+        "name": 2,
+        "x": 163,
+        "y": 439
+      },
+      {
+        "name": 81,
+        "x": 83,
+        "y": 266,
+        "angle": 20,
+        "size": 1
+      },
+      {
+        "name": 81,
+        "x": 255,
+        "y": 263,
+        "angle": -20,
+        "size": 1
+      },
+      {
+        "name": 51,
+        "x": 101,
+        "y": 176
+      },
+      {
+        "name": 3,
+        "x": 164,
+        "y": 298,
+        "timeout": 6
+      },
+      {
+        "name": 3,
+        "x": 167,
+        "y": 141,
+        "timeout": 7
+      },
+      {
+        "name": 3,
+        "x": 169,
+        "y": 79,
+        "timeout": 8
+      },
+      {
+        "name": 82,
+        "x": 220,
+        "y": 64,
+        "angle": 20,
+        "size": 2
+      },
+      {
+        "name": 82,
+        "x": 120,
+        "y": 64,
+        "angle": -20,
+        "size": 2
+      },
+      {
+        "name": 100,
+        "x": 261,
+        "y": 446,
+        "length": 100,
+        "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": 68,
+        "y": 444,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "kicked": false,
+        "radius": -1,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0,1 when candy_still for 0.3\npop_bubble 0 when candy_y < 190",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:42:35.574Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1 when candy_still for 0.3\npop_bubble 0 when candy_y < 190",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1 when candy_still for 0.3\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-025\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- 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-026.json

@@ -0,0 +1,210 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-026",
+  "level_file": "data/task/rope-026.json",
+  "video": {
+    "path": "data/video/rope-026.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,
+      "object_10": 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": 52,
+        "x": 238,
+        "y": 178
+      },
+      {
+        "name": 100,
+        "x": 163,
+        "y": 101,
+        "length": 80,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 296,
+        "y": 98,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 138,
+        "y": 435,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 253,
+        "y": 118,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 253,
+        "y": 336,
+        "timeout": -1
+      },
+      {
+        "name": 2,
+        "x": 75,
+        "y": 421
+      },
+      {
+        "name": 54,
+        "x": 252,
+        "y": 228
+      },
+      {
+        "name": 56,
+        "x": 259,
+        "y": 390,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 158,
+        "y": 428,
+        "group": 0,
+        "angle": 180
+      },
+      {
+        "name": 10,
+        "x": 112,
+        "y": 256,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 38,
+        "y": 273,
+        "locale": "en",
+        "text": "Keep the candy away from spikes",
+        "width": 150
+      }
+    ],
+    "ru": [
+      {
+        "name": 10,
+        "x": 112,
+        "y": 256,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 14,
+        "y": 269,
+        "locale": "ru",
+        "text": "Не дайте леденцу разбиться о шипы",
+        "width": 200
+      }
+    ],
+    "fr": [
+      {
+        "name": 10,
+        "x": 111,
+        "y": 257,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 34,
+        "y": 268,
+        "locale": "fr",
+        "text": "Garde le bonbon loin des pointes",
+        "width": 150
+      }
+    ],
+    "de": [
+      {
+        "name": 10,
+        "x": 112,
+        "y": 256,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 26,
+        "y": 273,
+        "locale": "de",
+        "text": "Pass auf, dass der Bonbon nicht in die Nähe der Spikes kommt",
+        "width": 250
+      }
+    ],
+    "levelId": "00-19",
+    "levelEditedAt": "2026-05-05T10:29:46.378Z"
+  },
+  "reference_solution": "cut_rope 0,1 \npop_bubble 0 when candy_y < 290",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:30:04.840Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1 \npop_bubble 0 when candy_y < 290",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1 \npop_bubble 0 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-026\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-027.json

@@ -0,0 +1,159 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-027",
+  "level_file": "data/task/rope-027.json",
+  "video": {
+    "path": "data/video/rope-027.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": 5,
+      "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": 50,
+        "x": 30,
+        "y": 180
+      },
+      {
+        "name": 100,
+        "x": 27,
+        "y": 61,
+        "length": 85,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 30,
+        "y": 304,
+        "length": 85,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 154,
+        "y": 424
+      },
+      {
+        "name": 51,
+        "x": 252,
+        "y": 142
+      },
+      {
+        "name": 3,
+        "x": 284,
+        "y": 242,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 93,
+        "y": 179,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 157,
+        "y": 338,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 156,
+        "y": 63,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 189,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 285,
+        "y": 61,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "mirroredFrom": "05-06.json"
+  },
+  "reference_solution": "cut_rope 4\ncut_rope 2 when right_candy_near 1120,400,40 times 2\ncut_rope 0 when right_candy_near 720,600,50\ncut_rope 1,3 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T16:22:49.603Z"
+  },
+  "answer": {
+    "commands": "cut_rope 4\ncut_rope 2 when right_candy_near 1120,400,40 times 2\ncut_rope 0 when right_candy_near 720,600,50\ncut_rope 1,3 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 4\ncut_rope 2 when right_candy_near 1120,400,40 times 2\ncut_rope 0 when right_candy_near 720,600,50\ncut_rope 1,3 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-027\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: 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- 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-028.json

@@ -0,0 +1,153 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-028",
+  "level_file": "data/task/rope-028.json",
+  "video": {
+    "path": "data/video/rope-028.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": 72,
+        "y": 65
+      },
+      {
+        "name": 100,
+        "x": 137,
+        "y": 219,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 288,
+        "y": 218,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 52,
+        "x": 207,
+        "y": 317
+      },
+      {
+        "name": 3,
+        "x": 212,
+        "y": 244,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 66,
+        "y": 252,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 150,
+        "y": 122,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 207,
+        "y": 182
+      },
+      {
+        "name": 54,
+        "x": 211,
+        "y": 242
+      },
+      {
+        "name": 54,
+        "x": 64,
+        "y": 310
+      },
+      {
+        "name": 55,
+        "x": 269,
+        "y": 160,
+        "angle": 180
+      },
+      {
+        "name": 56,
+        "x": 213,
+        "y": 408,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 66,
+        "y": 371,
+        "group": 0,
+        "angle": 270
+      }
+    ]
+  },
+  "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-028\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-029.json

@@ -0,0 +1,164 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-029",
+  "level_file": "data/task/rope-029.json",
+  "video": {
+    "path": "data/video/rope-029.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": {
+      "bubble": 7,
+      "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": 320,
+        "height": 480
+      },
+      {
+        "name": 1,
+        "ropePhysicsSpeed": 1,
+        "twoParts": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 51,
+        "x": 272,
+        "y": 135
+      },
+      {
+        "name": 50,
+        "x": 44,
+        "y": 133
+      },
+      {
+        "name": 2,
+        "x": 73,
+        "y": 412
+      },
+      {
+        "name": 54,
+        "x": 163,
+        "y": 322
+      },
+      {
+        "name": 54,
+        "x": 217,
+        "y": 271
+      },
+      {
+        "name": 54,
+        "x": 104,
+        "y": 276
+      },
+      {
+        "name": 54,
+        "x": 260,
+        "y": 216
+      },
+      {
+        "name": 54,
+        "x": 58,
+        "y": 217
+      },
+      {
+        "name": 54,
+        "x": 220,
+        "y": 160
+      },
+      {
+        "name": 54,
+        "x": 96,
+        "y": 161
+      },
+      {
+        "name": 3,
+        "x": 197,
+        "y": 227,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 121,
+        "y": 224,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 264,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 253,
+        "length": 140,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 158,
+        "y": 253,
+        "length": 140,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 55,
+        "x": 249,
+        "y": 416,
+        "angle": 180
+      }
+    ],
+    "levelId": "05-12",
+    "levelEditedAt": "2026-05-05T16:31:03.290Z",
+    "mirroredFrom": "05-12.json"
+  },
+  "reference_solution": "pop_bubble 3 when right_candy_x < 1120 \npop_bubble 4 when left_candy_x > 820\npop_bubble 5 when right_candy_y <300 \npop_bubble 6 when left_candy_y < 300\npop_bubble 0 when candy_still for 0.3\nactivate_pump 0 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T16:36:31.795Z"
+  },
+  "answer": {
+    "commands": "pop_bubble 3 when right_candy_x < 1120 \npop_bubble 4 when left_candy_x > 820\npop_bubble 5 when right_candy_y <300 \npop_bubble 6 when left_candy_y < 300\npop_bubble 0 when candy_still for 0.3\nactivate_pump 0 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "pop_bubble 3 when right_candy_x < 1120 \npop_bubble 4 when left_candy_x > 820\npop_bubble 5 when right_candy_y <300 \npop_bubble 6 when left_candy_y < 300\npop_bubble 0 when candy_still for 0.3\nactivate_pump 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-029\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: 7\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-030.json

@@ -0,0 +1,124 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-030",
+  "level_file": "data/task/rope-030.json",
+  "video": {
+    "path": "data/video/rope-030.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,
+      "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": 29,
+        "y": -9
+      },
+      {
+        "name": 2,
+        "x": 258,
+        "y": 296
+      },
+      {
+        "name": 54,
+        "x": 142,
+        "y": 418
+      },
+      {
+        "name": 100,
+        "x": 93,
+        "y": 315,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 100,
+        "x": 257,
+        "y": 161,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R",
+        "hidePath": false
+      },
+      {
+        "name": 3,
+        "x": 192,
+        "y": 86,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 31,
+        "y": 197,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 257,
+        "y": 243,
+        "timeout": -1
+      }
+    ],
+    "mirroredFrom": "01-19.json"
+  },
+  "reference_solution": "cut_rope 0 when candy_x > 1050\npop_bubble 0 when candy_still for 0.3 \ncut_rope 1 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T13:18:57.463Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_x > 1050\npop_bubble 0 when candy_still for 0.3 \ncut_rope 1 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_x > 1050\npop_bubble 0 when candy_still for 0.3 \ncut_rope 1 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-030\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-031.json

@@ -0,0 +1,206 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-031",
+  "level_file": "data/task/rope-031.json",
+  "video": {
+    "path": "data/video/rope-031.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": 8,
+      "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": 58,
+        "y": 215
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 43,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 50,
+        "x": 167,
+        "y": 140
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 261,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 51,
+        "x": 149,
+        "y": 358
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 450,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 232,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 257,
+        "y": 433,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 62,
+        "y": 76,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 156,
+        "y": 185,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 258,
+        "y": 433
+      },
+      {
+        "name": 100,
+        "x": 63,
+        "y": 141,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 254,
+        "y": 141,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 63,
+        "y": 355,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 254,
+        "y": 355,
+        "length": 60,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "levelId": "05-15",
+    "levelEditedAt": "2026-05-05T16:40:05.829Z",
+    "mirroredFrom": "05-15.json"
+  },
+  "reference_solution": "cut_rope 1,6 when right_candy_still for 0.3\ncut_rope 2,7 when right_candy_still for 0.3 \ncut_rope 3,5 when candy_still for 0.3 \npop_bubble 0 when candy_still for 0.3 and candy_near 750,100,60\ncut_rope 0 when candy_y > 310",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T16:42:37.450Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1,6 when right_candy_still for 0.3\ncut_rope 2,7 when right_candy_still for 0.3 \ncut_rope 3,5 when candy_still for 0.3 \npop_bubble 0 when candy_still for 0.3 and candy_near 750,100,60\ncut_rope 0 when candy_y > 310",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1,6 when right_candy_still for 0.3\ncut_rope 2,7 when right_candy_still for 0.3 \ncut_rope 3,5 when candy_still for 0.3 \npop_bubble 0 when candy_still for 0.3 and candy_near 750,100,60\ncut_rope 0 when candy_y > 310",
+  "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-031\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: 8\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-032.json

@@ -0,0 +1,143 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-032",
+  "level_file": "data/task/rope-032.json",
+  "video": {
+    "path": "data/video/rope-032.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": 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": 2,
+        "x": 450,
+        "y": 395
+      },
+      {
+        "name": 52,
+        "x": 246,
+        "y": 183
+      },
+      {
+        "name": 100,
+        "x": 183,
+        "y": 67,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 187,
+        "y": 244,
+        "length": 180,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 459,
+        "y": 68,
+        "length": 320,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 323,
+        "y": 70,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 362,
+        "y": 338,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 415,
+        "y": 221,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 307,
+        "y": 218,
+        "timeout": -1
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0\ncut_rope 3 when candy_x > 1100\ncut_rope 1 when candy_y > 870",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T13:16:32.813Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0\ncut_rope 3 when candy_x > 1100\ncut_rope 1 when candy_y > 870",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0\ncut_rope 3 when candy_x > 1100\ncut_rope 1 when candy_y > 870",
+  "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-032\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-033.json

@@ -0,0 +1,217 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-033",
+  "level_file": "data/task/rope-033.json",
+  "video": {
+    "path": "data/video/rope-033.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": 4,
+      "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": 2,
+        "x": 508,
+        "y": 424
+      },
+      {
+        "name": 100,
+        "x": 224,
+        "y": 177,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 441,
+        "y": 341,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 335,
+        "y": 256,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 45,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 52,
+        "x": 178,
+        "y": 103
+      },
+      {
+        "name": 3,
+        "x": 222,
+        "y": 269,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 428,
+        "y": 426,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 328,
+        "y": 345,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 180,
+        "y": 31,
+        "length": 50,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 4,
+        "x": 273,
+        "y": 39,
+        "locale": "en",
+        "text": "Automatic ropes appear when candy gets into their area",
+        "width": 140
+      },
+      {
+        "name": 8,
+        "x": 316,
+        "y": 164,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "ru": [
+      {
+        "name": 8,
+        "x": 325,
+        "y": 173,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 234,
+        "y": 68,
+        "locale": "ru",
+        "text": "Авто-веревки появляются, когда леденец попадает в их радиус",
+        "width": 220
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 213,
+        "y": 41,
+        "locale": "fr",
+        "text": "Les cordes automatiques apparaissent lorsqu'un bonbon arrive dans leur zone",
+        "width": 260
+      },
+      {
+        "name": 8,
+        "x": 298,
+        "y": 144,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 250,
+        "y": 47,
+        "locale": "de",
+        "text": "Sobald der Bonbon in ihrer Nähe ist, erscheinen automatisch Seile",
+        "width": 200
+      },
+      {
+        "name": 8,
+        "x": 329,
+        "y": 152,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "levelId": "01-10",
+    "levelEditedAt": "2026-05-05T12:58:50.336Z"
+  },
+  "reference_solution": "cut_rope 3 \ncut_rope 0 when candy_still for 0.3\ncut_rope 2 when candy_still for 0.3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T12:59:09.942Z"
+  },
+  "answer": {
+    "commands": "cut_rope 3 \ncut_rope 0 when candy_still for 0.3\ncut_rope 2 when candy_still for 0.3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 3 \ncut_rope 0 when candy_still for 0.3\ncut_rope 2 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-033\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-034.json

@@ -0,0 +1,130 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-034",
+  "level_file": "data/task/rope-034.json",
+  "video": {
+    "path": "data/video/rope-034.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": {
+      "bouncer": 1,
+      "bubble": 1,
+      "left_candy": 1,
+      "pump": 1,
+      "right_candy": 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": true
+      }
+    ],
+    "objects": [
+      {
+        "name": 54,
+        "x": 199,
+        "y": 435
+      },
+      {
+        "name": 51,
+        "x": 205,
+        "y": 441
+      },
+      {
+        "name": 50,
+        "x": 292,
+        "y": 125
+      },
+      {
+        "name": 55,
+        "x": 436,
+        "y": 148,
+        "angle": 180
+      },
+      {
+        "name": 56,
+        "x": 176,
+        "y": 68,
+        "group": 0,
+        "angle": 90
+      },
+      {
+        "name": 56,
+        "x": 435,
+        "y": 406,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 2,
+        "x": 433,
+        "y": 259
+      },
+      {
+        "name": 3,
+        "x": 197,
+        "y": 122,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 435,
+        "y": 348,
+        "timeout": -1
+      },
+      {
+        "name": 82,
+        "x": 291,
+        "y": 199,
+        "angle": 0,
+        "size": 2
+      },
+      {
+        "name": 3,
+        "x": 245,
+        "y": 124,
+        "timeout": -1
+      }
+    ],
+    "levelId": "06-18",
+    "levelEditedAt": "2026-05-05T17:32:43.877Z"
+  },
+  "reference_solution": "activate_pump 0 when right_candy_y<400",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:33:48.589Z"
+  },
+  "answer": {
+    "commands": "activate_pump 0 when right_candy_y<400",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "activate_pump 0 when right_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-034\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  - 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- 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-035.json

@@ -0,0 +1,222 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-035",
+  "level_file": "data/task/rope-035.json",
+  "video": {
+    "path": "data/video/rope-035.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,
+      "object_13": 1,
+      "object_4": 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": 219,
+        "y": 170
+      },
+      {
+        "name": 50,
+        "x": 88,
+        "y": 171
+      },
+      {
+        "name": 100,
+        "x": 156,
+        "y": 146,
+        "length": 150,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 155,
+        "y": 428
+      },
+      {
+        "name": 3,
+        "x": 133,
+        "y": 327,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 93,
+        "y": 58,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 222,
+        "y": 59,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 156,
+        "y": 146,
+        "length": 150,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 4,
+        "x": 241,
+        "y": 219,
+        "locale": "en",
+        "text": "Combine two pieces of candy before feeding it to Om Nom",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 265,
+        "y": 267,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 3,
+        "x": 181,
+        "y": 328,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 157,
+        "y": 369,
+        "timeout": -1
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 79,
+        "y": 219,
+        "locale": "ru",
+        "text": "Соедините половинки леденца прежде чем скормить его АмНяму",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 55,
+        "y": 267,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 78,
+        "y": 213,
+        "locale": "fr",
+        "text": "Associe deux morceaux de bonbon avant de le donner à Om Nom pour qu'il le mange",
+        "width": 230
+      },
+      {
+        "name": 13,
+        "x": 55,
+        "y": 267,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 56,
+        "y": 219,
+        "locale": "de",
+        "text": "Bring zwei Bonbons zusammen, bevor Du sie an Om Nom verfütterst",
+        "width": 200
+      },
+      {
+        "name": 13,
+        "x": 39,
+        "y": 268,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "mirroredFrom": "05-01.json"
+  },
+  "reference_solution": "cut_rope 1,2\ncut_rope 0 when candy_near 950,750,60\ncut_rope 3",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:51:24.671Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1,2\ncut_rope 0 when candy_near 950,750,60\ncut_rope 3",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1,2\ncut_rope 0 when candy_near 950,750,60\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-035\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-036.json

@@ -0,0 +1,262 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-036",
+  "level_file": "data/task/rope-036.json",
+  "video": {
+    "path": "data/video/rope-036.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,
+      "object_13": 1,
+      "object_4": 2,
+      "object_8": 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": 159,
+        "y": 425
+      },
+      {
+        "name": 52,
+        "x": 63,
+        "y": 238
+      },
+      {
+        "name": 3,
+        "x": 97,
+        "y": 311,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 140,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 66,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 63,
+        "y": 123,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 159,
+        "y": 245,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 54,
+        "x": 159,
+        "y": 337
+      },
+      {
+        "name": 4,
+        "x": 195,
+        "y": 289,
+        "locale": "en",
+        "text": "The bubble will lift the candy up",
+        "width": 110
+      },
+      {
+        "name": 8,
+        "x": 213,
+        "y": 367,
+        "locale": "en",
+        "angle": 50,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 185,
+        "y": 155,
+        "locale": "en",
+        "text": "Click to pop the bubble",
+        "width": 120
+      },
+      {
+        "name": 13,
+        "x": 241,
+        "y": 143,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 169,
+        "y": 284,
+        "locale": "ru",
+        "text": "Пузырь подымет леденец вверх",
+        "width": 160
+      },
+      {
+        "name": 8,
+        "x": 210,
+        "y": 350,
+        "locale": "ru",
+        "angle": 20,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 162,
+        "y": 161,
+        "locale": "ru",
+        "text": "Нажмите, чтобы лопнуть пузырь",
+        "width": 160
+      },
+      {
+        "name": 13,
+        "x": 239,
+        "y": 153,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 166,
+        "y": 255,
+        "locale": "fr",
+        "text": "La bulle va soulever le bonbon dans les airs",
+        "width": 150
+      },
+      {
+        "name": 8,
+        "x": 210,
+        "y": 350,
+        "locale": "fr",
+        "angle": 15,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 185,
+        "y": 155,
+        "locale": "fr",
+        "text": "Clique pour éclater la bulle",
+        "width": 160
+      },
+      {
+        "name": 13,
+        "x": 241,
+        "y": 147,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 194,
+        "y": 283,
+        "locale": "de",
+        "text": "Die Blase lässt die Süßigkeit schweben",
+        "width": 130
+      },
+      {
+        "name": 8,
+        "x": 212,
+        "y": 360,
+        "locale": "de",
+        "angle": 50,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 175,
+        "y": 154,
+        "locale": "de",
+        "text": "Klicke, um die Seifenblase platzen zu lassen",
+        "width": 140
+      },
+      {
+        "name": 13,
+        "x": 241,
+        "y": 146,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0\ncut_rope 1 when candy_near 950,300,30 for 1\npop_bubble 0 when candy_y<120",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T07:00:41.641Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0\ncut_rope 1 when candy_near 950,300,30 for 1\npop_bubble 0 when candy_y<120",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0\ncut_rope 1 when candy_near 950,300,30 for 1\npop_bubble 0 when candy_y<120",
+  "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-036\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-037.json

@@ -0,0 +1,125 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-037",
+  "level_file": "data/task/rope-037.json",
+  "video": {
+    "path": "data/video/rope-037.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": 529,
+        "y": 345
+      },
+      {
+        "name": 3,
+        "x": 344,
+        "y": 260,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 418,
+        "y": 358,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 319,
+        "y": 89,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 219,
+        "y": 61,
+        "length": 60,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 52,
+        "x": 220,
+        "y": 150
+      },
+      {
+        "name": 55,
+        "x": 119,
+        "y": 153,
+        "angle": 0
+      },
+      {
+        "name": 100,
+        "x": 390,
+        "y": 261,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 50,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "levelId": "03-19",
+    "levelEditedAt": "2026-05-05T15:19:29.844Z",
+    "mirroredFrom": "03-19.json"
+  },
+  "reference_solution": "activate_pump 0\ncut_rope 0 when candy_x > 920\nactivate_pump 0 when candy_x > 1250\ncut_rope 1 when candy_y > 630",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:19:35.838Z"
+  },
+  "answer": {
+    "commands": "activate_pump 0\ncut_rope 0 when candy_x > 920\nactivate_pump 0 when candy_x > 1250\ncut_rope 1 when candy_y > 630",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "activate_pump 0\ncut_rope 0 when candy_x > 920\nactivate_pump 0 when candy_x > 1250\ncut_rope 1 when candy_y > 630",
+  "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-037\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-038.json

@@ -0,0 +1,131 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-038",
+  "level_file": "data/task/rope-038.json",
+  "video": {
+    "path": "data/video/rope-038.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": 2,
+        "x": 340,
+        "y": 431
+      },
+      {
+        "name": 52,
+        "x": 314,
+        "y": 334
+      },
+      {
+        "name": 100,
+        "x": 317,
+        "y": 195,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 199,
+        "y": 401,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 434,
+        "y": 405,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 391,
+        "y": 173,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 220,
+        "y": 181,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 316,
+        "y": 375,
+        "timeout": -1
+      }
+    ],
+    "levelId": "03-16",
+    "levelEditedAt": "2026-05-05T14:51:11.811Z"
+  },
+  "reference_solution": "cut_rope 1 when candy_still for 0.3 and candy_near 960,760,60\ncut_rope 2 \ncut_rope 0 when candy_near 750,600,60",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:51:18.915Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1 when candy_still for 0.3 and candy_near 960,760,60\ncut_rope 2 \ncut_rope 0 when candy_near 750,600,60",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1 when candy_still for 0.3 and candy_near 960,760,60\ncut_rope 2 \ncut_rope 0 when candy_near 750,600,60",
+  "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-038\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-039.json

@@ -0,0 +1,217 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-039",
+  "level_file": "data/task/rope-039.json",
+  "video": {
+    "path": "data/video/rope-039.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": 3,
+      "object_12": 1,
+      "object_4": 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": 271,
+        "y": 152
+      },
+      {
+        "name": 100,
+        "x": 155,
+        "y": 250,
+        "length": 95,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 153,
+        "y": 89,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 155,
+        "y": 430
+      },
+      {
+        "name": 3,
+        "x": 51,
+        "y": 214,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 152,
+        "y": 41,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 276,
+        "y": 254,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 252,
+        "y": 367
+      },
+      {
+        "name": 4,
+        "x": 17,
+        "y": 301,
+        "locale": "en",
+        "text": "You can restart the level by pressing the",
+        "width": 200
+      },
+      {
+        "name": 12,
+        "x": 85,
+        "y": 373,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 85,
+        "y": 348,
+        "locale": "en",
+        "text": "button",
+        "width": 100
+      },
+      {
+        "name": 100,
+        "x": 320,
+        "y": 117,
+        "length": 10,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ],
+    "ru": [
+      {
+        "name": 4,
+        "x": 17,
+        "y": 304,
+        "locale": "ru",
+        "text": "Вы можете перезапустить уровень нажав кнопку",
+        "width": 200
+      },
+      {
+        "name": 12,
+        "x": 188,
+        "y": 376,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 27,
+        "y": 285,
+        "locale": "fr",
+        "text": "Tu peux recommencer le niveau en touchant le bouton",
+        "width": 180
+      },
+      {
+        "name": 12,
+        "x": 171,
+        "y": 361,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 12,
+        "x": 212,
+        "y": 329,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 31,
+        "y": 282,
+        "locale": "de",
+        "text": "Lass den Bonbon nicht aus der Schachtel",
+        "width": 175
+      }
+    ],
+    "levelId": "01-08",
+    "levelEditedAt": "2026-05-05T11:35:35.231Z"
+  },
+  "reference_solution": "cut_rope 2\ncut_rope 1 when candy_near 700,470,60\ncut_rope 0 when candy_still for 0.3 \npop_bubble 0 when candy_y < 130",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T11:36:29.406Z"
+  },
+  "answer": {
+    "commands": "cut_rope 2\ncut_rope 1 when candy_near 700,470,60\ncut_rope 0 when candy_still for 0.3 \npop_bubble 0 when candy_y < 130",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 2\ncut_rope 1 when candy_near 700,470,60\ncut_rope 0 when candy_still for 0.3 \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-039\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: 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-040.json

@@ -0,0 +1,143 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-040",
+  "level_file": "data/task/rope-040.json",
+  "video": {
+    "path": "data/video/rope-040.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": 2,
+      "bubble": 2,
+      "grab_or_rope_anchor": 1,
+      "left_candy": 1,
+      "pump": 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": 82,
+        "x": 63,
+        "y": 144,
+        "angle": 180,
+        "size": 2
+      },
+      {
+        "name": 82,
+        "x": 238,
+        "y": 54,
+        "angle": 180,
+        "size": 2
+      },
+      {
+        "name": 51,
+        "x": 232,
+        "y": 115
+      },
+      {
+        "name": 54,
+        "x": 241,
+        "y": 112
+      },
+      {
+        "name": 54,
+        "x": 78,
+        "y": 210
+      },
+      {
+        "name": 50,
+        "x": 73,
+        "y": 213
+      },
+      {
+        "name": 2,
+        "x": 167,
+        "y": 424
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 53,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 160,
+        "y": 101,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 159,
+        "y": 152,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 160,
+        "y": 275,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "kickable": false,
+        "radius": 60,
+        "invisible": false,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 55,
+        "x": 250,
+        "y": 386,
+        "angle": 180
+      }
+    ],
+    "mirroredFrom": "06-10.json"
+  },
+  "reference_solution": "pop_bubble 0\nactivate_pump 0 when right_candy_near 1120,850,50\ncut_rope 0 when candy_still for 0.3 and candy_near 960,370,60\npop_bubble 1 when candy_y < 170",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:22:25.471Z"
+  },
+  "answer": {
+    "commands": "pop_bubble 0\nactivate_pump 0 when right_candy_near 1120,850,50\ncut_rope 0 when candy_still for 0.3 and candy_near 960,370,60\npop_bubble 1 when candy_y < 170",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "pop_bubble 0\nactivate_pump 0 when right_candy_near 1120,850,50\ncut_rope 0 when candy_still for 0.3 and candy_near 960,370,60\npop_bubble 1 when candy_y < 170",
+  "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-040\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: 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- 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-041.json

@@ -0,0 +1,116 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-041",
+  "level_file": "data/task/rope-041.json",
+  "video": {
+    "path": "data/video/rope-041.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,
+      "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": 179,
+        "y": 427
+      },
+      {
+        "name": 52,
+        "x": 238,
+        "y": 523
+      },
+      {
+        "name": 3,
+        "x": 238,
+        "y": 378,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 169,
+        "y": 282,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 170,
+        "y": 28,
+        "timeout": 20
+      },
+      {
+        "name": 53,
+        "x": 54,
+        "y": 109
+      },
+      {
+        "name": 100,
+        "x": 171,
+        "y": 105,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 54,
+        "x": 237,
+        "y": 525
+      }
+    ],
+    "levelId": "08-02",
+    "levelEditedAt": "2026-05-05T17:54:26.919Z",
+    "mirroredFrom": "08-02.json"
+  },
+  "reference_solution": "toggle_gravity when candy_still for 0.3 and candy_near 620,-10,60\ncut_rope 0 when candy_still for 0.3 and candy_near 630,480,60",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:55:27.734Z"
+  },
+  "answer": {
+    "commands": "toggle_gravity when candy_still for 0.3 and candy_near 620,-10,60\ncut_rope 0 when candy_still for 0.3 and candy_near 630,480,60",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "toggle_gravity when candy_still for 0.3 and candy_near 620,-10,60\ncut_rope 0 when candy_still for 0.3 and candy_near 630,480,60",
+  "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-041\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  - 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- 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-042.json

@@ -0,0 +1,180 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-042",
+  "level_file": "data/task/rope-042.json",
+  "video": {
+    "path": "data/video/rope-042.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": 4,
+      "sock": 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": 2,
+        "x": 145,
+        "y": 288
+      },
+      {
+        "name": 56,
+        "x": 400,
+        "y": 387,
+        "group": 0,
+        "angle": -120
+      },
+      {
+        "name": 56,
+        "x": 194,
+        "y": 123,
+        "group": 0,
+        "angle": 0
+      },
+      {
+        "name": 52,
+        "x": 318,
+        "y": 134
+      },
+      {
+        "name": 100,
+        "x": 180,
+        "y": 197,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 467,
+        "y": 34,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 178,
+        "y": 33,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 465,
+        "y": 198,
+        "length": 110,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 56,
+        "x": 245,
+        "y": 390,
+        "group": 1,
+        "angle": -55
+      },
+      {
+        "name": 56,
+        "x": 396,
+        "y": 123,
+        "group": 1,
+        "angle": 180
+      },
+      {
+        "name": 3,
+        "x": 317,
+        "y": 302,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 385,
+        "y": 335,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 257,
+        "y": 334,
+        "timeout": -1
+      },
+      {
+        "name": 54,
+        "x": 318,
+        "y": 305
+      }
+    ],
+    "levelId": "04-24",
+    "levelEditedAt": "2026-05-05T15:50:13.741Z"
+  },
+  "reference_solution": "cut_rope 1,3\npop_bubble 0 when candy_near 900,500,40",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:50:20.475Z"
+  },
+  "answer": {
+    "commands": "cut_rope 1,3\npop_bubble 0 when candy_near 900,500,40",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 1,3\npop_bubble 0 when candy_near 900,500,40",
+  "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-042\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  - 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-043.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-043",
+  "level_file": "data/task/rope-043.json",
+  "video": {
+    "path": "data/video/rope-043.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": 158,
+        "y": 424
+      },
+      {
+        "name": 51,
+        "x": 203,
+        "y": 204
+      },
+      {
+        "name": 50,
+        "x": 81,
+        "y": 207
+      },
+      {
+        "name": 3,
+        "x": 196,
+        "y": 292,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 66,
+        "y": 298,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 140,
+        "y": 360,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 83,
+        "y": 72,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 210,
+        "y": 78,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 149,
+        "y": 140,
+        "length": 136,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 152,
+        "y": 143,
+        "length": 154,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "generatorTemplate": "split_drop",
+    "mirroredFrom": "18-05.json"
+  },
+  "reference_solution": "cut_rope 0,1 \ncut_rope 2 when candy_still for 0.3 and candy_near 920,700,60\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 920,700,60\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 920,700,60\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-043\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-044.json

@@ -0,0 +1,129 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-044",
+  "level_file": "data/task/rope-044.json",
+  "video": {
+    "path": "data/video/rope-044.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": 2,
+        "x": 299,
+        "y": 421
+      },
+      {
+        "name": 52,
+        "x": 143,
+        "y": 135
+      },
+      {
+        "name": 3,
+        "x": 139,
+        "y": 290,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 139,
+        "y": 389,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 138,
+        "y": 340,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 72,
+        "y": 38,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 229,
+        "y": 41,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 397,
+        "y": 44,
+        "length": 380,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0,1",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T14:43:10.214Z"
+  },
+  "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-044\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-045.json

@@ -0,0 +1,193 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-045",
+  "level_file": "data/task/rope-045.json",
+  "video": {
+    "path": "data/video/rope-045.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,
+      "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": 3,
+        "x": 159,
+        "y": 306,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 161,
+        "y": 308,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 162,
+        "y": 308,
+        "timeout": -1
+      },
+      {
+        "name": 52,
+        "x": 53,
+        "y": 128
+      },
+      {
+        "name": 100,
+        "x": 56,
+        "y": 46,
+        "length": 50,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      },
+      {
+        "name": 2,
+        "x": 226,
+        "y": 334
+      },
+      {
+        "name": 100,
+        "x": 256,
+        "y": 45,
+        "length": 260,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L",
+        "hidePath": false
+      }
+    ],
+    "ru": [
+      {
+        "name": 8,
+        "x": 126,
+        "y": 401,
+        "locale": "ru",
+        "angle": 90,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 146,
+        "y": 392,
+        "locale": "ru",
+        "text": "Чтобы повернуть пилу, нажмите кнопку",
+        "width": 150
+      }
+    ],
+    "w": [
+      {
+        "name": 4,
+        "x": 143,
+        "y": 393,
+        "locale": "en",
+        "text": "Turn the blades with this button",
+        "width": 150
+      },
+      {
+        "name": 8,
+        "x": 126,
+        "y": 401,
+        "locale": "en",
+        "angle": 90,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "de": [
+      {
+        "name": 4,
+        "x": 150,
+        "y": 389,
+        "locale": "de",
+        "text": "Drehe die Sägeblätter mit diesem Schaltfeld",
+        "width": 180
+      },
+      {
+        "name": 8,
+        "x": 126,
+        "y": 403,
+        "locale": "de",
+        "angle": 90,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ],
+    "fr": [
+      {
+        "name": 4,
+        "x": 146,
+        "y": 391,
+        "locale": "fr",
+        "text": "Faites tourner les lames avec ce bouton",
+        "width": 150
+      },
+      {
+        "name": 8,
+        "x": 128,
+        "y": 407,
+        "locale": "fr",
+        "angle": 90,
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T18:23:19.317Z"
+  },
+  "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-045\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-046.json

@@ -0,0 +1,170 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-046",
+  "level_file": "data/task/rope-046.json",
+  "video": {
+    "path": "data/video/rope-046.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": 6,
+      "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": 155,
+        "y": 210
+      },
+      {
+        "name": 100,
+        "x": 163,
+        "y": 37,
+        "length": 50,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 52,
+        "x": 157,
+        "y": 129
+      },
+      {
+        "name": 54,
+        "x": 244,
+        "y": 394
+      },
+      {
+        "name": 100,
+        "x": 159,
+        "y": 326,
+        "length": 100,
+        "wheel": false,
+        "radius": 60,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 258,
+        "y": 37,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 68,
+        "y": 35,
+        "length": 100,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 72,
+        "y": 167,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 242,
+        "y": 290,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 160,
+        "y": 417,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 291,
+        "y": 257,
+        "length": 180,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 100,
+        "x": 32,
+        "y": 255,
+        "length": 180,
+        "wheel": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      }
+    ]
+  },
+  "reference_solution": "cut_rope 0 when candy_y < 280\ncut_rope 2\ncut_rope 3 when candy_still for 0.3 and candy_near 810,380,60\ncut_rope 5 when candy_y > 780",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T15:39:53.262Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0 when candy_y < 280\ncut_rope 2\ncut_rope 3 when candy_still for 0.3 and candy_near 810,380,60\ncut_rope 5 when candy_y > 780",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0 when candy_y < 280\ncut_rope 2\ncut_rope 3 when candy_still for 0.3 and candy_near 810,380,60\ncut_rope 5 when candy_y > 780",
+  "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-046\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  - 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-047.json

@@ -0,0 +1,157 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-047",
+  "level_file": "data/task/rope-047.json",
+  "video": {
+    "path": "data/video/rope-047.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": 4,
+      "left_candy": 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": 50,
+        "x": 294,
+        "y": 222
+      },
+      {
+        "name": 100,
+        "x": 215,
+        "y": 136,
+        "length": 80,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 2,
+        "x": 429,
+        "y": 434
+      },
+      {
+        "name": 100,
+        "x": 366,
+        "y": 137,
+        "length": 80,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 51,
+        "x": 436,
+        "y": 228
+      },
+      {
+        "name": 54,
+        "x": 431,
+        "y": 223
+      },
+      {
+        "name": 3,
+        "x": 428,
+        "y": 311,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 430,
+        "y": 81,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 294,
+        "y": 332,
+        "length": 90,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "L"
+      },
+      {
+        "name": 3,
+        "x": 428,
+        "y": 283,
+        "timeout": -1
+      },
+      {
+        "name": 100,
+        "x": 431,
+        "y": 354,
+        "length": 100,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      }
+    ],
+    "levelId": "05-24",
+    "levelEditedAt": "2026-05-05T16:52:20.225Z"
+  },
+  "reference_solution": "cut_rope 0,2 \ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_still for 0.3 \npop_bubble 0 when candy_y < 230",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T17:04:52.358Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,2 \ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_still for 0.3 \npop_bubble 0 when candy_y < 230",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,2 \ncut_rope 1 when candy_still for 0.3 \ncut_rope 3 when candy_still for 0.3 \npop_bubble 0 when candy_y < 230",
+  "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-047\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-048.json

@@ -0,0 +1,211 @@
+{
+  "schema_version": 1,
+  "level_id": "rope-048",
+  "level_file": "data/task/rope-048.json",
+  "video": {
+    "path": "data/video/rope-048.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,
+      "object_10": 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": 52,
+        "x": 82,
+        "y": 178
+      },
+      {
+        "name": 100,
+        "x": 157,
+        "y": 101,
+        "length": 80,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 100,
+        "x": 24,
+        "y": 98,
+        "length": 70,
+        "wheel": false,
+        "gun": false,
+        "radius": -1,
+        "moveLength": -1,
+        "moveVertical": false,
+        "moveOffset": 0,
+        "spider": false,
+        "part": "R"
+      },
+      {
+        "name": 3,
+        "x": 182,
+        "y": 435,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 67,
+        "y": 118,
+        "timeout": -1
+      },
+      {
+        "name": 3,
+        "x": 67,
+        "y": 336,
+        "timeout": -1
+      },
+      {
+        "name": 2,
+        "x": 245,
+        "y": 421
+      },
+      {
+        "name": 54,
+        "x": 68,
+        "y": 228
+      },
+      {
+        "name": 56,
+        "x": 61,
+        "y": 390,
+        "group": 0,
+        "angle": 270
+      },
+      {
+        "name": 56,
+        "x": 162,
+        "y": 428,
+        "group": 0,
+        "angle": 0
+      },
+      {
+        "name": 10,
+        "x": 208,
+        "y": 256,
+        "locale": "en",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 282,
+        "y": 273,
+        "locale": "en",
+        "text": "Keep the candy away from spikes",
+        "width": 150
+      }
+    ],
+    "ru": [
+      {
+        "name": 10,
+        "x": 112,
+        "y": 256,
+        "locale": "ru",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 14,
+        "y": 269,
+        "locale": "ru",
+        "text": "Не дайте леденцу разбиться о шипы",
+        "width": 200
+      }
+    ],
+    "fr": [
+      {
+        "name": 10,
+        "x": 111,
+        "y": 257,
+        "locale": "fr",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 34,
+        "y": 268,
+        "locale": "fr",
+        "text": "Garde le bonbon loin des pointes",
+        "width": 150
+      }
+    ],
+    "de": [
+      {
+        "name": 10,
+        "x": 112,
+        "y": 256,
+        "locale": "de",
+        "moveSpeed": 100,
+        "rotateSpeed": 100
+      },
+      {
+        "name": 4,
+        "x": 26,
+        "y": 273,
+        "locale": "de",
+        "text": "Pass auf, dass der Bonbon nicht in die Nähe der Spikes kommt",
+        "width": 250
+      }
+    ],
+    "levelId": "00-19",
+    "levelEditedAt": "2026-05-05T10:29:46.378Z",
+    "mirroredFrom": "00-19.json"
+  },
+  "reference_solution": "cut_rope 0,1 \npop_bubble 0 when candy_y < 290",
+  "reference": {
+    "won": true,
+    "stars": 3,
+    "updated_at": "2026-05-05T10:30:04.840Z"
+  },
+  "answer": {
+    "commands": "cut_rope 0,1 \npop_bubble 0 when candy_y < 290",
+    "reason": "reference solution",
+    "confidence": 1.0
+  },
+  "legacy_answer": "cut_rope 0,1 \npop_bubble 0 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-048\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."
+  }
+}