{"data":{"kind":"file","path":"README.md","version_id":"q3ne1jsitflf7om9phzjrkvq","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1389,"modified_at":"2026-11-06T05:33:50.326000","content_hash":"d5f6d22482648ad93adf4965dc68caec3fc80ecf04775c6109e4a5013a033bf4"},"entries":[],"content":"# paprika3d-arcpoint-pile\n\n### Overview\n- **Environment ID:** `jessicayjung/paprika3d-arcpoint-pile`\n- **Short description:** one of Arcpoint's ten mouse-only, ARC-style 3D puzzle games by Keeno Corp (made with paprika3d). Nothing explains the rules: the agent must discover them by playing from pixels.\n- **Tags:** computer-use, multimodal, 3d, abstract-reasoning, arc-style, eval, game, browser\n\n### What it tests\nRule discovery from interaction (ARC-AGI-3 style): the agent sees only the frame and acts by clicking, dragging and pressing keys. Graded on game state; seeded and exactly reproducible; reference solution 1, no-op 0.\n\n### Results (blind, frame-only, held-out levels across all ten Arcpoint games)\n| Model | Held-out levels solved |\n|---|---|\n| Claude Opus | 4/30 |\n| Claude Sonnet | 2/30 |\n| Claude Haiku | 0/30 |\n\n### Quickstart\n```bash\nprime env install jessicayjung/paprika3d-arcpoint-pile\nuv run vf-eval paprika3d-arcpoint-pile -n 2\n```\nThis package holds the two public levels (level-1, level-5) only.\n\n### Held-out tasks and the full environment\nHeld-out levels run on hosted scoring only, so they stay uncontaminated: $1,200 per task. The full game environment: $3,000 per game (all ten Arcpoint games: $30,000). Contact founders@keeno.dev.\n\n### Licence\nThe public tasks here are CC BY-NC 4.0 (see LICENSE); everything else is under contract. Credits in CREDITS.md.\n","encoding":"utf-8","truncated":false,"total_bytes":1389},"status":null}