{"data":{"kind":"file","path":"README.md","version_id":"r69gq9qhlkcflai3j9zsstkw","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2212,"modified_at":"2026-08-03T14:07:32.072000","content_hash":"fa7e725bb86b056c9f92a85269e936c8109da019f0bf31f228dbded507274456"},"entries":[],"content":"# Minesweeper Agent\n\nA verifiers environment that evaluates an LLM's ability to play Minesweeper through logical reasoning, risk assessment, and strategic decision-making.\n\n## Overview\n\nThe model is presented with partially-revealed Minesweeper boards of varying difficulty and must:\n1. Analyze the visible board state\n2. Deduce which border cells are safe to click vs. which contain mines\n3. Provide step-by-step logical reasoning\n4. Make strategic recommendations\n\n## Difficulty Levels\n\n| Level | Grid Size | Mines | Description |\n|-------|-----------|-------|-------------|\n| 0 | 6x6 | 4 | Beginner |\n| 1 | 8x8 | 8 | Intermediate |\n| 2 | 9x9 | 12 | Advanced |\n| 3 | 10x10 | 18 | Expert |\n\n## Evaluation Metrics\n\n| Metric | Weight | Description |\n|--------|--------|-------------|\n| Move Accuracy | 0.40 | Correct identification of safe cells and mines |\n| Reasoning Quality | 0.25 | Logical deduction, step-by-step analysis |\n| Format Compliance | 0.15 | Structured output with clear recommendations |\n| Risk Assessment | 0.20 | Probability reasoning and tradeoff analysis |\n\n## Configuration\n\n```python\nimport verifiers as vf\n\nenv = vf.load_environment(\"raal/minesweeper-agent\")\n# or with custom args:\nenv = vf.load_environment(\"raal/minesweeper-agent\", num_examples=50, max_difficulty=3)\n```\n\n### Parameters\n\n- `num_examples` (int, default=30): Number of puzzles to generate\n- `max_difficulty` (int, default=2): Maximum difficulty level (0-3)\n\n## Example Board State\n\n```\n     0  1  2  3  4  5\n    ------------------\n 0 | 1  1  1  .  .  .\n 1 | 1  .  2  1  .  .\n 2 | 1  2  3  .  .  .\n 3 | .  1  .  .  .  .\n 4 | .  1  1  1  .  .\n 5 | .  .  .  1  1  .\n```\n\nThe model must analyze numbered cells' constraints to deduce which unrevealed cells (`.`) are safe to click and which are mines.\n\n## How It Works\n\n1. **Board Generation**: Random mine placement with guaranteed safe first-click zone\n2. **Flood-fill reveal**: Starting area is revealed, creating the puzzle state\n3. **Constraint analysis**: Border cells are classified as definitely safe or definitely mine using basic constraint solving\n4. **Evaluation**: Model output is parsed for coordinate mentions and compared against ground truth\n\n## License\n\nMIT\n","encoding":"utf-8","truncated":false,"total_bytes":2212},"status":null}