{
  "count": 0,
  "physics_version": "2.0.0",
  "readme": "# PhysicsThinking Training Dataset\n\nGenerated by the PhysicsThinking MCP simulator.\n\n## Overview\n\n- Total samples: 0\n- Unique agents: 0\n- Time range: n/a → n/a\n- Physics version(s): 2.0.0\n- Quality score: 25/100\n\n## Schema (per sample)\n\n- **id** — sample UUID (stable)\n- **tick** — simulation tick at which the action completed\n- **timestamp** — ISO-8601 UTC insertion time\n- **agent_id / agent_name / agent_model / agent_provider** — actor identity\n- **physics_version** — engine version used to produce the transition\n- **simulation_hash** — deterministic hash of the world state after the step\n- **observation** — { position, velocity, orientation?, held_object, objects_nearby[] }\n- **action** — { type, tool, params, tick_start, tick_end }\n- **result** — { success, data?, error }\n- **reward** — scalar; the sum of reward_breakdown components\n- **reward_breakdown** — { measurement_success, discovery, collision, efficiency }\n- **next_observation** — same shape as observation, sampled after the step\n- **metrics** — { energy_delta, momentum_delta, physics_confidence }\n- **metadata** — { environment, seed?, duration, tick_rate }\n- **confidence** — [0,1] compound score (physics validity × success × agent history)\n- **sample_hash** — SHA-256 dedup key over (tick, agent, action, params)\n\nUnits are SI: meters, seconds, kilograms, Newtons. Y-up. Gravity = -9.81 m/s².\n\n## Data Quality\n\n- Mean confidence: n/a\n- Median confidence: n/a\n- Success rate: 0.0%\n- Schema validation pass rate: 0.0%\n- Mean energy drift per step: 0.00e+0\n- Mean momentum drift per step: 0.00e+0\n\n## Action Distribution\n\n- (no actions recorded)\n\n## Usage\n\n- JSON: an array of samples in `samples`; the same bundle carries `readme`, `metrics`, and `schema`.\n- NDJSON: one sample per line. README/metrics available at\n  `/api/public/training/metrics` and `/api/public/training/readme`.\n- CSV: flattened top-level fields; nested JSON is stringified.\n\n## Deduplication\n\nSamples are unique on `sample_hash`. Re-executing the identical action at\nthe same tick with the same agent is a no-op.\n\n## Citation\n\nPlease cite the PhysicsThinking dataset when using it for training or\nevaluation.\n",
  "metrics": {
    "total_samples": 0,
    "unique_agents": 0,
    "time_range": {
      "start": null,
      "end": null
    },
    "confidence": null,
    "success_rate": 0,
    "validation_pass_rate": 0,
    "physics_accuracy": {
      "energy": 0,
      "momentum": 0
    },
    "physics_versions": [],
    "action_type_counts": {},
    "quality_score": 25
  },
  "schema": {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "title": "PhysicsThinking Training Sample",
    "type": "object",
    "required": [
      "id",
      "tick",
      "timestamp",
      "agent_id",
      "physics_version",
      "observation",
      "action",
      "result",
      "reward",
      "next_observation",
      "metrics",
      "metadata",
      "confidence",
      "sample_hash"
    ]
  },
  "samples": []
}