{
  "schema_version": 2,
  "manifest_ref": "post-training:pto",
  "algorithm": "pto",
  "seed": 42,
  "dataset": "arithmetic-smoke",
  "baseline": {
    "accuracy": 0.1953125,
    "mean_reward": 0.3596891,
    "entropy": 1.7917595,
    "kl_from_reference": 0.0
  },
  "final": {
    "accuracy": 0.4453125,
    "mean_reward": 0.5820224,
    "entropy": 1.767852,
    "kl_from_reference": 0.0239075
  },
  "training": {
    "steps": 120,
    "last_diagnostics": {
      "preference_tree_nodes": 12.0,
      "lookahead_depth": 1.0,
      "oracle_comparisons": 6.0
    }
  },
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single-seed mini-suite"
  },
  "provenance": {
    "artifact_path": "docs/post-training/2608.12062-pto/metrics/arithmetic-smoke-seed42.json",
    "original_code_commit": "working tree",
    "dataset_fingerprint": "deterministic built-in mini-suite"
  }
}