{
  "schema_version": 2,
  "manifest_ref": "post-training:roft",
  "method": "roft",
  "dataset": "deterministic public mechanism mini-suite",
  "seeds": [
    42,
    43,
    44
  ],
  "metrics": {
    "masked_ce": {
      "mean": 4.723277568817139,
      "std": 0.3799676184522421
    },
    "retrospection_tokens": {
      "mean": 3.0,
      "std": 0.0
    },
    "context_target_tokens": {
      "mean": 0.0,
      "std": 0.0
    }
  },
  "seed_results": [
    {
      "seed": 42,
      "masked_ce": 5.138889789581299,
      "retrospection_tokens": 3,
      "context_target_tokens": 0
    },
    {
      "seed": 43,
      "masked_ce": 4.810451030731201,
      "retrospection_tokens": 3,
      "context_target_tokens": 0
    },
    {
      "seed": 44,
      "masked_ce": 4.220491886138916,
      "retrospection_tokens": 3,
      "context_target_tokens": 0
    }
  ],
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "diagnostic_only": true,
    "claim_policy": "executable mechanism and invariant checks only"
  },
  "provenance": {
    "artifact_path": "docs/post-training/2609.35741-roft/metrics/mechanism-seeds42-44.json",
    "dataset_fingerprint": "deterministic public mechanism suite sep30-v1",
    "original_code_commit": "working tree"
  }
}
