{
  "schema_version": 2,
  "manifest_ref": "agent-research:certified-selective-eval",
  "method": "certified-selective-eval",
  "dataset": "deterministic public mechanism mini-suite",
  "seeds": [
    42,
    43,
    44
  ],
  "diagnostic_only": true,
  "seed_results": [
    {
      "seed": 42,
      "upper_error": 0.0,
      "certified": true,
      "tasks": 3
    },
    {
      "seed": 43,
      "upper_error": 0.5,
      "certified": true,
      "tasks": 3
    },
    {
      "seed": 44,
      "upper_error": 0.0,
      "certified": true,
      "tasks": 3
    }
  ],
  "aggregate_metrics": {
    "seed_mean": 43.0,
    "upper_error_mean": 0.16666666666666666,
    "tasks_mean": 3.0
  },
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "formal_comparison": false,
    "diagnostic_only": true,
    "claim_policy": "mechanism execution and invariants only",
    "seeds": [
      42,
      43,
      44
    ]
  },
  "provenance": {
    "artifact_path": "docs/agent-research/2609.34320-certified-selective-eval/metrics/mechanism-seeds42-44.json",
    "dataset_fingerprint": "sep30-closure-public-fixtures-v1",
    "original_code_commit": "working tree"
  }
}