{
  "schema_version": 2,
  "manifest_ref": "agent-research:tfrbench",
  "adapter_key": "tfrbench",
  "evaluation_tier": "l0_interface",
  "diagnostic_only": true,
  "promotion_eligible": false,
  "evolve_integrated": false,
  "formal_comparison": false,
  "seed": 42,
  "device": "cpu",
  "dataset": {
    "id": "self-authored-protocol-fixtures",
    "public_benchmark": false,
    "revision": "google16-evaluation-fixtures-v1"
  },
  "metrics": {
    "fixture_candidates": 2,
    "fixture_selected_mase": 0.0,
    "fixture_summary_mase": 0.0
  },
  "full_experiment_status": "not_reproduced_high_cost",
  "evaluation_protocol": {
    "tier": "l0_interface",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "bounded component/protocol only, not paper capability"
  },
  "provenance": {
    "artifact_path": "docs/agent-research/2604.05364-tfrbench/metrics/protocol-cpu-seed42.json",
    "code_commit": "6873b033ffa97bee066d08436b16fac53bb42ccc",
    "dataset_fingerprint": "bef370cfe12f2cebda1bc612b7b20cd679a4227b8d2e48e7cd7f3b050f2ba7fb",
    "raw_result_path": "docs/research-audits/google16-components-20261021/evidence/google16-tfrbench-final-seed42.json",
    "raw_result_sha256": "756bbc1b34aeb8a70e36ca896e8c2e639caab72a8c749f4ee63e1d8527263731",
    "source_binding": "curator_verified_frozen_files"
  },
  "fixture_demo": true,
  "model_calls": 0,
  "report": {
    "protocol": "outcome_conditioned_benchmark_construction_not_heldout_evaluation",
    "diagnostic_only": true,
    "promotion_eligible": false,
    "future_events_allowed_in_construction": true,
    "candidates": [
      {
        "reasoning": "continue trend",
        "forecast": [
          [
            5.0
          ],
          [
            6.0
          ]
        ],
        "consistency": 4,
        "mase": 0.0,
        "iteration": 1
      },
      {
        "reasoning": "flat",
        "forecast": [
          [
            4.0
          ],
          [
            4.0
          ]
        ],
        "consistency": 4,
        "mase": 1.5,
        "iteration": 1
      }
    ],
    "selected": {
      "reasoning": "continue trend",
      "forecast": [
        [
          5.0
        ],
        [
          6.0
        ]
      ],
      "consistency": 4,
      "mase": 0.0,
      "iteration": 1
    },
    "reference_accepted": true,
    "summary": "continue trend",
    "summary_accepted": true,
    "summary_mase": 0.0,
    "selection_policy": "explicit_callback_paper_does_not_fix_composite_weights"
  },
  "status": "completed"
}