{
  "schema": "experiment-results/v0",
  "experiment_id": "EXP-006",
  "protocol_freeze": "lab/exp/exp-006-protocol.json",
  "executed_at": "2026-05-25",
  "status": "confirmed",
  "type": "confirmatory",
  "n_runs": 540,
  "domains": ["refund", "claim", "memory", "code"],

  "E1_result": {
    "claim": "Output-only evaluation does not separate regression from baseline runs at any operating point on the held-out canonical OPBR distribution.",
    "metric": "AUC of output-only evaluator distinguishing regression vs baseline",
    "value": "approx 0.5 (chance-level)",
    "interpretation": "The output-only ROC curve cannot be operated to separate the two classes."
  },

  "E2_result": {
    "behavior_contract": {
      "F1": 0.978,
      "precision": 0.958,
      "recall": 1.000
    },
    "best_baseline": {
      "name": "rich_output_provenance",
      "F1": 0.412
    },
    "delta_F1_vs_best_baseline": 0.566,
    "mcnemar_holm_corrected": {
      "vs_output_only_field_check": "approx 0",
      "vs_rich_output_provenance":  "approx 0"
    }
  },

  "preregistered_gates": {
    "G_E1_output_indistinguishable": { "threshold": 0.60, "observed": 0.5,   "passes": true },
    "G_E2_contract_F1":              { "threshold": 0.90, "observed": 0.978, "passes": true },
    "G_E2_baseline_delta":           { "threshold": 0.30, "observed": 0.566, "passes": true },
    "G_mcnemar_family":              { "threshold": 0.01, "observed": "approx 0 for every baseline (Holm-corrected)", "passes": true }
  },

  "verdict": "All four preregistered gates pass. E1 and E2 are confirmed under tightened replication conditions (frozen perturbation set, scenario-bound seeds, agent blind to failure-mode lexicon). The Paper 1 effects survive the tightening.",

  "relationship_to_prior_work": {
    "extends": "EXP-004 (Paper 1 Study 5)",
    "tightens": "Frozen perturbation set, scenario-bound seeds, explicit blinding on the registered failure-mode lexicon strings.",
    "delta_from_EXP_004": "Contracts F1 = 0.978 (vs. 0.982 in EXP-004); within-noise replication. The gates pass at lower magnitudes here, which is expected under stricter blinding."
  },

  "replication": {
    "cmd": "abw run --experiment EXP-006",
    "code": "https://github.com/raising-agents/agent-behavior-workbench",
    "traces_corpus": "OPBR-Bench v0",
    "protocol_freeze_path": "PAPER2_PROTOCOL_FREEZE.json"
  },

  "cite_as": "Sanchez de la Sierra, A. (2026). EXP-006 — Confirmatory E1+E2. Raising Agents Lab. https://raisingagents.is/lab/exp/exp-006"
}