{
  "schema": "experiment-results/v0",
  "experiment_id": "EXP-001",
  "protocol_freeze": "lab/exp/exp-001-protocol.json",
  "executed_at": "2026-04-12",
  "status": "confirmed",
  "n_runs": 1080,
  "domains": ["refund", "claim", "memory"],

  "primary_result": {
    "induced_runs": {
      "n": 180,
      "output_pass_at_contract_fail": 180,
      "rate": 1.000,
      "contract_pass_at_contract_fail": 0
    },
    "drift_paths": {
      "n": 60,
      "output_plus_operational_pass_at_contract_fail": 60,
      "rate": 1.000
    }
  },

  "preregistered_gates": {
    "G_induced_output_pass":  { "threshold": 0.95, "observed": 1.000, "passes": true },
    "G_drift_composite_pass": { "threshold": 0.80, "observed": 1.000, "passes": true }
  },

  "verdict": "Both gates pass at the maximum observable rate. Induced fast-lane condition produces 100% disagreement between output evaluation and behavior contracts. Drift paths are also operationally clean — skipping a precondition does not produce latency spikes, retries, or errors, so the operational layer adds no signal. Only the behavior layer separates regressions from baselines under this regime.",

  "mechanism_interpretation": "The output is well-formed, the operational metrics are clean, and the failure is invisible to both. The failure mode lives in the *order* of tool calls relative to the consequential commit, not in any field-level or operational signal. This is the empirical evidence that the OPBR category is structurally distinct from output errors.",

  "replication": {
    "cmd": "abw run --experiment EXP-001",
    "code": "https://github.com/raising-agents/agent-behavior-workbench",
    "traces_corpus": "OPBR-Bench v0"
  },

  "cite_as": "Sanchez de la Sierra, A. (2026). EXP-001 — Controlled mechanism. Raising Agents Lab. https://raisingagents.is/lab/exp/exp-001"
}