{
  "schema": "anvil-serving.benchmark-summary/v1",
  "verdict": "no improvement established under acceptance rule",
  "evidence_labels": [
    "functional",
    "historical-invalid"
  ],
  "decision": "no-promotion",
  "retained_default": 4096,
  "task_runs": 24,
  "valid_runs": 16,
  "invalid_runs": 8,
  "promotion_authorized": false,
  "promoted": false,
  "arms": {
    "A": {
      "completion_cap": 4096,
      "valid_tasks": 8,
      "successful_tasks": 6,
      "truncations": 2,
      "empty_finals": 2,
      "unexpected_tool_failures": 0,
      "expected_transient_errors": 2,
      "total_task_seconds": 77.9515251670091
    },
    "B": {
      "completion_cap": 16384,
      "valid_tasks": 8,
      "successful_tasks": 7,
      "truncations": 1,
      "empty_finals": 1,
      "unexpected_tool_failures": 0,
      "expected_transient_errors": 2,
      "total_task_seconds": 143.67185614799382
    }
  },
  "limitations": [
    "Coding coverage invalid; no live rerun beyond 24 tasks.",
    "B did not rescue the cap-related reasoning fixture in both repetitions.",
    "Timings are diagnostic: warm/cache state uncontrolled.",
    "No new full 393216-token or concurrency qualification."
  ]
}