{
  "schema_version": 1,
  "measurement": "deterministic correctness controls",
  "run_id": "6a3a469a349e449bbf45a3c5769d781a",
  "suite_sha256": "d3cfd1f21f68fb26a2e8ff84dd0ca8aa0714dbf4dc888840228b1b52305698b7",
  "seed": 1729,
  "controls": {
    "oracle": {
      "total": 152,
      "passed": 152,
      "failed": 0,
      "error": 0,
      "skipped": 0,
      "scored": 152,
      "pass_rate": 1.0,
      "valid": true
    },
    "nop": {
      "total": 152,
      "passed": 0,
      "failed": 152,
      "error": 0,
      "skipped": 0,
      "scored": 152,
      "pass_rate": 0.0,
      "valid": true
    },
    "unsigned": {
      "total": 152,
      "passed": 117,
      "failed": 35,
      "error": 0,
      "skipped": 0,
      "scored": 152,
      "pass_rate": 0.7697368421052632,
      "valid": true
    },
    "identity": {
      "total": 152,
      "passed": 0,
      "failed": 152,
      "error": 0,
      "skipped": 0,
      "scored": 152,
      "pass_rate": 0.0,
      "valid": true
    }
  },
  "failure": {
    "id": "negative-d",
    "input": "123D",
    "expected": "-123",
    "actual": "123",
    "status": "failed"
  },
  "model_performance": null,
  "browser_verified": false,
  "project": "Adversarial correctness benchmark",
  "release_profile": "Verified locally; no hosting or browser-interaction claim"
}
