{
  "model": "Qwen/Qwen2.5-7B-Instruct",
  "api_base": "http://localhost:8000/v1",
  "timestamp": "2026-05-01T14:22:32.244924",
  "total_evaluated": 10,
  "avg_latency_seconds": 21.64,
  "api_errors": 0,
  "summary": {
    "total_episodes": 10,
    "detection_accuracy": 0.4,
    "miss_rate": 0,
    "false_alarm_rate": 0.5,
    "precision": 0.0,
    "recall": 0,
    "f1_security": 0,
    "threat_type_accuracy": 0,
    "parse_failure_rate": 0.1,
    "breakdown": {
      "TP": {
        "total": 0,
        "correct": 0,
        "accuracy": 0
      },
      "FP": {
        "total": 10,
        "correct": 4,
        "accuracy": 0.4
      },
      "TN": {
        "total": 0,
        "correct": 0,
        "accuracy": 0
      }
    },
    "per_sq": {
      "SQ1": {
        "total": 10,
        "accuracy": 0.4,
        "miss_count": 0,
        "false_alarm_count": 5
      }
    }
  },
  "errors": {
    "total_errors": 6,
    "error_distribution": {
      "PARSE_FAIL": 1,
      "FALSE_ALARM": 5
    },
    "error_by_sq": {
      "SQ1": {
        "PARSE_FAIL": 1,
        "FALSE_ALARM": 5
      },
      "SQ2": {},
      "SQ3": {},
      "SQ4": {},
      "SQ5": {}
    },
    "error_by_category": {}
  }
}