{
  "model": "Qwen/Qwen3.6-35B-A3B-FP8",
  "api_base": "http://localhost:8000/v1",
  "timestamp": "2026-05-04T21:03:36.738976",
  "total_evaluated": 60,
  "summary": {
    "total_episodes": 60,
    "detection_accuracy": 0.5666666666666667,
    "miss_rate": 0.21428571428571427,
    "false_alarm_rate": 0.625,
    "precision": 0.5238095238095238,
    "recall": 0.7857142857142857,
    "f1_security": 0.6285714285714286,
    "threat_type_accuracy": 0.4090909090909091,
    "parse_failure_rate": 0.0,
    "breakdown": {
      "TP": {
        "total": 28,
        "correct": 22,
        "accuracy": 0.7857142857142857
      },
      "FP": {
        "total": 28,
        "correct": 10,
        "accuracy": 0.35714285714285715
      },
      "TN": {
        "total": 4,
        "correct": 2,
        "accuracy": 0.5
      }
    },
    "per_sq": {
      "SQ1": {
        "total": 13,
        "accuracy": 0.6923076923076923,
        "miss_count": 3,
        "false_alarm_count": 1
      },
      "SQ2": {
        "total": 6,
        "accuracy": 0.6666666666666666,
        "miss_count": 0,
        "false_alarm_count": 2
      },
      "SQ3": {
        "total": 17,
        "accuracy": 0.5294117647058824,
        "miss_count": 1,
        "false_alarm_count": 7
      },
      "SQ4": {
        "total": 17,
        "accuracy": 0.47058823529411764,
        "miss_count": 1,
        "false_alarm_count": 8
      },
      "SQ5": {
        "total": 7,
        "accuracy": 0.5714285714285714,
        "miss_count": 1,
        "false_alarm_count": 2
      }
    }
  },
  "errors": {
    "total_errors": 39,
    "error_distribution": {
      "FALSE_ALARM": 20,
      "MISS": 6,
      "WRONG_TYPE": 13
    },
    "error_by_sq": {
      "SQ1": {
        "FALSE_ALARM": 1,
        "MISS": 3
      },
      "SQ2": {
        "FALSE_ALARM": 2,
        "WRONG_TYPE": 2
      },
      "SQ3": {
        "FALSE_ALARM": 7,
        "WRONG_TYPE": 5,
        "MISS": 1
      },
      "SQ4": {
        "FALSE_ALARM": 8,
        "WRONG_TYPE": 5,
        "MISS": 1
      },
      "SQ5": {
        "FALSE_ALARM": 2,
        "WRONG_TYPE": 1,
        "MISS": 1
      }
    },
    "error_by_category": {
      "device_fault": 3,
      "intrusion": 3,
      "water_damage": 2,
      "behavioral_anomaly": 4,
      "child_specific": 4,
      "fire_gas": 2,
      "elderly_specific": 1
    }
  },
  "pipeline": "EGPv2"
}