{
  "model": "Qwen/Qwen3.6-35B-A3B-FP8",
  "api_base": "http://localhost:8000/v1",
  "timestamp": "2026-05-05T22:39:45.145700",
  "total_evaluated": 60,
  "summary": {
    "total_episodes": 60,
    "detection_accuracy": 0.5833333333333334,
    "miss_rate": 0.8571428571428571,
    "false_alarm_rate": 0.03125,
    "precision": 0.8,
    "recall": 0.14285714285714285,
    "f1_security": 0.24242424242424243,
    "threat_type_accuracy": 0.75,
    "parse_failure_rate": 0.0,
    "breakdown": {
      "TP": {
        "total": 28,
        "correct": 4,
        "accuracy": 0.14285714285714285
      },
      "FP": {
        "total": 28,
        "correct": 27,
        "accuracy": 0.9642857142857143
      },
      "TN": {
        "total": 4,
        "correct": 4,
        "accuracy": 1.0
      }
    },
    "per_sq": {
      "SQ1": {
        "total": 13,
        "accuracy": 0.6923076923076923,
        "miss_count": 4,
        "false_alarm_count": 0
      },
      "SQ2": {
        "total": 6,
        "accuracy": 0.5,
        "miss_count": 3,
        "false_alarm_count": 0
      },
      "SQ3": {
        "total": 17,
        "accuracy": 0.47058823529411764,
        "miss_count": 8,
        "false_alarm_count": 1
      },
      "SQ4": {
        "total": 17,
        "accuracy": 0.5294117647058824,
        "miss_count": 8,
        "false_alarm_count": 0
      },
      "SQ5": {
        "total": 7,
        "accuracy": 0.8571428571428571,
        "miss_count": 1,
        "false_alarm_count": 0
      }
    }
  },
  "errors": {
    "total_errors": 26,
    "error_distribution": {
      "MISS": 24,
      "FALSE_ALARM": 1,
      "WRONG_TYPE": 1
    },
    "error_by_sq": {
      "SQ1": {
        "MISS": 4
      },
      "SQ2": {
        "MISS": 3
      },
      "SQ3": {
        "FALSE_ALARM": 1,
        "MISS": 8
      },
      "SQ4": {
        "MISS": 8
      },
      "SQ5": {
        "WRONG_TYPE": 1,
        "MISS": 1
      }
    },
    "error_by_category": {
      "device_fault": 4,
      "intrusion": 4,
      "fire_gas": 4,
      "water_damage": 1,
      "behavioral_anomaly": 4,
      "child_specific": 4,
      "elderly_specific": 4
    }
  },
  "pipeline": "EGPv3.2"
}