{
  "model": "Qwen/Qwen3.6-35B-A3B-FP8",
  "api_base": "http://localhost:8000/v1",
  "timestamp": "2026-05-05T15:51:06.920454",
  "total_evaluated": 60,
  "summary": {
    "total_episodes": 60,
    "detection_accuracy": 0.6166666666666667,
    "miss_rate": 0.7857142857142857,
    "false_alarm_rate": 0.03125,
    "precision": 0.8571428571428571,
    "recall": 0.21428571428571427,
    "f1_security": 0.34285714285714286,
    "threat_type_accuracy": 0.5,
    "parse_failure_rate": 0.0,
    "breakdown": {
      "TP": {
        "total": 28,
        "correct": 6,
        "accuracy": 0.21428571428571427
      },
      "FP": {
        "total": 28,
        "correct": 28,
        "accuracy": 1.0
      },
      "TN": {
        "total": 4,
        "correct": 3,
        "accuracy": 0.75
      }
    },
    "per_sq": {
      "SQ1": {
        "total": 13,
        "accuracy": 0.5384615384615384,
        "miss_count": 5,
        "false_alarm_count": 1
      },
      "SQ2": {
        "total": 6,
        "accuracy": 0.6666666666666666,
        "miss_count": 2,
        "false_alarm_count": 0
      },
      "SQ3": {
        "total": 17,
        "accuracy": 0.6470588235294118,
        "miss_count": 6,
        "false_alarm_count": 0
      },
      "SQ4": {
        "total": 17,
        "accuracy": 0.5882352941176471,
        "miss_count": 7,
        "false_alarm_count": 0
      },
      "SQ5": {
        "total": 7,
        "accuracy": 0.7142857142857143,
        "miss_count": 2,
        "false_alarm_count": 0
      }
    }
  },
  "errors": {
    "total_errors": 26,
    "error_distribution": {
      "MISS": 22,
      "FALSE_ALARM": 1,
      "WRONG_TYPE": 3
    },
    "error_by_sq": {
      "SQ1": {
        "MISS": 5,
        "FALSE_ALARM": 1
      },
      "SQ2": {
        "MISS": 2
      },
      "SQ3": {
        "MISS": 6,
        "WRONG_TYPE": 2
      },
      "SQ4": {
        "MISS": 7,
        "WRONG_TYPE": 1
      },
      "SQ5": {
        "MISS": 2
      }
    },
    "error_by_category": {
      "device_fault": 5,
      "intrusion": 4,
      "water_damage": 1,
      "behavioral_anomaly": 4,
      "child_specific": 4,
      "elderly_specific": 4,
      "fire_gas": 3
    }
  },
  "pipeline": "EGPv2.1"
}