{
  "model": "Qwen/Qwen3.6-35B-A3B-FP8",
  "api_base": "http://localhost:8000/v1",
  "timestamp": "2026-05-05T16:55:43.552603",
  "total_evaluated": 60,
  "summary": {
    "total_episodes": 60,
    "detection_accuracy": 0.6,
    "miss_rate": 0.2857142857142857,
    "false_alarm_rate": 0.5,
    "precision": 0.5555555555555556,
    "recall": 0.7142857142857143,
    "f1_security": 0.6250000000000001,
    "threat_type_accuracy": 0.3,
    "parse_failure_rate": 0.0,
    "breakdown": {
      "TP": {
        "total": 28,
        "correct": 20,
        "accuracy": 0.7142857142857143
      },
      "FP": {
        "total": 28,
        "correct": 15,
        "accuracy": 0.5357142857142857
      },
      "TN": {
        "total": 4,
        "correct": 1,
        "accuracy": 0.25
      }
    },
    "per_sq": {
      "SQ1": {
        "total": 13,
        "accuracy": 0.6923076923076923,
        "miss_count": 2,
        "false_alarm_count": 2
      },
      "SQ2": {
        "total": 6,
        "accuracy": 0.6666666666666666,
        "miss_count": 1,
        "false_alarm_count": 1
      },
      "SQ3": {
        "total": 17,
        "accuracy": 0.35294117647058826,
        "miss_count": 4,
        "false_alarm_count": 7
      },
      "SQ4": {
        "total": 17,
        "accuracy": 0.6470588235294118,
        "miss_count": 1,
        "false_alarm_count": 5
      },
      "SQ5": {
        "total": 7,
        "accuracy": 0.8571428571428571,
        "miss_count": 0,
        "false_alarm_count": 1
      }
    }
  },
  "errors": {
    "total_errors": 38,
    "error_distribution": {
      "FALSE_ALARM": 16,
      "MISS": 8,
      "WRONG_TYPE": 14
    },
    "error_by_sq": {
      "SQ1": {
        "FALSE_ALARM": 2,
        "MISS": 2
      },
      "SQ2": {
        "FALSE_ALARM": 1,
        "MISS": 1,
        "WRONG_TYPE": 1
      },
      "SQ3": {
        "FALSE_ALARM": 7,
        "MISS": 4,
        "WRONG_TYPE": 3
      },
      "SQ4": {
        "FALSE_ALARM": 5,
        "WRONG_TYPE": 7,
        "MISS": 1
      },
      "SQ5": {
        "FALSE_ALARM": 1,
        "WRONG_TYPE": 3
      }
    },
    "error_by_category": {
      "device_fault": 2,
      "intrusion": 4,
      "water_damage": 2,
      "behavioral_anomaly": 4,
      "child_specific": 4,
      "elderly_specific": 3,
      "fire_gas": 3
    }
  },
  "pipeline": "EGPv3"
}