{
  "total_fixtures": 50,
  "total_correct": 31,
  "accuracy": 0.62,
  "benign_total": 25,
  "benign_pass_through": 25,
  "benign_false_positives_by_category": {
    "context-smuggling": 0,
    "encoding-trick": 0,
    "instruction-override": 0,
    "multi-turn-ramp": 0,
    "prefix-injection": 0,
    "role-play": 0
  },
  "per_category": {
    "context-smuggling": {
      "tp": 3,
      "fp": 0,
      "fn": 5,
      "tn": 67,
      "precision": 1.0,
      "recall": 0.375,
      "f1": 0.5455
    },
    "encoding-trick": {
      "tp": 6,
      "fp": 1,
      "fn": 2,
      "tn": 66,
      "precision": 0.8571,
      "recall": 0.75,
      "f1": 0.8
    },
    "instruction-override": {
      "tp": 6,
      "fp": 2,
      "fn": 2,
      "tn": 65,
      "precision": 0.75,
      "recall": 0.75,
      "f1": 0.75
    },
    "multi-turn-ramp": {
      "tp": 4,
      "fp": 0,
      "fn": 3,
      "tn": 68,
      "precision": 1.0,
      "recall": 0.5714,
      "f1": 0.7273
    },
    "prefix-injection": {
      "tp": 6,
      "fp": 1,
      "fn": 3,
      "tn": 65,
      "precision": 0.8571,
      "recall": 0.6667,
      "f1": 0.75
    },
    "role-play": {
      "tp": 6,
      "fp": 1,
      "fn": 4,
      "tn": 64,
      "precision": 0.8571,
      "recall": 0.6,
      "f1": 0.7059
    }
  }
}