{
  "scope": "Matched 277-case engine v2 runs. Program-evidence baseline separates classes by construction; results are an upper bound, not a production estimate.",
  "outcome_meanings": {
    "true_positive": "called attack",
    "false_positive": "called harmless",
    "needs_context": "sent to a person"
  },
  "runs": [
    {
      "model": "qwen3.8-27b",
      "run_file": "rec2-qwen3.8-27b-D.json",
      "sha256": "ec657877ed1af9497a75c2c29045b0c1ff60c626e3e36ffeeabe91341d964556",
      "cases": 277,
      "outcomes": {
        "attack": {
          "true_positive": 23,
          "false_positive": 0,
          "needs_context": 3
        },
        "benign": {
          "true_positive": 0,
          "false_positive": 204,
          "needs_context": 47
        }
      },
      "errors": 0
    },
    {
      "model": "fdtn-ai/foundation-sec-1.1-8b-instruct",
      "run_file": "rec2-D-program-recipes.json",
      "sha256": "df6e6ae12a1f853c02c8bac8622d3da499c0bc13ff8b21cf0a5edd1aafb34124",
      "cases": 277,
      "outcomes": {
        "attack": {
          "true_positive": 23,
          "false_positive": 0,
          "needs_context": 3
        },
        "benign": {
          "true_positive": 0,
          "false_positive": 64,
          "needs_context": 187
        }
      },
      "errors": 0
    },
    {
      "model": "qwen3.8-27b",
      "run_file": "rec2-qwen3.8-27b-C3.json",
      "sha256": "1386d9f112800d5d499f66771e970577399bf24ceebacd9b56ae0c9aa974a371",
      "cases": 277,
      "outcomes": {
        "attack": {
          "true_positive": 21,
          "false_positive": 0,
          "needs_context": 5
        },
        "benign": {
          "true_positive": 0,
          "false_positive": 208,
          "needs_context": 43
        }
      },
      "errors": 0
    },
    {
      "model": "fdtn-ai/foundation-sec-1.1-8b-instruct",
      "run_file": "rec2-C3-host-level-recipes.json",
      "sha256": "e5cc625132bc02c122531f2dcf05ab3c47e81c45fad4cf9e2fae5035e7ef3097",
      "cases": 277,
      "outcomes": {
        "attack": {
          "true_positive": 23,
          "false_positive": 0,
          "needs_context": 3
        },
        "benign": {
          "true_positive": 0,
          "false_positive": 60,
          "needs_context": 191
        }
      },
      "errors": 0
    }
  ]
}