{
  "model": "0.3.0",
  "extension": "0.4.1",
  "settings": "Balanced",
  "date": "2026-10-30",
  "records": [
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "legacy-regression",
      "datasetLabel": "Earlier regression",
      "role": "post",
      "primary": true,
      "ai": 5620,
      "human": 24416,
      "aiHidden": 5075,
      "aiVisible": 545,
      "humanHidden": 236,
      "humanVisible": 24180,
      "aiCapture": 0.9030249110320284,
      "aiMissed": 0.09697508896797158,
      "humanFalsePositive": 0.009665792922673657,
      "humanKept": 0.9903342070773263
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "legacy-regression",
      "datasetLabel": "Earlier regression",
      "role": "comment",
      "primary": true,
      "ai": 5620,
      "human": 24416,
      "aiHidden": 4367,
      "aiVisible": 1253,
      "humanHidden": 15,
      "humanVisible": 24401,
      "aiCapture": 0.7770462633451958,
      "aiMissed": 0.22295373665480422,
      "humanFalsePositive": 0.0006143512450851901,
      "humanKept": 0.9993856487549148
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "development-test",
      "datasetLabel": "RAID development",
      "role": "post",
      "primary": true,
      "ai": 2094,
      "human": 2094,
      "aiHidden": 1211,
      "aiVisible": 883,
      "humanHidden": 295,
      "humanVisible": 1799,
      "aiCapture": 0.5783190066857689,
      "aiMissed": 0.42168099331423115,
      "humanFalsePositive": 0.1408787010506208,
      "humanKept": 0.8591212989493792
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "development-test",
      "datasetLabel": "RAID development",
      "role": "comment",
      "primary": true,
      "ai": 2094,
      "human": 2094,
      "aiHidden": 712,
      "aiVisible": 1382,
      "humanHidden": 37,
      "humanVisible": 2057,
      "aiCapture": 0.34001910219675263,
      "aiMissed": 0.6599808978032473,
      "humanFalsePositive": 0.01766953199617956,
      "humanKept": 0.9823304680038204
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "generator-challenge",
      "datasetLabel": "Harder RAID",
      "role": "post",
      "primary": true,
      "ai": 2293,
      "human": 2293,
      "aiHidden": 653,
      "aiVisible": 1640,
      "humanHidden": 323,
      "humanVisible": 1970,
      "aiCapture": 0.2847797645006542,
      "aiMissed": 0.7152202354993458,
      "humanFalsePositive": 0.14086349760139555,
      "humanKept": 0.8591365023986044
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "generator-challenge",
      "datasetLabel": "Harder RAID",
      "role": "comment",
      "primary": true,
      "ai": 2293,
      "human": 2293,
      "aiHidden": 355,
      "aiVisible": 1938,
      "humanHidden": 48,
      "humanVisible": 2245,
      "aiCapture": 0.15481901439162668,
      "aiMissed": 0.8451809856083733,
      "humanFalsePositive": 0.020933275185346708,
      "humanKept": 0.9790667248146533
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "mage-observed/test.csv",
      "datasetLabel": "MAGE standard",
      "role": "post",
      "primary": true,
      "ai": 27957,
      "human": 28594,
      "aiHidden": 5581,
      "aiVisible": 22376,
      "humanHidden": 2903,
      "humanVisible": 25691,
      "aiCapture": 0.19962800014307686,
      "aiMissed": 0.8003719998569232,
      "humanFalsePositive": 0.10152479541162482,
      "humanKept": 0.8984752045883752
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "mage-observed/test.csv",
      "datasetLabel": "MAGE standard",
      "role": "comment",
      "primary": true,
      "ai": 27957,
      "human": 28594,
      "aiHidden": 1660,
      "aiVisible": 26297,
      "humanHidden": 322,
      "humanVisible": 28272,
      "aiCapture": 0.05937690023965375,
      "aiMissed": 0.9406230997603462,
      "humanFalsePositive": 0.011261103728054837,
      "humanKept": 0.9887388962719451
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "mage-observed/test_ood_set_gpt.csv",
      "datasetLabel": "MAGE unfamiliar domains",
      "role": "post",
      "primary": true,
      "ai": 677,
      "human": 562,
      "aiHidden": 270,
      "aiVisible": 407,
      "humanHidden": 51,
      "humanVisible": 511,
      "aiCapture": 0.3988183161004431,
      "aiMissed": 0.601181683899557,
      "humanFalsePositive": 0.09074733096085409,
      "humanKept": 0.9092526690391459
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "mage-observed/test_ood_set_gpt.csv",
      "datasetLabel": "MAGE unfamiliar domains",
      "role": "comment",
      "primary": true,
      "ai": 677,
      "human": 562,
      "aiHidden": 138,
      "aiVisible": 539,
      "humanHidden": 8,
      "humanVisible": 554,
      "aiCapture": 0.2038404726735598,
      "aiMissed": 0.7961595273264401,
      "humanFalsePositive": 0.014234875444839857,
      "humanKept": 0.9857651245551602
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "beemo",
      "datasetLabel": "BEEMO instruction writing",
      "role": "post",
      "primary": true,
      "ai": 2165,
      "human": 2165,
      "aiHidden": 1221,
      "aiVisible": 944,
      "humanHidden": 614,
      "humanVisible": 1551,
      "aiCapture": 0.563972286374134,
      "aiMissed": 0.436027713625866,
      "humanFalsePositive": 0.2836027713625866,
      "humanKept": 0.7163972286374134
    },
    {
      "model": "current",
      "label": "Installed detector (model 0.3.0)",
      "dataset": "beemo",
      "datasetLabel": "BEEMO instruction writing",
      "role": "comment",
      "primary": true,
      "ai": 2165,
      "human": 2165,
      "aiHidden": 816,
      "aiVisible": 1349,
      "humanHidden": 261,
      "humanVisible": 1904,
      "aiCapture": 0.376905311778291,
      "aiMissed": 0.623094688221709,
      "humanFalsePositive": 0.12055427251732101,
      "humanKept": 0.879445727482679
    }
  ]
}