{
  "version": "2.0",
  "date": "2026-04-02",
  "datasetVersion": "v2-categorized",
  "sampleCount": 4245,
  "maliciousSamples": 270,
  "benignSamples": 3881,
  "edgeCaseSamples": 94,
  "adapters": {
    "hma-static-pipeline": {
      "adapterId": "hma-static-pipeline",
      "adapterName": "HMA Static Patterns (no NanoMind)",
      "flagged": 153,
      "flagRate": 0.036,
      "truePositives": 138,
      "falsePositives": 1,
      "trueNegatives": 3880,
      "falseNegatives": 132,
      "precision": 0.993,
      "recall": 0.511,
      "f1": 0.675,
      "fpr": 0.0003
    },
    "hma-tme-only": {
      "adapterId": "hma-tme-only",
      "adapterName": "NanoMind TME v0.5.0 (model only)",
      "flagged": 294,
      "flagRate": 0.069,
      "truePositives": 243,
      "falsePositives": 32,
      "trueNegatives": 3849,
      "falseNegatives": 27,
      "precision": 0.884,
      "recall": 0.900,
      "f1": 0.892,
      "fpr": 0.0082
    },
    "hma-pipeline": {
      "adapterId": "hma-pipeline",
      "adapterName": "HMA Full Pipeline (AST + NanoMind v0.5.0)",
      "flagged": 436,
      "flagRate": 0.103,
      "truePositives": 270,
      "falsePositives": 124,
      "trueNegatives": 3757,
      "falseNegatives": 0,
      "precision": 0.685,
      "recall": 1.000,
      "f1": 0.813,
      "fpr": 0.032,
      "avgScanTimeMs": 4692.0
    }
  },
  "dvaa": {
    "totalScenarios": 70,
    "detected": 61,
    "detectionRate": 0.871,
    "missed": 9
  },
  "paperComparison": {
    "paper": "Holzbauer et al., arXiv:2603.16572, March 2026",
    "table2FlagRates": {
      "Socket": 0.0379,
      "Snyk": 0.0769,
      "agent-trust-hub": 0.1376,
      "Cisco Skill Scanner (Skills.sh)": 0.1404,
      "Cisco Skill Scanner (ClawHub)": 0.1674,
      "GPT 5.3-based (Skills.sh)": 0.2728,
      "VirusTotal": 0.362,
      "GPT 5.3-based (ClawHub)": 0.388,
      "OpenClaw Scanner": 0.4193
    },
    "keyFinding": "Only 0.12% consensus across 5 scanners on 27K skills"
  }
}