{
  "title": "DeAIze production-path rewrite observation, 16 September 2026",
  "generated_at": "2026-09-16T22:46:06+0800",
  "source": "docs/benchmark-detectors.json",
  "samples": 6,
  "measurement": "Per-model AI scores before and after a rewrite. Scores are model outputs, not calibrated authorship probabilities.",
  "classification_threshold": null,
  "threshold_note": "No binary pass/fail threshold is used in these before/after comparisons.",
  "model_versions": null,
  "version_note": "The stored report names models but does not record exact model revisions; this is a historical observation, not a fully reproducible current benchmark.",
  "summary": {
    "roberta_openai": {"n": 6, "mean_pre": 0.0009, "mean_post": 0.443, "fell": 1, "rose": 4, "unchanged_score": 1},
    "hc3_roberta": {"n": 6, "mean_pre": 0.1873, "mean_post": 0.5022, "fell": 2, "rose": 4, "unchanged_score": 0},
    "mage": {"n": 6, "mean_pre": 0.6681, "mean_post": 0.833, "fell": 3, "rose": 1, "unchanged_score": 2},
    "fast_detectgpt": {"n": 6, "mean_pre": 0.3063, "mean_post": 0.3061, "fell": 4, "rose": 2, "unchanged_score": 0},
    "perplexity_burstiness": {"n": 6, "mean_pre": 0.0198, "mean_post": 0.0, "fell": 1, "rose": 0, "unchanged_score": 5}
  },
  "limits": [
    "The six generated drafts are not a random sample of customer work.",
    "No human-written control group, reader-quality ratings or binary accuracy measurements are included.",
    "This is not a Turnitin, GPTZero, Copyleaks or other commercial detector evaluation.",
    "The historical rewrite provider and current product behavior may differ.",
    "Neither a lower score nor a semantic similarity check proves accuracy, authorship or institutional acceptance."
  ]
}