evaluation/thresholds.json

{
  "status": "fitted_internal_only",
  "method": "group-weighted balanced error threshold search; nested profiles",
  "fit_rows": 828,
  "fit_groups": 638,
  "thresholds": {
    "light": {
      "harassment": null,
      "harassment_threatening": 0.12,
      "hate": 0.26,
      "hate_threatening": 0.01,
      "self_harm": null,
      "self_harm_instructions": 0.23,
      "self_harm_intent": null,
      "sexual": null,
      "sexual_minors": 0.02,
      "violence": null,
      "violence_graphic": 0.02
    },
    "medium": {
      "harassment": 0.45,
      "harassment_threatening": 0.03,
      "hate": 0.060000000000000005,
      "hate_threatening": 0.01,
      "self_harm": 0.12,
      "self_harm_instructions": 0.01,
      "self_harm_intent": 0.02,
      "sexual": 0.59,
      "sexual_minors": 0.01,
      "violence": 0.09999999999999999,
      "violence_graphic": 0.02
    },
    "high": {
      "harassment": 0.24000000000000002,
      "harassment_threatening": 0.03,
      "hate": 0.05,
      "hate_threatening": 0.01,
      "self_harm": 0.03,
      "self_harm_instructions": 0.01,
      "self_harm_intent": 0.02,
      "sexual": 0.34,
      "sexual_minors": 0.01,
      "violence": 0.09999999999999999,
      "violence_graphic": 0.02
    },
    "corporate": {
      "harassment": 0.24000000000000002,
      "harassment_threatening": 0.01,
      "hate": 0.05,
      "hate_threatening": 0.01,
      "self_harm": 0.02,
      "self_harm_instructions": 0.01,
      "self_harm_intent": 0.01,
      "sexual": 0.29000000000000004,
      "sexual_minors": 0.01,
      "violence": 0.09,
      "violence_graphic": 0.02
    }
  },
  "standard_category_thresholds": {
    "harassment": 0.25,
    "harassment_threatening": 0.02,
    "hate": 0.03,
    "hate_threatening": 0.01,
    "self_harm": 0.01,
    "self_harm_instructions": 0.01,
    "self_harm_intent": 0.01,
    "sexual": 0.11,
    "sexual_minors": 0.01,
    "violence": 0.09,
    "violence_graphic": 0.02
  },
  "probabilities_calibrated": false,
  "checkpoint_sha256": "4218fb73825ff2fbbdd897a4346dd67141b15fc63d1811413929dddb88d408d3",
  "fit_ids_sha256": "852388954817cf5ef3047eae79e8a1480cc60a1311673e439a02fbb9511cca23",
  "policy": "Reference four-profile policy plus finite profanity rule. Research intent checks scored separately.",
  "limitations": "Weak teacher-derived and synthetic labels. No new detector heads were trained."
}