Content moderation
Classify against a stated policy and retain human appeal.
Give the model the actual policy categories. Treat quoted user content as evidence to classify, never as instructions that can rewrite the policy.
{
"model": "ren-zero-2026-11-12",
"type": "choice",
"situation": "Policy: direct personal threats are prohibited; ordinary criticism is allowed. Comment under review: Your service is slow and I want an explanation.",
"question": "How should this comment be handled under the stated policy?",
"options": [
{"id": "allow", "description": "Clearly permitted criticism without a direct threat"},
{"id": "review", "description": "Meaning, context, or policy applicability is uncertain"},
{"id": "restrict", "description": "Clearly contains a prohibited direct personal threat"}
],
"threshold": 0.95
}Route every ask_human and review decision to a reviewer. For restrictive actions, apply your approval, notice, and appeal policy even if the gate returns act.
Evaluate quotations, counterspeech, sarcasm, dialect differences, and attempts to inject instructions. A confident label does not replace a complete moderation policy or establish that a user violated the law.