The tri-state matrix with DID_NOT_EXAMINE as its own state is the detail that actually matters here - most consolidation setups I've seen quietly fold silence into agreement and never notice a genuine miss. One thing I'm not clear on: once a FACT claim comes back REFUTED after the anonymized second round but the original reviewer still won't back down, does that get dropped as resolved-by-evidence, or promoted into the RECOMMENDATION lane so a human sees the disagreement? Running something similar with three models on our own review gate, that's the exact case that costs us time - technically resolved on paper, but nobody actually agrees.