How /consensus works 3 stages per round, repeats up to 5 rounds until the models agree STAGE 1 Parallel verdicts Claude posts a blind verdict, then GPT + Gemini + Grok review in parallel: APPR / RC / REJECT. STAGE 2 Blind cross-review Each reviewer rates the OTHER 2 anonymized answers as viable or not-viable. (conditional - fires on disagreement) STAGE 3 Arbiter adjudicates Claude reconciles everything: accept / dismiss / defer each issue, revises plan. Converged? Stop. Else next round. If not converged, loop back - up to 5 rounds (hard cap) Stage 2 cross-review pattern adapted from karpathy/llm-council Click the detailed version for bias guards and per-model flow.