consensuslisted

# Consensus - third-model arbitration Two models disagreed and neither rebuttal landed. Before a human has to arbitrate blind, run ONE stance-steered debate through a third vendor and synthesize the result. This exists for genuinely contested calls - if you can settle the question with a test, run the test instead. ## Procedure 1. **Check availability.** The `mcp__pal__consensus` tool must be present (pal-mcp-server installed and a Gemini key configured). If it is not, say so in one line and stop - do not simulate a third opinion yourself. 2. **State the question neutrally.** One contested item per invocation. Strip any hint of which model held which position; include the concrete context both sides agreed on (the code, the failure scenario, the constraint) in the prompt. 3. **Run one round.** Call `mcp__pal__consensus` with two steered stances - one arguing FOR the proposition, one AGAINST - and the neutral question. ONE round only: no follow-ups, no re-rolls until it agrees with you. 4. **Synthesize a verdict.** One paragraph: what the arbiter concluded, the single strongest argument on each side, and a concrete recommendation (accept / reject / needs-a-test) with how confident the arbiter was. Label it clearly as a third-model recommendation for the human, not ground truth - it is never auto-applied. ## Guardrails - Never use this to GENERATE findings or designs - it arbitrates one existing disagreement, nothing else. - One round,