Run multiple LLMs in parallel and surface the answer they agree on
Consensus answer + per-model breakdown