LLM Ensembles

If your code-review agents share a base model and a prompt, the majority vote you trust isn't canceling error — it's laundering it into false confidence, and the math says so.