How to Run an LLM Review Panel: Use It to Check, Not to Write
An LLM review panel, or council, is several models with personas scoring a draft cold. The usual loop is to fix what the critics say and run them again. That loop barely moves the score. Across 270 logged rounds, edits of taste moved it 0.04 points out of 10, error fixes 0.62, and new proof was what reliably moved verdicts. Here is a method that uses the panel for what it does well: catching errors and naming the evidence a reader needs.