code-reviewlisted

# Code Review The reviewer is a **read-only role** (`references/core/agents.md`). That is not a limitation on thoroughness — it is the property that makes review safe to automate: a reviewer who can also change the code will fix what it thinks is wrong, and a wrong fix applied confidently is worse than a finding raised and rejected. **Reviewers reopen. The human elects.** ## Requesting a review — give the contract, withhold the conclusion Hand over the diff **and what it is supposed to satisfy** (the pin, the criterion, the contract). Do **not** hand over your conclusion that it does. A reviewer told "this correctly implements X" evaluates your sentence, not your code — the agreement it produces is worthless because it was primed. State the claim as a question, or state nothing. Include: the pins in scope, how you verified it, and what you are unsure about. Uncertainty is the most useful thing you can pass on and the easiest thing to omit. Withholding the conclusion is half of it; the other half is on the reviewer. Re-running the tests the author already passed proves nothing an author building *to* those tests could not have staged — so where the stakes justify it, the reviewer **exercises the behavior itself** against the elected criterion, rather than trusting the author's own run. Withholding your conclusion keeps the review honest; verifying the behavior, not the artifact built to pass, keeps it real. ## Giving a review — precedence, first match wins Work in this