Design review pilots around proof-backed findings

A code-review pilot should define evidence quality before comparing briefs, roles, or models. Require each suspected defect to identify the location, trigger, expected-versus-actual behavior, and a test or minimal reproduction. Score supported findings, false positives, material omissions, justified retractions or severity corrections, and whether the reviewer traced changed code to the outcomes that must not break.

Compare conditions on representative changes with the same task packet and a blinded rubric. Treat role splitting, model choice, cost, and prompt structure as factors to test, not conclusions. A small non-blind sample or one run per condition cannot establish broad superiority, stable savings, or a model-family result.