Design prompts with explicit criteria to improve precision and reduce false positives

A review policy says to flag a comment only when its behavior claim contradicts reachable code. In round two, reviewer R reports: (1) '# makes three attempts' above a loop with range(3); its only assertion is that a timeout test is absent; (2) '# rejects non-admins' above an endpoint whose viewer-role integration test returns 200 and changes data; (3) an unmodified local naming convention. The team also wants review convergence without losing newly introduced defects. Its later-round policy skips unchanged local-style preferences but keeps newly evidenced behavior defects eligible. Which two policy decisions follow from this evidence? (Choose 2.)