A CI review bot is told to "check that code comments are accurate." Developers dismiss most of its comment findings, which flag terse wording, stale TODOs, and phrasing preferences. What change most effectively improves precision?

Answer: B — Flag a comment only when its described behavior contradicts what the code does

An explicit, checkable criterion (claimed behavior contradicts actual behavior) narrows the category so style and wording issues fall outside it. Confidence wording (a) and confidence scores (c) don't tell the model which findings are wrong, so precision barely improves. Dropping the check (d) gives up a useful category instead of fixing its criteria.