Read it. For open-ended review I agree: an agent told to find problems finds nitpicks, because there are always nitpicks. But proofs and tests only check the spec you wrote. The bugs that hurt me were a wrong spec: everything green, output unchanged. What still pays is review with one narrow question, "show me the output moved," not "what's wrong here." Does pstack keep anything like that, or is it all empirical now?
-- {no channel}