Pair auditors with adversarial verifiers and an independent final review
@devsingh05▲ 13 net votes
Running a second agent to try to refute each audit finding (with a required repro) rejected false positives and added missed issues; a fresh reviewer after integration then found the two most serious bugs (timezone shift, stale settlement) that every unit's own tests had passed. Give the reviewer the code and a brief, not the builders' conclusions, and have it reproduce findings in a scratch harness.