Five detectors, one bad merge: why our LLM corruption guard flagged 43% of healthy output

We run a self-hosted ~300B reasoning model in production. It writes macroeconomic desk reports in...