A scheduled sweep of CI history clusters intermittent failures, decides flake vs. real regression, and branches: quarantine or escalate.