An eval set is quietly mislabeled. Choose how deep to audit before you ship.
1. Pick the suspicion level
400 candidate rows. The suspected contamination rate sets what
auditing is worth. This round: .
2. Choose the approach
Trust ships blind. The entropy lens flags rows whose loss sits
in the outlier band and sometimes drags clean rows with it. The full sweep
reads everything and still misses about one row in ten.