A model silently degraded for six weeks before anyone noticed. Walk me through the postmortem.

The structural insight that separates seniors is that traditional postmortems assume a discrete onset, and silent degradation has none, so 'when did it start' is itself a finding about your detection lag, not a timestamp you can look up. Interviewers listen for whether you distinguish detection failure from the underlying defect, because the most important action items usually fix detection, not the model. A weak answer reaches straight for 'retrain'; the strong one first asks whether the degradation was drift, an upstream feature change, a skew bug, or a label-pipeline problem, because retraining only fixes one of those.