Tell me about a model that passed your evals but failed in production. What happened?

The disqualifier is a story where the root cause is never actually found ('we retrained and it got better'), because the interviewer is checking whether you can separate symptom from cause under a live failure. The strongest version names a specific failure class (train/serve skew, label leakage, covariate shift) with the evidence that pointed there, and ends with a prevention that is institutional, a monitor or a gate, not a personal resolution to be more careful. Expect 'how long was it failing before you noticed?', which probes whether you had online monitoring at all or were flying on offline metrics.