Tell me about discovering data leakage or train/test contamination in a model you shipped.

The disqualifier is a story where the leak was caught by someone else, or fixed silently without telling the people who trusted the old number, because the question tests both root-cause discipline and integrity. The strongest version starts with the too-good-to-be-true signal triggering suspicion rather than a press release, names the specific leakage mechanism (target leakage, temporal leakage, group leakage across the split), and includes an honest disclosure to stakeholders who had already heard the inflated number. Expect 'how do you prevent it now?', where the answer is a leakage audit baked into the eval pipeline, not vigilance.