A model you promoted last night is serving and quietly losing money. It's 9am. What do you do?

This is graded as decision-making under uncertainty, not as a rollback how-to; the interviewer wants to see you act on a proxy before you have proof, because waiting for delayed labels means weeks more bleeding. The strongest candidates state the asymmetry out loud: reverting a fine model costs a little, leaving a bad one costs a lot, so the bar for rollback is a credible proxy, not certainty. The trap is treating rollback as automatically safe; the held-back follow-up is 'what if the old model can't serve anymore', and a senior answer has already checked schema compatibility before pulling the trigger.