Two months before launch, the model is 20% below target on a critical edge case. How do you triage, brief the business, and resolve it?

This screens whether the candidate triages with a diagnostic split (data vs retrieval vs generation vs the target itself) instead of thrashing, whether they communicate a forecast and options rather than false confidence, and whether they understand that 'ship with a guardrail on the edge case' is often the right answer rather than 'close the gap or slip.' The answer that fails jumps straight to prompt-tuning or promises to hit the number. The held-back follow-ups: 'is the target even right' and 'what if you can't close it by launch', and they want a containment plan (route the edge case to a human, scope it out of v1) that lets the launch happen safely.