1. A model's answer quality drops but latency and error rates look normal. What kind of failure is this?
2. Why is request tracing valuable when debugging an agent incident?
3. After a bad prompt change ships, what is the fastest safe mitigation?
4. An upstream provider silently updates its model and outputs shift. What practice detects this early?
5. What is the purpose of a canary during a model rollout incident review?
6. During an incident, why capture the exact inputs that triggered a failure?
7. A rise in outputs hitting a guardrail block is best interpreted as what signal?