How would you automate the process of trying to make your model "break" or "hallucinate"?

When would you evaluate a model without having a "correct" answer to compare it against? (e.g., checking for tone or politeness).

Guardrails add an extra check. How do you evaluate if the safety benefit of a guardrail outweighs the 200ms latency penalty it adds?

How do you calculate the ROI of a prompt change? If a new prompt is 5% more accurate but 50% more expensive in tokens, how do you decide if it’s worth it?