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

If your retriever returns 5 documents but only 1 was actually related to answering the question, how do you penalize the retriever for the "noise"?

You’ve updated your system prompt to fix a specific bug. How do you ensure this "fix" didn't break 10 other things the model was previously doing correctly?

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