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

Instead of checking for exact words, how would you use BERTScore or Cosine Similarity of embeddings to evaluate if an LLM's summary is accurate?

Why is a standard unit test (asserting that output == "expected") often a bad way to test an LLM? How do you handle a model that gives three different, but correct, answers to the same prompt?

If your model’s accuracy suddenly drops by 10% on Tuesday, how do you determine if the Model changed (API update), the Data changed (new documents in RAG), or User Behavior changed?