What is a common mistake when working with LLM evaluation?

Using only subjective manual testing or a single aggregate score can hide important failure modes.