“LLMs often fail at automated error analysis because they lack the necessary product context to identify certain failures, such as hallucinating a feature that does not exist.”