When you use large language models to analyze legal documents, one question inevitably comes up: How do you know the output is actually good? In most domains, an imperfect answer is an inconvenience.