It would be nice if LLMs could do that without being wrong about what the code does and doesn't do.