Yes but it takes much longer to trace them. Because the LLM code almost always gravitates toward data blobs and highly dynamic objects and spaghetti that takes a ton of cognitive load to understand what their failure modes are. Even when it does document them.