They can guarantee that certain small changes won't create a big change in performance (provably, verified through tests, or trivially). An LLM can't.