OK, but in that scenario the prompter doesn't understand the output and can't judge whether the LLM is going off the rails or not. As we know from coding, this

OK, but in that scenario the prompter doesn't understand the output and can't judge whether the LLM is going off the rails or not. As we know from coding, this is a vital step in the process. Experienced coders produce better results, faster, from LLMs than inexperienced coders. I'm not a mathematician so I can't judge the results (which is kinda my point).