I wonder if this can be solved by adding a disconnected, 2nd layer which can answer meta-questions about the bot’s answers, such as “does this answer follow the

I wonder if this can be solved by adding a disconnected, 2nd layer which can answer meta-questions about the bot’s answers, such as “does this answer follow the rules?”, and this can be used to tweak the answers. Perhaps it can also be used to answer questions that require some sort of chain-of-thought, like “what would be the first step to answering the question? What is the next one?”, etc, then compile these “thoughts” to a final answer. Basically you have the LLM that answers questions, and a “supervisor” LLM which can help guide it. Does something like this exist?