I think what I am saying is tangential to TDD. I am not really even concerned about the ability of LLM to function as desired, and its verification.

I was rather concerned about a broader fundamental question - how does post-hoc guided generation interfere with the potential benefits of instruction-tuning?