Not really because trust was never how we ensured software quality in the first place. What matters is the process, doing things like code reviews, having a good test harness, coming up with specifications, doing manual testing, and so on. LLMs don't really change anything here.

I'd also love to hear how you actually use LLMs and what your workflow is, because they absolutely do get closer when you set up constraints correctly and run them in the agentic loop in the harness. I also don't know what you refer to by puzzle generator here, clearly you mean something simpler than I assumed. If you want to write out what the specific requirements are, I can show you what Qwen 3.8 comes up with on my machine for them.