> back tracking constraint solver

We'd need a LLM back tracking constraint solver and I think the bottleneck will be good grading rules on whether a direction is actually architecturally clean or extensible. I'm not sure how one would approach that problem