[unprovable]

This is cool - Seeing a lot more usefulness being found for things like ASTs, SMT solvers, and other automated reasoning tools within LLM workflows. Things like this are useful as it gives an potential extra layer of confidence. Nice one.

[MuddySheep]

Thanks. Interesting for me is that the LLM and the parser have inverse failure mode, the LLM nails intent and misses structure, and the parser is the opposite.