If I wanted to be a silly pedant, I’d say that Turing machines are language specifications and thus it’s theoretically impossible for an LLM or any program to validate output formats in general.
in _general_ sure, but if you restricted each token to conform to a Kleene-star grammar you should be able to guarantee that you get something that parses according to a context-free grammar
