This makes me wonder if there are cases where one would want the LLM to generate a syntactically invalid response (which could be identified as such) rather than guarantee syntactic validity at the potential cost of semantic accuracy.