thanks a ton!

constrained decoding (OpenAI-style structured outputs) make models dumber unfortunately - the short+dense version is that simply masking logits is insufficient because if ever a model was assigning probability to an invalid token, the model is by definition confused. you'd be better off erroring IMO