PROTRAILBLAZER

Caveats · where this simplifies

• The model is a seeded toy. Every probability here is illustrative, generated from hand-set logits plus seeded noise, not measured from a real LLM.
• Real tokenizers emit subword pieces. A production grammar engine must track legality across partial lexemes. Each token here is a whole lexeme so the state machine stays readable.
• This grammar fixes key order and a closed value vocabulary. Real JSON Schema allows any key order and open-ended strings, which makes the compiled automaton much larger.
• Masking logits is one implementation. Some systems compile the schema to a token-level automaton ahead of time, others repair or re-ask after generation instead of intervening during sampling.
• Constraints guarantee structure, not truth. A schema-valid document can still contain a wrong or useless value.