Mechanically, schema-constrained decoding masks the logits at each step so only tokens the grammar allows can be sampled. The model still generates a token at a time. A JSON object with six fields costs six fields’ worth of sequential decoding, and any confidence figure has to be reconstructed from logprobs afterwards.
A System One model differs at exactly that step. Nothing is generated and then constrained, because the option set you passed in is the answer space, and a calibrated probability over those options comes back as part of the answer. TypeSafe’s public claim is that Jev “Generates all outputs in a single query”.
Almeida’s objection to the older approach, as reported by agentpedia.codes, is that OpenAI-style structured outputs “make models dumber… simply masking logits is insufficient.” No published evaluation backs that up, and no independent benchmark of Jev existed as of 2026-09-18.
The tradeoff runs the other way too. Structured outputs can return free-text fields and arbitrary nesting; Jev returns neither, and its own jaggedness page states that “Jev-1.13 is not trained to generate text.” The comparison is worked through in System One vs LLM, and the generation question in non-autoregressive.