Jev vs structured outputs
Structured outputs (JSON schema, tool-call arguments, grammar-constrained decoding) make an LLM emit tokens that parse. The model is still generating. You still pay output tokens. You still wait on a decode loop. You may still get a schema-valid object that is semantically wrong.
Jev does not emit the document. You declare questions. You receive typed fields and distributions. TypeSafe’s launch post says they also wrap LLMs in a “System One LLM wrapper” to force structured decisions for comparisons — and that this wrapper is slower and more expensive than giving decisions without generating probabilities as text.
When structured outputs remain the right tool
- You need a record with open string fields (names, emails, free comments).
- You need a reply the user will read.
- You need code or SQL.
Jev can choose among candidate strings you already extracted. It should not invent the strings. That is the jaggedness “Generation” section.
When Jev is the better shape
- Closed-set classification with a distribution.
- Rubric scoring with a legend.
- Yes/no with a calibrated-ish probability.
- Many such judgments on one state.
You can use both: Jev routes or verifies; an LLM fills the open fields.
Sources
Public TypeSafe or adjacent documentation only. No private claims.