Typed Decisions with Jev
When the output is a decision
A support ticket needs a team, an urgency estimate, and an answer to whether the customer asked for a person. None requires a paragraph of generated prose. The engineering question is whether a model can return those bounded judgments directly, with a contract the application can consume.
Jev, from TypeSafe AI, is a concrete example. The provider calls its architecture System One and describes training for calibrated decisions as RLCD. Those are vendor terms and claims about the model, not established guarantees about your workflow. The transferable idea is to choose a decision interface when the task has a finite answer space.
Three primitives cover different semantics. Choice selects one named option and returns its distribution. Score gives an expected index on an ordered rubric, with probabilities over levels. Noul returns a probability of yes. A generated explanation, open-ended extraction, or long reasoning trace needs another interface.
A typed answer can still be wrong. Choosing the allowed value billing for a technical issue satisfies the schema and misroutes the ticket. Shape validity and semantic accuracy require separate evidence, just as the Structured Outputs lesson separates parsing from business validation.
Everything in these two lessons runs locally. Ticket labels, answer probabilities and timing are synthetic teaching fixtures; no API key or paid request is needed.
Sources and review dates
- TypeSafe: System One and Jev · reviewed 2026-10-11
- Choice contract · reviewed 2026-10-11
- Score contract · reviewed 2026-10-11
- Noul contract · reviewed 2026-10-11
- HTTP API · reviewed 2026-10-11
- JavaScript SDK · reviewed 2026-10-11
- Model versions, pricing and limits · reviewed 2026-10-11
- Parallel question fan-out · reviewed 2026-10-11
- Jev 1.13 limitations · reviewed 2026-10-11
8 components9 connections0:00
Recording…