Learning Jev: a model that returns judgments, not text

A model that answers with a typed value and a probability instead of a paragraph.

I've been reading up on Jev, TypeSafe's first "System One" model, and it's a different shape from the LLMs I'm used to. You don't ask it to write or reason. You hand it state (JSON: the ticket, the document, the candidate list), a question, and the allowed answers, and it returns one of them with a calibrated probability. No prose, no reasoning trace, fast enough to call many times per request. Code keeps the workflow; the model supplies the semantic judgment ordinary code can't make.

Three primitives:
- Choice: pick one of a defined set, with a distribution over the options
- Noul: probability that a condition holds (a yes/no with a number attached)
- Score: a position on ordered levels you define, probability-weighted

Independent questions over the same state run in parallel, so you can ask ten narrow questions at once instead of one big prompt.

Where it seems to land in practice:
- Routing: pick the handler for a request and fill its typed arguments (support triage, intent routing)
- Select-instead-of-generate: find candidate values in code, let the model pick the right one (invoice and form extraction, structure recovery from messy text)
- Reranking and hierarchical classification for search and RAG
- Composite scoring: score dimensions once, let code own the weights and thresholds (lead scoring, content ranking)
- Verification and escalation: check a claim against its evidence, send low-confidence cases to a human or a reasoning model (citation checks, extraction cascades)
- Agent guardrails: bounded next-step selection over changing state

The bit I find most useful: the output is typed so the interface is guaranteed, but the docs are blunt that typed doesn't mean true. You still validate calibration on your own data.

I haven't shipped anything on it yet. If you have, what did you put it behind, and where did the probabilities mislead you?

Docs: https://docs.typesafe.ai