JevStation

JevStation is the workspace for Jev, TypeSafe's System One model. Jev is built for a job that large language models handle badly: returning a judgement that software can act on.

A language model produces text meant for people to read. When your code needs a decision — is this ticket urgent, how negative is this review, how complex is this request — you have to coerce prose out of the model and parse it back into a value you can branch on. Jev works the other way round. You send a state, ask typed questions, and it returns typed answers directly. There is no text generation to parse and no prose to interpret.

Three question types cover most judgements. Choice asks for one option from a set you define. Score returns a numeric rating on your own scale. Noul returns a yes-or-no answer as a probability, so you can set a threshold instead of accepting a coin flip. Every answer arrives with a probability distribution and a confidence value, so a workflow can gate an action on how sure the model is.

JevStation wraps that contract in a workspace. The playground runs an evaluation in three steps so you can see the shape of the answers before writing any code. Saved threads keep past evaluations, and reusable context lets you pin standing instructions that apply to every run. Credit metering shows exactly what each evaluation costs.

Pricing is simple. You start with 2,000 free credits, no card required. A standard evaluation costs one credit, and an evaluation over 8,000 characters costs three. When the free credits run out you buy a one-time pack — 5,000 credits for $9.90 or 12,000 credits for $15.90 — with no subscription and no monthly commitment. Credits stay valid for twelve months and failed evaluations are refunded automatically. API access uses your own key.

Privacy is the default rather than an option: your threads are never used for training.

Teams use JevStation where a judgement has to be consistent and auditable — triaging support tickets for sentiment, urgency and complexity in a single round trip, scoring reviews and surveys against a rubric, or gating agent actions on confidence before an automated step runs. The documentation covers the request contract, the answer payload and the credit rules, and the blog covers patterns such as model routing, action gating and escalation.

Start free and run an evaluation in the playground.