Context in. A typed answer out.
Jev AI is TypeSafe AI’s first System One model. You provide a state—the information to evaluate—and define a question with a constrained answer space. Jev returns a decision your application can use directly.
State can be text, a JSON object, or an array of text. The current model accepts text input; it does not process images, audio, or video.
Three ways to ask
- Choice: select a named option. Route a billing question to Billing instead of Technical Support.
- Score: evaluate against ordered criteria. A three-level rubric can return a continuous value such as 1.4, rather than only an integer.
- Noul: evaluate a true-or-false question as a probability. Our interface calls this Yes / No and uses 0.5 as the decision threshold.
Probability and confidence are different
A probability assigns likelihood to a particular outcome. Confidence is a separate field where the primitive provides it. AIJev displays only the fields actually returned; it does not invent a confidence value from the winning probability.
Calibration is measured across groups of predictions. It does not guarantee that any individual prediction is correct. Validate thresholds on your own representative data before automating consequential actions.
Where it fits
Jev can sit between steps of a workflow: choosing an agent tool, routing incoming requests, checking a response, or scoring retrieved context. It is not a replacement for an LLM that drafts text, generates code, or reasons through an open-ended problem.
Read the Jev and LLM comparison, explore practical decision patterns, or start with the API guide.
Primary source
Continue with our practical guide to Choice, Score, and Noulfor a worked example, then learn how to route decisions using confidence.
This overview is based on TypeSafe’s System One documentation. AIJev is an independent resource.