Skip to main content

Jev — typed decisions

Jev is TypeSafe AI's "System One" model. It is not a chat model.

It takes unstructured state plus a typed question and returns a typed decision — a choice, a score, or a boolean — each with a confidence. No prose, no parsing, no "please respond in JSON" and hoping.

Why it matters for this stack​

Every agent pipeline contains decision points that a chat model handles badly:

DecisionToday, badlyWith a typed decision layer
Route a request to a branchprompt a chat model, parse the answera typed choice with confidence
Score against a rubricLLM-as-judge, noisy and slowa typed score
Verify a RAG answer is groundedanother generation passa typed boolean
Guardrail before a tool callregex, or hopea typed boolean with a threshold

The failure it removes is parse-and-pray: asking a generative model for structure, getting prose back some fraction of the time, and building retry logic around it.

Where it fits your work​

  • LangGraph routing — the conditional edge is a typed choice, not a prompt whose answer you JSON.parse in a try
  • RAG verification — "is this answer supported by the retrieved context" is a boolean with a confidence, which is exactly what you want to threshold on
  • Agent guardrails — a cheap classification before an expensive or irreversible tool call
  • Classification at volume — where a chat model is both overkill and unreliable

Honest caveats​

Worth stating because the ecosystem around it is young:

  • It is a decision layer, not a reasoning layer. It will not plan, write code, or explain itself.
  • The confidence score is only useful if you calibrate it against your own labels. An uncalibrated threshold is a guess with a decimal point.
  • awesome-jev's own README carries the right warning: "Treat same-day bulk submissions with particular care. Volume is not evidence of quality." Several repos published together sharing a scaffold can satisfy every inclusion rule and still be unproven.

Read​