Question 1

Should Jev replace my LLM or sit in front of it?

Accepted Answer

Use Jev as the instinct layer and the LLM as fallback: Jev answers typed choice, score, and noul questions in about 100ms, and only low-confidence cases go to GPT-4 or Claude. Code then acts on structured JSON instead of parsing prose.

Question 2

What do I actually get to run, not just read?

Accepted Answer

Ten folders covering triage, trading signals, guardrails, DOM navigation, model routing, moderation, healthcare triage, incident routing, contract review, and fraud detection. Each ships a README, a Python script, and a payload.json pasteable into the console.

Question 3

How do the three primitives differ in practice?

Accepted Answer

Choice picks one criterion, score places input on an ordered rubric as a probability-weighted float, and noul returns a boolean probability. Schemas use criteria throughout, and score floats can land between levels rather than snapping to integers.

Question 4

Which pattern fits a guardrail versus a router?

Accepted Answer

Composite scoring for atomic questions combined in code, confidence-gated routing for automate versus escalate, speculative fan-out for parallel next-action evaluation, and a hybrid loop pairing Jev structure with LLM text. THEORY.md adds anti-patterns and calibration notes.

Question 5

Can I try Jev without writing any SDK code?

Accepted Answer

Yes. The quick start uses the TypeSafe console at console.typesafe.ai, which has two input panes for state and questions and a response pane, no SDK needed. For the raw API, the cookbook posts to https://api.typesafe.ai/v1/systemone with the model jev-1.13.0 and a Bearer key.

Question 6

What are the four architectural patterns it documents?

Accepted Answer

Composite scoring breaks a judgment into atomic questions and combines them in code; confidence-gated routing automates high-confidence answers and escalates the rest; speculative fan-out evaluates all next actions in one parallel request; and the hybrid agent loop pairs Jev for structure with an LLM for text.

Question 7

What is the most common schema mistake it warns about?

Accepted Answer

Always use criteria, never options or min/max: Choice takes a criteria dict, Score takes an ordered criteria list of 2 to 10 levels, and Noul takes instructions. A score returns a float, the probability-weighted expected value across levels, not an integer.

Jev Cookbook — paramjeetn

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community