Should follow-up questions go in the same call or a second one?

Put them in the same call when they are independent, since one answer never affects another in a call. Make a second call only when its questions depend on an answer from the first.

How should I read a Noul answer sitting near 0.5?

It means Jev cannot tell, not medium intensity. Use a score question when you need to measure how much rather than a yes-or-no probability.

Can I trust a Choice answer with low confidence?

Choice is the most probable option even when it barely wins, and confidence drops when options are close. Gate actions by confidence and route low-confidence cases to a human or queue.

Will adding more context to the state improve accuracy?

No. Accuracy drops as unrelated detail grows, so include only what the decision depends on and point questions at parts of a record with backtick paths.

How do I install it and set the key?

The module is a single Nushell file. git clone the repository, use jev.nu/jev, then set $env.TYPESAFE_API_KEY = "apikey_..." from console.typesafe.ai/keys. Put those last two lines in config.nu to keep them. Nothing is compiled and there is no package to install; jev noul, jev choice and jev score only build questions, and nothing reaches the network until jev ask runs.

How should I word a question, and its score levels?

Jev never sees the question id, so the whole question belongs in the instructions. It also judges each score level without seeing the others, so describe a level as a situation, for example a workaround exists, rather than as a degree. A word like moderate gives it nothing to match. Instructions and descriptions can be records or lists, which is where a rubric with examples goes.

How do I judge every row of a table without spending a fortune?

jev ask judges one state per call, and a list counts as one state, so judging every row means calling it once per row with insert. The README example trims a pasted log first with update body { str substring 0..3000 } so a long body does not eat the token budget, then adds jev as a column. Questions that only matter for some rows are still asked on all of them, because a second request costs more than the wasted answers.

CommunityJev Practices & Patterns

Jev.nu

Nushell module for Jev that pipes state into jev ask with noul, choice, and score builders, returning numeric answers that filter and sort in the pipeline.

See highlights

Open github.com

Stars: 8
Language: Nushell
License: MIT
Published by: Community
Author: cablehead
Last verified: Oct 2026

cablehead/jev.nu

User repository on GitHub

View on GitHub

Nushell module for the TypeSafe System One API: typed decisions with calibrated probabilities

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A Nushell module for jev, TypeSafe's model that answers questions with probabilities instead of text. Pipe in content and ask questions about it. The answers are numbers, so you can filter and sort on them. The module is a single Nushell file. Put the last two lines in config.nu to keep them. Jev reads text like an LLM but never writes any. You list the possible answers, and it gives each one a probability. TypeSafe calls this a System One model. A request has three parts:

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import nushell

Works forJev for Classification

Used inJev for Software & DevTools

Highlights

Single Nushell file; clone, use jev.nu/jev, and set TYPESAFE_API_KEY to start.
Builders only construct questions; nothing is sent until jev ask runs.
Asks everything in one parallel call where one answer never affects another.
Supports record state with backtick paths and per-row judging via insert.

Watch out

Needs a TypeSafe key and network access; answers shift slightly between runs, so leave room around thresholds and pin the model once tuned. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev_local

Local unofficial /v1/systemone adapter that serves Choice, Score, and Noul probabilities via first-token LLM judgments or a fine-tuned ModernBERT scorer, including Japanese docs.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

Argos1111/jev_localPython

jev-on-a-laptop

Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.