Typesafe Jev — Testing skill for Claude Code

typesafe-jev


Field notes, runnable scripts and an agent skill for [TypeSafe Jev](https://docs.typesafe.ai), the System One decision model. Everything here was measured on `jev-1.13.0` in September 2026 while building and testing a production, Arabic-first business application. When a newer model ships, re-run the evals before trusting any number on this page.

Jev does one thing. You send it text and a closed set of options, and it returns a choice with a probability for every option and a confidence score, for a fraction of a cent. We measured 200 to 500 ms on a warm connection and 600 to 1,100 ms cold. It cannot write, count, do arithmetic or look at a screenshot. Used as a decision layer under code you control, it fits the places where a keyword rule is too brittle and a full LLM call is too slow or too costly.


That chart is the reason to care. On 60 synthetic routing phrases in English, Standard Arabic, Arabic dialect and mixed script, raw accuracy was 83%. All 10 wrong answers came back with confidence under 0.8, and the highest was 0.72. Act above the gate and do what you did before below it, and in this sample no wrong answer would have reached a user.

Read that for what it is. Sixty phrases and zero observed errors above the gate is an encouraging result, not a calibration study and not a guarantee about your traffic. With 37 gated answers and no errors, the 95% upper bound on the true gated error rate is still about 8%. Run the eval on your own cases, and keep shadow mode on until you have hundreds.

Five-minute start

Requires Node 18+ and Python 3.9+. The eval, the text judge and the tree test need no packages. The browser loop