open-alternative-jev

What it is

A Python package that gives typed, calibrated decisions from any open-weights LLM in one forward pass, running locally on your own GPU. For developers who want Jev-style decision-making without the official SDK.

How it uses Jev

Jev reads a state once and answers each typed question (Choice, Score, Noul) from the next-token distribution at its own position, restricted to the options you give. The result is a choice with a calibrated probability, used directly in code.

Primitives:choicescorenoul

Technique worth stealing

Packing writes the shared state once instead of once per question, cutting tokens and boosting throughput.

Try it

pip install open-alternative-jev, then use Decider.from_pretrained with backend='hf'.

View on GitHub