jevbooks

← All projects

AnyJev

nokia-applied-research/AnyJev · Homepage

Turn any LLM into a Jev-style decision model with typed decisions and real probabilities.

Classification & Routing57%Runner-up: Infra / SDKs / IntegrationsQuestionnaireCheck-up

What it is

AnyJev is a Python library that turns any open transformers or vLLM model into a Jev-style decision model. It lets you ask typed questions (Choice/Score/Noul) and get calibrated probabilities from one prefill, with no training. It is for developers who need reliable, thresholdable decisions from LLMs.

How it uses Jev

AnyJev implements Jev-style decisions by reading the next-token distribution from one prefill. It uses cyclic-shift marginalization and a label-free prior estimate to reduce option-order sensitivity. Results are labeled raw, L0, or L1 based on debiasing and calibration. L2 uses a closed-form head on hidden states for better accuracy.

Primitives:choicescorenoul

Technique worth stealing

Cyclic-shift marginalization and label-free prior estimate reduce option-order sensitivity without labels.

Try it

pip install "anyjev[hf]" then run python -m demo.jev_mode --backend fake

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.95
  • Shows a System One patternno0.33
  • Handles uncertaintyno0.07
  • Measuredno0.30
  • Runnableyes0.86
  • Worth recommendingunclear0.58
  • Model replicayes0.90
  • Problem scopescore on a 0–2 scale1.29
  • About Jevyes0.94

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 23 Sept 2026.