local-jev

Run the Jev/System One pattern locally, without a hosted model call.

Source Releases Install

What It Is

local-jev is a small local service for typed decisions: yes or no, multiple choice, and ordered ratings. It packages the Jev-style idea into a local tool that downloads an open GGUF model, runs llama.cpp, and reads answer probabilities from logits instead of generating prose or JSON. It is not official TypeSafe Jev and does not include TypeSafe weights.

Local by defaultRuns on your machine after downloading a GGUF model.

Jev-style decisionsUse fixed answer options for noul, choice, and score.

System One APIAccepts POST /v1/systemone requests on localhost.

Agent-friendlyRoute tasks, check policy, classify relevance, or gate tool calls.

Logits, Not Parsed Text

A normal LLM wrapper asks the model to write text, then tries to parse that text into JSON. local-jev takes a shorter path. It builds a prompt with evidence, a question, and fixed answer options labelled A, B, C, and so on. Then it runs the local GGUF model once and reads only the next-token logits for those allowed letters.

Logits are the raw scores a language model gives to possible next tokens. local-jev applies softmax to the selected answer-letter logits, so they become an option distribution. The model never has to write {"choice":"billing"}. It only has to make one allowed letter more likely than the others.

Why It Exists

Not every model call needs a chat answer. Agents and automation often need a small structured decision: is this true, which route fits, or how severe is this? local-jev gives that workflow a local package and a familiar HTTP surface.

Quickstart

git clone https://github.com/ra-yavuz/local-jev
cd local-jev
bin/local-jev setup
bin/local-jev serve

The default server listens on 127.0.0.1:8010.

Example

curl -s http://127.0.0.1:8010/v1/systemone \
  -H 'content-type: application/json' \
  -d '{"state":"The deployment completed and health checks passed.",
       "questions":{"success":{"type":"noul",
       "instructions":"Did the deployment succeed?"}}}'

Kev Comparison

Kev is a separate open Jev-like model family built on Qwen3.5, with trained adapters, a decision readout, evaluation data, and a compatible server. local-jev is smaller: it uses regular local GGUF weights and reads answer-letter logits directly. The repository includes scripts/try-kev.sh so you can run Kev beside local-jev and send both servers the same sample request.

Benchmark Smoke Test

The repository includes scripts/benchmark-routing.py, a 100-case support-routing smoke test. On one local CPU run with the default Qwen3 0.6B Q4_K_M preset, it returned 42 correct and 58 wrong answers. Median latency was 997 ms, and the 95th percentile latency was 2389 ms. The five answers with confidence at or above 0.75 were all correct, but the middle confidence bands were noisy. Treat the scores as validation signals, not calibrated truth.

Disclaimer

local-jev downloads and runs third-party model files and uses model output to score decisions. It is provided as is, without warranty of any kind. You alone are responsible for installing, running, and acting on its output. Do not rely on it for safety-critical, legal, medical, financial, employment, law-enforcement, or similarly sensitive decisions without independent review.

Author

Published by Ramazan Yavuz as an independent open-source project.