Is its confidence calibrated like the hosted Jev model?

No. It reproduces the API surface and decision semantics, but its confidence is a shape statistic. The defaults were the best measured setup, and raising temperature worsened both accuracy and calibration.

Does the prefix-caching layout actually save measurable time?

State plus questions share an identical byte prefix so SGLang and vLLM prefix caches reuse the prefill. Measured warm prefixes add little cost while cold prefills cost roughly per thousand tokens.

Can existing Jev clients point at it without rewrites?

Yes. It answers choice, noul, and score with Jev-shaped probabilities, confidence, legends, and usage, so existing clients point at the new URL with the backend chosen by environment.

What breaks if I tune temperature or logit bias?

Defaults are temperature zero with no logit bias. Nonzero temperature rescales every reported probability because the backend applies it before computing logprobs, and both values are echoed by health.

How do I install it, and which backends can it sit in front of?

pip install git+https://github.com/TOSUKUi/jev-bridge.git, then set JEVB_BACKEND_BASE_URL and JEVB_BACKEND_MODEL and run jev-bridge serve. It fronts any OpenAI-compatible server such as SGLang, vLLM, llama.cpp, LMDeploy or TGI, and can also point at the official OpenAI API with a key.

How does it turn next-token logprobs into typed answers?

It puts the state in the first user turn with each question after it, appends the options as single-token labels, requests max_tokens=1 with logprobs and top_logprobs, and computes a restricted softmax over the labels so probabilities sum to 1 over the declared options. No fine-tuning, calibration training or output-token decoding is involved.

What are the concurrency and confidence defaults?

JEVB_MAX_CONCURRENCY defaults to 8 in-flight question calls per request, and JEVB_TOP_K to 20 minimum top_logprobs. Confidence uses linear by default, (K*p_max - 1)/(K - 1), with max_prob and entropy as alternatives; the bridge states this is a distribution-shape statistic, not Jev's RLCD-calibrated confidence.

CommunityJev Tools & CLIs

Jev Bridge

Jev-style /v1/systemone API over any OpenAI-compatible backend, reading single-token label logprobs with a restricted softmax instead of generating JSON.

See highlights

Open github.com

Stars: 2
Language: Python
License: MIT
Published by: Community
Author: TOSUKUi
Last verified: Oct 2026

TOSUKUi/jev-bridge

User repository on GitHub

View on GitHub

Jev-style /v1/systemone API in front of any OpenAI-compatible LLM server (one-token logprob scoring, MIT)

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A Jev-style /v1/systemone API in front of any OpenAI-compatible LLM server. TypeSafe AI's Jev is a "System One" model: you send it a piece of state plus typed questions, and it returns typed answers with probabilities — no generated prose. Jev itself is a closed, hosted model. jev-bridge reproduces its API surface and decision semantics on top of any OpenAI-compatible backend (SGLang, vLLM, llama.cpp, LMDeploy, TGI, …) by reading the model's next-token distribution for the answer labels instead of generating JSON.

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

Tagscommunity github

Works forJev for Scoring

Used inJev for Software & DevTools

Languages: Python

Highlights

Shared byte prefix across questions lets SGLang radix cache and vLLM prefix caching reuse the state prefill.
Single-token labels with max_tokens one plus logprobs feed a restricted softmax summing to one over declared options.
Default linear confidence maps uniform to zero and one-hot to one, matching published Jev example values.
Backend, model, concurrency, temperature, and thinking controls come from JEVB environment variables.

Watch out

Confidence here is a distribution-shape statistic, not RLCD-calibrated confidence, so thresholds drift; dict key reordering changes the cached prefix. 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

brier

Open local decision model on Qwen3-4B with LoRA adapters that answers choice, score, and noul questions with probabilities, calibrated confidence, and conformal sets, offline with no per-call cost.

Jev Tools & CLIs#community#github

Community

Mangaba-ai/brierPython

paper-radar-jev

Python CLI that pulls recent arXiv papers, scores relevance with TypeSafe and writes daily Markdown and JSON rankings with resume-safe storage.

Jev Tools & CLIs#community#github

Community

LYchoon/paper-radar-jevPython

rwkv-jev

RWKV-7 decision-model server exposing a Jev-compatible /v1/systemone endpoint, with constrained decoding over answer tokens and zero output tokens.

Jev Tools & CLIs#community#github

Community

XingQiPan/rwkv-jevPython

pyGPT

A desktop AI assistant with a Jev / System One inline plugin for classification, routing, verification, and scoring without chat-style output.

Jev Tools & CLIs#community#python#desktop

Community

szczyglis-dev/py-gptPython

judgekit

CLI and Python library that runs declarative YAML judgment tasks - classify, route, score, verify - natively on Jev Choice/Score/Noul, on any OpenAI-compatible backend, or on local keyword rules, with confidence gates, fallbacks and per-decision cost records.

Jev Tools & CLIs#community#yaml#classification

Community

lexingtonhibiki/judgekitPython

jev-tmmluplus-eval

Harness that sits Jev for the 66-subject Traditional Chinese TMMLU+ suite, asking every item as one four-way A/B/C/D choice with calibration reporting.

Jev Tools & CLIs#community#github

Community

lianghsun/jev-tmmluplus-evalPython

From the Jev guides

Original write-ups on what this entry is for.

Jev on the Gateways: Vercel, Workers AI, and OpenRouter

Oct 12, 2026

The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.

Jev Tools & CLIs#gateway#vercel#cloudflare

Read article

Exporting Decision Models: GGUF, ONNX, and Device Runtimes

Oct 10, 2026

Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.

Jev Tools & CLIs#open-models#local#onnx

Read article

Records at Rest: Documents, Inboxes, and Risk Queues

Sep 28, 2026

Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.

Jev Tools & CLIs#classification#email#security

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai