User repository on GitHub
Jev-style typed decisions (Choice / Noul / Score) from any model whose API exposes logprobs, plus a local /v1/systemone server
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Jevless
Independent toolkit that reproduces Jev-style Choice, Noul, and Score decisions on any model exposing log-probabilities, plus a local server speaking the same systemone request shape.
- Stars
- 2
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- jbpayton
- Last verified
- Oct 2026
About this project
Jev-style typed decisions from any model whose API exposes log-probabilities. An agent makes lots of small decisions: route this ticket, is this memory relevant, did that step succeed, how urgent is this. Asking a model to write "yes" and parsing it is slow and gives no probability. jevless asks the question once and reads the probabilities of the answer letters at the first output position. That's one prefill and one token, and you get a probability for every option.
Quoted from the project's README. Read the full README.
Highlights
- Reads answer-letter probabilities at the first output position for one-prefill decisions.
- Runs on a local 9B via LM Studio, vLLM, llama.cpp, or hosted APIs including OpenAI.
- Evaluates both option orders in parallel with keepalive connections and streamed answer mode.
- Reports 60 labeled basic and hard decisions with per-model time, token, and cost tables.
Watch out
Independent and unaffiliated; Claude models run in answer mode without logprobs and reasoning effort changes results. 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.
Mostly. The same 60 labeled decisions in basic and hard sets went to every reachable model, with median timing network included and list-price costing. Run-to-run accuracy moves a question or two and small sets compare models rather than certify them.
Claude exposes no logprobs API, so jevless runs in answer mode and takes the reply letter, stopping streamed output at the letter. Reasoning effort then affects time and tokens.
Both orders is the default for stability. Adaptive second order reads the reversed order only when the first top option falls below the auto threshold, saving input tokens at higher log loss.
No. It is an independent open implementation of the same decision idea over ordinary model APIs. Pick a local 9B for free local calls or a hosted model for higher hard-set accuracy.
At the first output position it reads the probability of every answer letter, so a choice or noul is one prefill and one token and you get a probability for each option, rather than generating text and parsing a reply.
Any model whose API exposes logprobs, run through a local server that speaks /v1/systemone: a local 9B in LM Studio, vLLM or llama.cpp, or a hosted API such as OpenAI. Models without logprobs still run in answer mode.
It sends both option orders at once, which halves the time per decision at no cost, keeps connections alive to save the handshake after the first request, and streams in answer mode, stopping at the answer letter.
More Jev resources like this
1GitHub stars
Condor drop-in where Jev Noul fan-out ranks Meteora pools and a Score sizes positions, while math clamps enforce rug cards, cost tiers, dry-run, and one-sided walls.
Jev Agents & MCP#community#github
10GitHub stars
Reproducible study testing whether Jev zero-shot can match specialist rerankers on fixed MS MARCO passage, MS MARCO document, and TREC-1 WSJ candidates, reporting ranking quality, time, and API cost.
Jev Agents & MCP#community#github
Community
carlaiau/jev-rerankingPython
carlaiau/jev-rerankingPython6GitHub stars
Open-weight 4B decision model on Qwen3.5-4B-Base serving a Jev-compatible /v1/systemone API, returning per-option probabilities with adjustable thinking.
Jev Agents & MCP#community#github
Community
org2AI/wald-4bPython
org2AI/wald-4bPython3GitHub stars
A Python decision gateway and stdio MCP server that exposes Jev through OpenRouter's alpha decisions endpoint. Five tools cover yes/no checks, classification, scoring, multi-question evaluation and health, with bounded retries, validated answers and local diagnostics.
Jev Agents & MCP#community#mcp#openrouter
Community
ctmx/openrouter-jev-mcpPython
ctmx/openrouter-jev-mcpPython3GitHub stars
MCP server that exposes Jev classify, score, yes-no check, risk gate, and multi-question decide tools to coding agents through one Jev call per tool.
Jev Agents & MCP#community#github
Community
codaaiteam/jev-mcpJavaScript
codaaiteam/jev-mcpJavaScript2GitHub stars
Home Assistant conversation integration that routes Assist utterances with Jev Choice and Noul scoring, binding intents and entities locally before falling back on harder requests.
Jev Agents & MCP#community#github
From the Jev community
Posts from builders shipping with Jev right now.