User repository on GitHub
A local Jev: calibrated yes/no, choice and score judgments from small NLI models, in the browser or as a Node server.
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
Vej
A local Jev clone: calibrated noul, choice, and score judgments from small NLI cross-encoders, running in the browser or as a Node server with Jev's request and response shapes.
- Stars
- 0
- Language
- JavaScript
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- draeder
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- JavaScript
- Platforms
- Browser
- Capabilities
- Open sourceRuns locallySelf-hosted
Highlights
- Same three primitives, request, and response shapes as Jev, with the weights on the machine that asked.
- Answers three questions in about 150 ms from one batched forward pass over NLI cross-encoders of 100–200M parameters.
- No decoder and no generation: text becomes a premise, each outcome becomes a claim, and arithmetic turns the entailment distribution into an answer.
- 36 tests run against fake runtimes with no weights, no network, and no key.
- Built in a night during Jev's invite-only period, per the madewithjev listing.
Quickstart
ts
import { noul, VejClient } from "vej";
const vej = new VejClient();
const { answers } = await vej.systemOne({
state: "We drove the new route on Saturday. The trip took about an hour.",
questions: { by_car: noul("The writer travelled by car.") },
});
answers.by_car.noul; // 0.99Watch out
Copy reviewed against the repo README on 2026-10-15; the README is longer than the fetched excerpt, so the routing-per-question section was only partially reviewed. Speed figures are the author's measurements, and the local models are NLI cross-encoders rather than Jev's own weights.
Questions about this entry
Short answers drawn from the linked project or page.
A local clone of Jev: a model that returns a calibrated judgment instead of a paragraph. It runs in a browser tab or as a Node server from the same source, keeps Jev's request and response shapes, and answers three questions in about 150 ms.
It is built on NLI cross-encoders, models trained to say whether a claim follows from a text. State becomes a premise, each outcome becomes a claim, one batched forward pass returns an entailment/contradiction/neutral distribution per claim, and arithmetic turns that into the answer.
A noul asks one claim and takes entailment over the whole label set, reading false as the remainder, so unrelated claims land in neutral and come out a no. Choice and score ask one claim per outcome and normalize the same number across them.
Because e/(e+c) has a hole: an option nothing supports and nothing contradicts has both values near zero, so it scores about 0.5, a coin flip. The README documents a roadworks text where an unmentioned option took 34% that way, while entailment over the full set keeps it at 0.1%.
Yes. No single small model measured is good enough and their failures are disjoint, so several models are routed per question. Their errors are one-sided, under-claiming entailment rather than inventing it, which is what makes combining them work.
No. The model is never asked to write anything, and the outcomes are never shown to it as a list to prefer the top of, so a choice cannot have position bias and generation is not possible.
Most of the plumbing: 36 tests run against fake runtimes with no weights, no network, and no key. State assembly, claim construction, and the arithmetic that turns distributions into answers all have no model in them.
More Jev resources like this
A 100% offline, zero-cost Rust drop-in for the TypeSafe Jev System One API: one native binary serves Choice, Score, Noul and Route decisions on CPU in under 18 ms with no API keys, cloud egress or rate limits.
Jev Repos & SDKs#community#catalog-import#madewithjev
Community
ronitgupta138/local-jevRust
ronitgupta138/local-jevRustHTTP server for the Mica v0.5 4B Preview decision model: GGUF weights run through llama.cpp and answer noul, choice and score questions with calibrated probabilities over the TypeSafe /v1/systemone format.
Jev Repos & SDKs#community#github#python
Community
akivet/MicaPython
akivet/MicaPythonSelf-hosted Python SDK and System One-compatible API for runtime-defined Choice, Boolean and Score decisions on pinned open models, across Torch, vLLM, MLX, llama.cpp and WebGPU. Each result records probabilities, token usage, timing, model revision and prompt version, with committed row-level benchmarks.
Jev Repos & SDKs#community#sdk#self-hosted
Community
chengyongru/fastjevPython
chengyongru/fastjevPythonA correctness-validated Laya runtime for Apple Silicon that runs upstream Laya on the MLX GPU and the Apple Neural Engine, using the ANE only after an artifact passes a parity gate on your machine, plus a local Jev-compatible server.
Jev Repos & SDKs#community#catalog-import#madewithjev
Community
tc3oliver/laya-applePython
tc3oliver/laya-applePythonServing code for the Jeff 1.0 family of Jev-style decision models: a state, a typed question and its options go in, and one forward pass returns a probability for each option. Four open models from 4B to 31B parameters are served the same way.
Jev Repos & SDKs#community#github#python
Community
j-geuter/jeff-servePython
j-geuter/jeff-servePythonLaya GGUF inference and an HTTP server written in Goish Rust: one Linux binary runs upstream Laya or OpenThai-SystemOne on your CPU and serves typed choose, score and check decisions over HTTP, the CLI or a JSON-lines process.
Jev Repos & SDKs#community#catalog-import#madewithjev
Community
cogentica-ai/laya-goishRust
cogentica-ai/laya-goishRustFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.