Is this actually Jev-compatible or just shaped like it?

The shape, yes; the behavior, no. The internal package is named jev for wire clarity, the served endpoint is /v1/systemone, and the official TypeScript SDK runs against it unchanged in smoke tests. The README says full behavioral and API compatibility is not established, that training targets include outputs from a pinned Jev version, and that it is not affiliated with or endorsed by TypeSafe.

Can I run this on Linux or on a GPU box?

Not today. Apple MLX is the primary training and serving backend and the quickstart states it requires Apple Silicon, with base models downloading from Hugging Face on first use. A Rust runtime and portable export are described as future options gated on profiling evidence, not as shipped paths.

How good is the small model, and how was that measured?

The numbers are the project's own. The 2B q8 tier is reported at +6.8 accuracy points over the 0.6B on a fresh reserved test split, CI [+4.5, +9.2]. A 24-question ladder the README calls an illustration rather than a benchmark puts the served 2B q8 at 22/24, the fine-tuned 2B at 22/24 and pinned Jev at 23/24, with per-example misses published in examples/.

Do I have to accept whatever probabilities it hands me?

No. Each served tier ships an identity-bound temperature.json, and POST /v1/calibrations fits a temperature to your own traffic from about 100 labeled decisions. The structured-v1 renderer preserves JSON and primitive identity, historical adapters fall back to legacy-v0 automatically, and incompatible overrides are rejected rather than silently accepted.

What actually ships in the repository?

A Python and Apple MLX research prototype: versioned readouts, LoRA training, matched experiments, fitted temperatures, and a bounded local HTTP service, with structured-v1 training runs complete for SmolLM2-135M and Qwen3-0.6B. Datasets and full-weight models are not bundled in git, but the selected LoRA adapters for both tiers and the experiment controls are.

What does the 24-question ladder show?

It is an authored illustration, not a benchmark. On it the bare 2B base and the fine-tuned 0.6B each score 17 of 24, the fine-tuned 2B and its 8-bit served build each score 22 of 24, and pinned Jev scores 23 of 24. Fine-tuning buys five questions on the same backbone, and 8-bit quantization costs nothing.

How do the eight use-case examples present themselves?

Each example, from support triage and model routing to citation checking and lead scoring, ships the actual output of five systems from the bare backbone through pinned Jev, including the honest misses. The repository also implements canonical data, split and provenance checks, grouped comparisons, independent Score levels, and a readout for up to 255 options.

CommunityJev Practices & Patterns

System One — mpuig

An open System One decision model on small fine-tuned backbones, served locally through Apple MLX and speaking Jev's request/answer shape. Two model tiers with identity-bound temperatures, calibration fitted to your own labeled decisions, and a comparison ladder that includes the misses. A research prototype.

See highlights

Open github.com

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

mpuig/system-one

User repository on GitHub

View on GitHub

An open-source System One decision model — Kahneman's term for the fast, automatic judgment faculty. Calibrated choice/score/noul probabilities in one forward pass on small fine-tuned open models. Local on Apple Silicon (MLX), Jev-compatible API, audited experiment log.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

This is one of several Jev listings for the same project. The canonical entry is System One — iamaamir.

About this project

An open learning project inspired by TypeSafe's Jev. The goal is a small fine-tuned model, built on an existing backbone such as SmolLM, Qwen, or MiniCPM, with task-specific changes and efficient local inference. Apple MLX is the primary training and serving backend; a Rust runtime is an option if profiling justifies it.

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

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: Python

Highlights

Two served tiers: a 0.6B volume model and a 2.5 GB MiniCPM5-2B q8 quality model, each with its own identity-bound temperature file.
A 24-question correctness ladder, described as an illustration rather than a benchmark: 22/24 for the served 2B tier against 23/24 for pinned Jev.
Per-workload calibration fits temperatures to your own traffic from roughly 100 labeled decisions, over one call to /v1/calibrations.
Eight runnable use-case examples print real output from five systems, from the bare backbone through to pinned Jev, including where each one misses.

Watch out

The quickstart needs Apple Silicon for MLX plus uv, and portable export and a Rust runtime are listed as later, evidence-gated options. Compatibility is wire-level only: the internal Python package is even named jev and the official TypeScript SDK runs unchanged in smoke tests, but full behavioral and API compatibility is explicitly not established, training targets include outputs from a pinned Jev version, and the project states it is not affiliated with or endorsed by TypeSafe. Calibration and unfamiliar-rubric generalization are research goals, not guarantees. The accuracy claim is a +6.8 point gap on a fresh reserved test split for one task mix, and the 24-question ladder is an authored illustration. 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

jev-rag-benchmark

Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

erendikmenn/jev-rag-benchmarkPython

local-jev

Local offline System One server speaking the hosted /v1/systemone wire format, so existing SDK code runs unchanged against small on-machine models.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

amithgc/local-jevPython

jevshield

Framework-agnostic guard layer that puts Jev in front of an agent's tool calls: a Choice risk tier, a Noul destructiveness check and a Score blast radius decide allow, deny or ask.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

lgy1027/jevshieldPython

jev-mcp-server

Python MCP server exposing Jev's three official question types plus compare, verify and batch classify, with an installer subcommand that writes the client config and a measured cost per call.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

wangkuangkuang/jev-mcp-serverPython

qwen-rlcd

An inference-only prototype of a Jev-style decision model on a sub-billion-parameter hybrid backbone, returning typed probability distributions in one forward pass, with its architecture treated as an explicit guess.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

shamazharikh/qwen-rlcdPython

typesafe-vs-deepseek

A Vercel-hosted comparison of TypeSafe (Jev) against DeepSeek-flash on speed, tokens, cost and accuracy, backed by per-task Python pipelines, fixture generators and a report script. Eighteen commits, nine issues still open.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

markfive-proto/typesafe-vs-deepseekPython

From the Jev guides

Original write-ups on what this entry is for.

Designing the State: What to Include and What to Drop

Oct 12, 2026

Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.

Jev Practices & Patterns#state#design#primitives

Read article

Noul, End to End: One Probability and a Dozen Thresholds

Oct 12, 2026

A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.

Jev Practices & Patterns#noul#primitives#confidence

Read article

Scoring Without Measuring: Bands, Rubrics, and Composite Triage

Oct 12, 2026

A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.

Jev Practices & Patterns#score#primitives#scoring

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai