Is this using TypeSafe Jev under the hood?

No. Jev names only the interface style of typed questions in and calibrated probabilities out. The engine is a diffusion LM with slot reads, temperature calibration, and an optional domain LoRA, all local.

Can I run tests without downloading weights?

Yes. pytest runs on stub tokenizer and stub reader with no model needed, plus ruff and mypy checks. Model downloads and calibration only matter for smoke, benchmark, and serve scripts.

Will the LoRA learn my domain or just templates?

That is what shift_eval and domain_eval address: the LoRA trains on slot-label cross-entropy with synthetic tickets carrying exact ground truth, then gets tested for policy learning versus template memorization.

What hardware and setup does serving need?

Use Apple silicon with 32GB+ memory, sync with uv, run smoke to fetch about 15GB, fit calibration with the benchmark, then serve FastAPI POST /v1/decide locally with optional adapter weights.

How does one decoder pass answer every question?

Each question gets one label-token slot in the diffusion canvas. One decoder pass produces every question's distribution, and only the label tokens are read, so an answer cannot leave the schema; questions are read in a single pass rather than generating an answer sentence or repairing JSON.

How does its calibration work?

It fits a per-type or per-question temperature on outcomes by negative log-likelihood and reports expected calibration error and Brier scores alongside the fitted values. The repo also ships a synthetic support-ticket generator with exact ground truth and a per-question calibration evaluation against it.

What license covers the code and the model files?

The code is Apache-2.0. The model weights and datasets are downloaded at runtime under their own licenses, so DiffusionGemma 26B-A4B and the Hugging Face ticket adapter carry their own terms rather than the repository's.

CommunityJev Models & Reimplementations

Dlm Jev

Jev-style local decision engine on DiffusionGemma 26B-A4B that reads one label-token slot per question in a single decoder pass to return distributions without generating text.

See highlights

Open github.com

Stars: 1
Language: Python
License: APACHE-2.0
Published by: Community
Author: seongyeon1
Last verified: Oct 2026

seongyeon1/dlm-jev

User repository on GitHub

View on GitHub

Jev-style typed decision engine on a diffusion LM (DiffusionGemma + MLX): one-pass slot reads, calibration, domain LoRA. Not affiliated with TypeSafe AI.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A Jev)-style typed decision engine built on a diffusion LM (DiffusionGemma 26B-A4B, MLX 4-bit). It doesn't generate text. It returns a probability distribution and a confidence for each typed question (noul / choice / score) about a given state. All questions are read in a single decoder pass. Slot read — each question gets one label-token slot in the diffusion canvas; one decoder pass gives every question's distribution, and only label tokens are read, so answers cannot leave the schema.

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

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Reads one label-token slot per question so answers cannot leave the schema in a single pass.
Fits per-type temperature on outcomes with ECE and Brier reporting for calibration.
Ships a support-ticket LoRA adapter with synthetic generator plus shift and interference tests.
Serves FastAPI POST /v1/decide on Apple silicon after smoke and benchmark calibration steps.

Watch out

Not affiliated with TypeSafe AI or Jev and uses no Jev weights, API, or data; needs 32GB+ Apple silicon and a large download. 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

gut

Python helper that turns judgment calls into one-liners like likely, classify and rate, defaulting to Jev while also running on local or OpenAI-compatible backends.