Are the benchmark numbers real?

They are the author's measurements on one idle M2 Max with 4-bit weights: question-first layout scored highest on the game sentences, with per-model accuracy, calibration, and latency tables for decision sets included.

How close is it to Jev?

On the author's comparison over public cases it trails Jev overall, roughly level on short texts but clearly behind on long documents requiring cross-checking across thousands of tokens.

What hardware and models do I need?

Apple Silicon with Python 3.10 or newer, about 20 GB free for the recommended mixture-of-experts model and less for the 4B, with Qwen 3.5-family MLX weights only.

What prompt lessons carry over?

Put the fixed question before the data for caching and accuracy, locate obstacles explicitly, avoid labels that collide with question words, and keep numeric routing like payment signs in code.

How do I start the server and point a client at it?

On Apple Silicon with Python 3.10 or newer, run pip install -e . then snapjudge-serve --model mlx-community/Qwen3.6-35B-A3B-4bit. It serves POST http://127.0.0.1:8724/v1/systemone in the TypeSafe System One format and a /game/ page, and models download from Hugging Face on first use.

How does the engine keep latency low?

The question head is prefilled once and cached as an LRU snapshot, so each request pays mostly for its state; a long state over 2,000 characters with several questions flips the order instead. Option keys become a token prefix tree, all branching points run as one batched forward pass, and the softmax runs over allowed tokens only, with the final projection in float32.

What debug and layout options exist?

Add debug: true to see timings, the top next tokens, and a coverage value for how much probability landed on allowed labels, and use layout to force state_first, question_first, or header. The default auto uses question-first, switching to state-first for a long state with several questions, and snapjudge-serve --api-key requires a bearer token.

CommunityJev Demos & Experiments

Snapjudge

Local Jev-compatible decision server for Macs that reads answer probabilities from Qwen next-token logits in one forward pass, with no text generation or JSON parsing.

See highlights

Open github.com

Stars: 14
Language: Python
License: MIT
Published by: Community
Author: Micha0827
Last verified: Oct 2026

Micha0827/snapjudge

User repository on GitHub

View on GitHub

Typed decisions (choice / score / yes-no) from local Qwen models on Apple Silicon. Probabilities come straight from the logits, no text generation. TypeSafe-compatible HTTP API, runs on MLX.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Typed decisions from local Qwen models on your Mac. Ask a question about some text, get back a probability for every allowed answer. No text generation, no JSON parsing: the probabilities are read straight from the model's next-token logits. The HTTP API follows the TypeSafe System One format (POST /v1/systemone, Choice / Score / Noul), so the same client code can talk to Jev or to your own machine.

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

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: Python

Highlights

Serves a System One-shaped endpoint with Choice, Score, and Noul so the same client code can target Jev or localhost.
Question-first prompt layout with cached prefixes keeps median latencies in the low hundreds of milliseconds.
Confidence is one minus normalized entropy, with debug output adding timings, top tokens, and label coverage.
Runs Apple Silicon Qwen 3.5-family MLX models, one request at a time, with models fetched on first use.

Watch out

Probabilities are not proven calibrated and long-document questions lag Jev clearly, so validate on your own labels first. 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

open-jev-typed-decision-engine

Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.