User repository on GitHub
A small open decision model: state + typed questions -> calibrated probabilities. A Jev / System One re-creation on Qwen3.5.
- Forks
- 19
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Practices & Patterns
Reflex — kshetrajna12
An open re-creation of Jev's decision model on an open-weights Qwen3.5-4B: it answers a list of typed questions in one forward pass over POST /v1/systemone in about 200 ms, never writing free text.
- Stars
- 169
- Language
- Python
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- kshetrajna12
- Last verified
- Sep 2026
About this project
A tiny "decision model" you can run on your own GPU. You give it some information (a support ticket, a document, a photo) and a list of questions with fixed answer options. It answers all the questions at once and tells you how sure it is about each option, as percentages. It never writes free text, so it can never make up an answer that isn't on your list.
Quoted from the project's README. Read the full README.
Works forJev for Evaluation
Used inJev for Software & DevTools
Highlights
- On public JevBench items it reports 1.000 easy, 0.917 standard, and 0.685 hard accuracy with hard-tier ECE 0.081 at about 200 ms.
- A frozen Qwen3.8-27B configuration reaches 0.958 standard, 0.766 hard, and ECE 0.061, at about 1 s per request.
- Official JevBench v1.2 placed the 4B configuration #5 of 36; fine-tuning in every form tried is reported not to help.
- What helped was the readout: lettered yes/no answers, Evidence/Criterion framing, and averaging two distinct option orders.
- A WebGPU browser demo runs a 650 MB Qwen3.5-0.8B locally in Chrome, Edge, or Safari 18+, uploading nothing.
Quickstart
bash
git clone https://github.com/kshetrajna12/reflex
cd reflex
git checkout stable
uv sync
uv run python examples/support_ticket.pyWatch out
MIT-licensed (Qwen weights keep Apache-2.0); needs a Linux machine with an NVIDIA GPU (8 GB for the default model), Python 3.12, and uv, and is not affiliated with TypeSafe.
Questions about this entry
Short answers drawn from the linked project or page.
Both kinds are published. The 1.000 / 0.917 / 0.685 table is self-run on a suite the author consulted throughout development, and they say so. The official JevBench v1.2 run - 534 decisions including held-out items, run by the benchmark's author on an H100 - placed the 4B configuration 5th of 36 with a score of 71.7.
On one tier, narrowly, at a cost. The 27B configuration reached 75.9 percent on the hard tier against Jev's 74.1 and took the board's best Intelligence and Calibration scores, but it was held back by cost and speed, and the 4B default sits below Jev on the standard and hard tiers.
Not the configuration that ships. It wants Linux, an NVIDIA GPU with 8 GB for the default model, Python 3.12 and uv. Two smaller paths exist: PyTorch MPS on Apple silicon (0.8B in about 130 ms on an M3 Pro) and a browser WebGPU demo, which the README calls a demonstration of the mechanism rather than of the served quality.
Yes. The request and response shapes match TypeSafe's hosted API, so a client aimed at api.typesafe.ai can be pointed at http://localhost:8008. To get exactly what the server serves, load the stable manifest rather than Engine.load with defaults, because the two option orders are part of that configuration.
A Linux machine with an NVIDIA GPU; 8 GB of memory is enough for the default 4B model, and any 16 GB CUDA GPU runs the recommended stable server.
On the official JevBench v1.2 of 534 decisions run on an H100, reflex placed #5 of 36 with score 71.7, and reflex-27b placed #22 with the highest Intelligence (90.5) and Calibration (86.2).
Fine-tuning in every form tried: four LoRA mixes and a 27B distillation won on training-shaped data then lost general judgement; wording ensembles and a reasoning cascade were measured and dropped too.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsOther listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
A Mac menu-bar app that routes every coding agent you run in herdr from one ask bar, phone remote or edge rail: requests go to Jev via jev-fabric as typed questions over self-built candidates — which session, which folder, side question or continuation — and come back as keys and probabilities in about a second.
Jev Tools & CLIs#community#github#swift
Community
monotykamary/reflexSwift
monotykamary/reflexSwiftCoding agent on Pi that pairs any OpenRouter LLM with TypeSafe Jev reflex checks on tool calls, turns and transcripts, where code thresholds decide allow, ask or block.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
kaustav1996/reflexTypeScript
kaustav1996/reflexTypeScriptMore Jev resources like this
An independent baseline that reads typed option probabilities straight from a frozen Qwen3.5-4B's logits in one forward pass, reproducing Jev's interface pattern with open models rather than Jev's undisclosed model or training.
Jev Practices & Patterns#community#python#open-models
Community
TheoLeeCJ/SemIfPython
TheoLeeCJ/SemIfPythonApache-licensed, locally runnable Jev-style decision models (0.8B, 4B, 9B on Qwen3.5) with released weights, training code, a System One-compatible server, frozen eval suites, and a playground.
Jev Practices & Patterns#community#python#open-models
Community
jaredpalmer/kevPython
jaredpalmer/kevPythonAn independent open reproduction of the System One model class: Qwen3.5-based 2B and 35B mixture-of-experts models that return typed Choice, Score, and Noul probabilities in one forward pass, with nothing distilled from Jev.
Jev Practices & Patterns#community#python#open-models
Community
Mapika/deciderPython
Mapika/deciderPythonAn encoder-side reproduction of Jev built on a retrained GLiClass ModernBERT base (151M): it evaluates multiple typed questions in one non-autoregressive forward pass, returning choices, ordinal scores, and probabilities under 35 ms.
Jev Practices & Patterns#community#python#open-models
Community
Heman10x-NGU/Verdict-open-jevPython
Heman10x-NGU/Verdict-open-jevPythonA from-scratch one-pass option scorer that maps a context and a changing list of text options to one probability per option, adding a hashed n-gram encoder, rival-aware attention, a gated head, and temperature scaling; JSONL data is drop-in compatible with jevlike.
Jev Practices & Patterns#community#python#calibration
Community
olanotolu/jevbetterPython
olanotolu/jevbetterPythonA ModernBERT-based open decision model that evaluates Choice, Score, and Noul schemas in one non-autoregressive pass; Verdict 2.0 reports 77.10% accuracy on 2,000 held-out enterprise decisions.
Jev Practices & Patterns#community#python#open-models
From 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.