Is this an official Jev model release?

No. It reproduces the interface pattern with open-weight models and does not reproduce the undisclosed Jev model or training. It is unaffiliated with TypeSafe.

How does one forward pass replace JSON generation?

Logits at the answer position are softmaxed over your labels only, so nothing outside the option set can win. No decoding loop, no JSON parsing, every answer carries a probability.

Can I trust the confidence numbers for gating?

Confidence is 1 minus normalized entropy, stated as a transparent approximation since TypeSafe does not publish its formula. Raw zero-shot probabilities run over-confident before calibration.

Which backends and models run today?

Hugging Face Transformers on CPU, CUDA, or MPS today, with a mock backend that needs no download. MLX, vLLM, and a TypeSafe reference mode are roadmap, and the REPL already streams bars and latency.

How do I install it and drive the REPL?

Python 3.10 or newer, then a uv venv at 3.11 and uv pip install -e ".[dev]" for core plus REPL. Plain openjev starts the mock backend instantly with no download, and openjev --backend hf --model Qwen/Qwen2.5-0.5B-Instruct runs real weights. In the REPL, /state sets the text, /choice, /score and /noul add questions, and /ask streams the answers.

What did the live experiments against real Jev measure?

python scripts/experiments.py all runs pelican, surface, latency and maze in roughly 130 calls with TYPESAFE_API_KEY in .env. Adding questions is nearly free at 290 ms for 1 question against 328 ms for 27, and the maze run solved a shortest path of 14 moves in 5.4 s with confidence dropping to 0.6 at corners and sitting at 0.99 in corridors.

Can the official SDKs point at it once a server ships?

That is the Phase 1 plan, not what runs today. Phase 0 ships the core primitives, the HF backend and the REPL, with Phase 1 adding openjev serve as a FastAPI POST /v1/systemone so typesafe-sdk or @typesafe-ai/sdk work by changing base_url. The playground, eval command and temperature calibration sit in Phase 2, and the MIT license is already in place.

CommunityJev Practices & Patterns

OpenJev — GPT-AGI

Open-source Jev-compatible decision engine with a Claude Code style REPL, answering Choice, Score, and Noul questions in one forward pass over open models.

See highlights

Open github.com

Stars: 19
Language: Python
License: MIT
Published by: Community
Author: GPT-AGI
Last verified: Oct 2026

GPT-AGI/OpenJev

User repository on GitHub

View on GitHub

Opensource Jev

Forks: 7
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is OpenJev — razorback16.

About this project

An open-source, Jev-compatible System One decision engine with a Claude Code style REPL. Typed decisions (Choice · Score · Noul) from open models in one forward pass. No JSON parsing, no hallucinated shapes, every answer comes with a probability. openjev play maze · left: live state · right: streaming /v1/systemone response, option probabilities, rolling confidence & latency. Replays real Jev decisions (14 moves, shortest path, 0 tokens generated).

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

Tagscommunity awesomejev catalog-import

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Same request and response shape as POST /v1/systemone for Choice, Score, and Noul.
REPL takes state and slash-command questions, then streams probability bars and latency.
Ships a reference harness replaying live Jev experiments for pelican, surface, and maze.
Python library accepts state plus questions and returns typed answers with usage.

Watch out

Phase 0 ships core primitives, HF backend, and REPL; serve mode, playground, calibration, and extra backends remain roadmap. 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_local

Local unofficial /v1/systemone adapter that serves Choice, Score, and Noul probabilities via first-token LLM judgments or a fine-tuned ModernBERT scorer, including Japanese docs.

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

Community

Argos1111/jev_localPython

jev-on-a-laptop

Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.