open-jev
From-first-principles PyTorch reconstruction of a Jev-like architecture with typed noul, choice and score heads; random weights only.
Overview
Open Jev is an unofficial research implementation in PyTorch that reconstructs the ideas behind TypeSafe's Jev from first principles. It encodes JSON-like state once with a bidirectional transformer and structural path embeddings, then answers every question through learned query slots that cross-attend into the shared state cache. Questions are isolated from each other and folded into the batch dimension for one parallel pass.
There are three typed heads: Noul returns the probability a statement is true, Choice softmaxes only over options supplied at runtime, and Score uses ordinal thresholds to return a distribution and expectation. An evidential head models epistemic confidence separately, and an RLCDLoss combines soft-target NLL, Brier, consistency, confidence and ECE terms. The README is explicit that weights are random, the tokenizer is a deterministic hash tokenizer unsuitable for training a useful model, and no results are claimed. It is educational.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Probability a statement about the state is true | true · false |
| Choice | Which runtime-supplied option applies | — |
| Score | Position on ordinal levels | — |
How it works
- Encode state once with the bidirectional encoder.
- Attach query slots per question.
- Read typed outputs from noul, choice and score heads.
- Train with RLCDLoss if you supply data and a real tokenizer.
Strengths
- Small, legible reference architecture.
- Shows typed heads and calibration-oriented loss terms.
Caveats
- Random weights; not a usable model.
- Hash tokenizer unsuitable for real training.
- No benchmark results.
Getting started
git clone https://github.com/kyegomez/open-jev && cd open-jev && python -m pip install "torch>=2.0"At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
Use cases and integrations
How curated lists describe it
an open-source, from-first-principles reconstruction of the ideas behind…
wh000wh000/awesome-jev-liveAn open-source, from-first-principles reconstruction of the ideas behind TypeSafe AI's Jev, written in pytorch.
hellogumbo/awesome-jev
Frequently asked questions
Is this the real Jev?
No, it is an unofficial reconstruction with random weights.
Can I use it in production?
No, it is for studying the architecture.
What is RLCDLoss here?
A combined loss of soft-target NLL, Brier, consistency, confidence and ECE terms.
Related projects
33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonInstruction dataset of 119 rows compiled from TypeSafe's public Jev documentation, split by Choice, Noul and Score.
ProjectresearchPythonKorean study notes and TypeScript mock-ups of Jev concepts, with a local Laya versus GPT-5 comparison and a calibration toolkit.
ProjectresearchTypeScriptLearning project that builds choose-an-answer decision models from scratch and from Qwen, with downloadable 0.8B and 4B releases.
LibraryresearchPythonMinimal training project for a Jev-style decision model: Qwen3-0.6B with a pointer head that outputs probabilities over options.
ModelresearchPythonPreprint survey of Jev-style typed decision models: their research roots, adaptation costs and what current evidence supports.
Modelresearch