Jevlet
From-scratch research reconstruction of a small System-One decision model, used to route and gate commands in a Windows palette.
Overview
Jevlet is an independent research reconstruction of a Jev-like System-One model plus a laptop 'daily driver' built on it: an always-on Windows command palette. It is not TypeSafe's implementation; a decisions document lists which public facts each design choice rests on. The model takes one state and several typed questions (Noul, Choice, Score) and returns probabilities over runtime-defined options in one packed pass, with no text generation.
The architecture packs the state and every question branch into one sequence; branches see the state but not each other, and options can be isolated so probabilities are order-invariant. A pretrained variant uses a BERT-family encoder (bge-small since v4) and a scratch byte-level variant is used for topology research. On 78 hand-written laptop tasks the README reports v4 at 97.4% route accuracy with ECE 0.056, a P(safe) of at least 0.9 required before running a task unattended, and about 35 ms median GPU latency.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which skill or route should handle this command? | — |
| Yes / no | Is it safe to run this task unattended? | — |
| Choice | Which on-screen element or app does this command refer to (grounding)?(inferred) | — |
How it works
- User types a natural-language command into the palette
- Jevlet encodes the state once and answers route, risk and grounding questions in one pass
- Task runs unattended only if P(safe) >= 0.9; otherwise it asks or blocks
- Planner maps the chosen route to an assistant command
Strengths
- Runs locally; about 40 ms on CPU for three questions per the README
- Order-invariant option scoring by construction
- Reported calibration (ECE) alongside accuracy
- Colab notebook and Hugging Face weights available
Caveats
- Research project on a small hand-written benchmark
- Not affiliated with or a copy of TypeSafe's Jev
- Windows-specific daily driver
Getting started
Open the Colab notebook notebooks/jevlet_colab.ipynb or load the weights from Hugging Face (NAME0x0/Jevlet).At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- Windows desktop app, PyTorch
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Jevlet: a small, calibrated System-One decision model (typed Noul/Choice/Score questions in, probabilities over live options out, no text generation) driving an always-on Windows command palette. Research reconstruction of TypeSafe's Jev idea.
hellogumbo/awesome-jev
Frequently asked questions
Is Jevlet the same as Jev?
No. It reconstructs the published input/output idea from scratch and makes no claim about Jev's internals.
What does it decide on the laptop?
Which route or skill a command goes to, whether it is safe to run unattended, and which element it refers to.
Does it need an API key?
No, the model runs locally.
Related projects
- Strands Decider 2B★ 612
AWS Strands Labs' 2B decision model: Qwen3.5-2B with a pointer head scoring options and calibrated confidence, plus a CLI.
ModelbetaPython - poorjev★ 10
Local Jev alternative on zero-shot NLI models with temperature scaling and conformal abstention; reports ECE 0.170 to 0.071, offline.
ServicebetaPython OneJev: open multimodal System One model (0.8B-27B) answering typed questions about screenshots, photos, video and text in one pass.
ModelbetaPython- CLM★ 3,014
CLM-8B, a contrastively trained System One model that embeds states and actions separately, served behind a TypeSafe-compatible API.
ServiceresearchPython 706K-parameter Jev-style option scorer that rates fill, check, click or skip for each form element in one parallel pass.
ModelresearchPythonspark-s1: local Jev-inspired decision models on Qwen3.5-4B for NVIDIA DGX Spark, trained on code- and solver-labelled decision rows.
ModelresearchPython