What is the JevDrive project?

A research prototype testing where Jev can help an existing driving policy choose continue, observe, replan or defer, while a local controller and feasibility checks keep ownership of vehicle execution. Jev receives structured text, not images.

What did the 27-image probe measure?

On 27 recorded simulator images from 9 source groups with Qwen2.5-VL-7B on one RTX 4090 and Jev 1.13.0: image to one-token choice scored 54/54 at a 275 ms median per image, while description to Jev succeeded on 47/54 attempts with a 774 ms median decision stage.

Does the probe prove Jev improves driving?

No. The README states there are no outcome labels, so disagreement is not accuracy, and these numbers establish no driving improvement; the result only directs the project toward occasional recovery decisions.

How does the architecture divide work?

A vision model produces observations, numerical facts and local feasibility come from the controller, a semantic supervisor combines them, and the existing policy or planner executes after host-side observation-age and candidate checks.

How do I run it locally?

With Python 3.10 or newer: clone the repo, pip install -e ., run the unit tests, then use python -m jev_drive demo with a scenario, provider and output path. The core package has no third-party runtime dependencies.

What do I need for the Jev API path?

A TYPESAFE_API_KEY set securely in your environment, after which python -m jev_drive replay can run an authored caption fixture. You can also follow the GPU experiment instructions for your own recorded RGB images.

What failures does it record?

The probe preserves API failures rather than hiding them: the measured run included five HTTP 503 responses, one transport timeout and one response rejected by validation. The client uses finite socket timeouts and no hidden retries.

Jev directory

CommunityJev Demos & Experiments

JevDrive

Research prototype that tests where Jev can help a driving policy choose continue, observe, replan or defer: Jev receives structured text, a separate vision model handles camera images, and a local controller owns vehicle execution.

See highlights Open github.com

Stars: 1
Language: Python
License: MIT
Reading time: 2 min
Published by: Community
Author: alpha-harper-franklin
Last verified: Oct 2026

alpha-harper-franklin/jev-drive

User repository on GitHub

View on GitHub

Jev + autonomous driving: structured decisions, multimodal baselines, recovery research, and measured API diagnostics.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity catalog-import logicrw autonomous-driving research robotics

Works forJev for Decision gating Jev for Realtime control

Used inJev for Robotics & Hardware

Languages: Python
Capabilities: Open sourceRequires API key

Highlights

Research prototype with a live, versioned Jev Choice client, response gate and evaluation utilities implemented; CARLA/BeamNG integration is not yet.
27-image diagnostic on one RTX 4090: the Qwen-description-to-Jev path succeeded 47/54 attempts with a 774 ms median decision stage.
Jev attempts included five HTTP 503 failures, one timeout and one validation rejection; the gate rejects stale observations and unavailable candidates.
There are no outcome labels, so disagreement is not accuracy and the README states the numbers establish no driving improvement.
The deterministic demo is a 2D bicycle/A* contract test with range-limited geometry, not a camera simulator or evidence of real-time performance.

Quickstart

bash

git clone https://github.com/Alpha-Harper-Franklin/jev-drive.git
cd jev-drive
python -m pip install -e .
python -m unittest discover -s tests -v

Watch out

Copy reviewed against the truncated repo README on 2026-10-15; the excerpt stops inside the implemented-features list. The README explicitly says CARLA/BeamNG closed-loop integration, outcome calibration and driving-performance improvements are not yet demonstrated.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

Hacker News: Pick Verify: a single decision at the warehouse pick station catches confusable picks.

Show HN: Verifying robot picks with a decision modelRead the thread on Hacker News

X: A dual-arm robot with Jev in the middle layer

YouTube: Dan Makes Things drives an ESP32 robot with Jev decisions.

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.

Jev Demos & Experiments#gaming#agents#control

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai