Jev in Practice: Browser Use and Agent Workflows

A developer argues that using Jev to make routing decisions before calling a large model may be its most valuable application.

An eval shows that when evidence is ambiguous, the order of choices becomes an implicit policy for Jev, and the effect disappears once evidence clearly separates the options, so the issue is insufficient state information rather than hallucination.

The author open-sources 2B and 9B decision models that reproduce JEV-style decisions using LoRA adapters and decision heads, releasing code, data and weights.

Breaks down a trading bot's layered architecture: state engine, Jev evaluating narrowly defined questions, policy layer, risk engine and execution layer, noting that stale answers are skipped.

Using four decisions from a failed deploy as an example, explains how Jev's Choice, Score and Noul primitives return typed answers with probabilities in parallel, and stresses that the probability matters more than the answer.

The author says he wanted to cover the basics first before writing an advanced piece on building a custom harness with Jev, and invites questions.

The author shows off a movie classifier he built with Jev and asks what others have built with it so far.

The author says open-ended questions can be built on top of Jev using only existing vLLM machinery like entity/number extraction and simple counts, with similar performance, though it may only work in DiffusionGemma.

The next article in the series will explain how to use Jev to build a System 1 plus System 2 agent harness, with more to be announced in the community.

After Jev launched, the open source community released Von, a 395M-parameter CPU decision model, but the author tried several local Jev replacements and found none matched Jev's capabilities.

Using Jev as the decision layer and Picsart as the execution layer, the author wired 188 models into one pipeline with eight stations like BRIEF, ROUTER, IMAGE and VIDEO to solve character drift across posts.

The author rejects AI doomerism: LLMs have no agency, and a language tool like Jev shows models are just tools that obey their users, so the real alignment problem is making tools follow the user.