Jev in Practice: Browser Use and Agent Workflows

A Claude Code setup where Sonnet implements, read-only Haiku subagents map the repo and docs, and Opus is consulted at three checkpoints; Jev is suggested for deterministic routing if already installed.

A Clairic and Jev workflow pre-reviews pull requests: fetch changes, categorize and cluster them, deep-check selected clusters with luna, and report at line-level granularity, finishing in under 10 seconds when no agent work is involved.

A demo wires Jev into a live second brain running 10,000 choice, score and gate decisions at ~250 ms median for $0.42 total, with the human review gate never auto-releasing posts.

The author spent spare time building a Siri-like personal assistant on TypeSafe's Jev, with a video demo of it running.

An open-source framework demoed in Nanjing uses an LLM for planning and Jev as the decision cerebellum, connecting robot dogs and arms via the MHS protocol and retraining while grasping.

AutoTrust launched a demo hub where you can play Mario, Rubik's Cube and Tetris with JEV-27B, test typed decisions with JEV-9B, and try live News and Shorts demos.

MLflow 3.17 lets you pass TypeSafe's Jev as the model for make_judge(), removing the manual API call and probability conversion inside @scorer; the author documents running it from a Databricks notebook with an OpenRouter key.

Pi 1.4.0 fixes summary truncation, Codemode worker exhaustion and broken sessions from truncated tool calls; models.classify() now accepts images, with TypeSafe Jev supporting vision input.

After weeks of running Jev in his pipelines, the developer concludes it is only good at simple classifications or as a probability handoff that narrows the scope for a real LLM.

A short explainer uses a bubble sheet to show how Jev only picks from preset options, why it is far faster and cheaper, where its zero-hallucination claim comes from, and what independent research found it gets wrong.

The author says early experiments show Jev is good at judging whether an agent is moving in the right direction, and uses it to guide and correct agents.

Pi v1.4.0 lets codemode's models.classify() pass images to Jev and Clef models on typesafe, Cloudflare Workers AI and OpenRouter, and fixes compaction summary failures.