Jev in Practice: Browser Use and Agent Workflows

A 10-minute breakdown shows how Jev returns typed decisions with calibrated probabilities instead of text, covering the Choice, Score and Noul question types plus measured speed and cost.

A developer says wiring Jev into the Temporal Agent Harness auto mode fit naturally along existing harness seams and greatly simplifies agent oversight.

The author argues Jev separates decision from generation and offers a training method for generalizable co-processing models, suggesting a frontier LLM could coordinate several Jev-like models for robot decision-making.

The author uses Jev as a front-end router to decide whether a task needs a large model; on 60 hard problems the setup matched a flagship model's accuracy at about 1/13 of the cost.

The author released an accessible demo site with four examples — spam classification, an NVIDIA buy-or-sell call, an agent tool-call guardrail, and inbox triage — showing how Jev fits into workflows.

Jared Palmer released Kev-0.6B/4B/8B, Qwen3-based models with LoRA and a pointer head that replicate Jev-style decision models, usable with the TypeSafe SDK via a base_url change; Kev-8B reaches 79.6% out of domain.

Jev founder and ex-OpenAI researcher Diogo Almeida explains why RLHF is no longer the future, how modern LLMs will be built, and outlines Jev's core idea and the path to becoming a Jev Engineer.

The author benchmarks Bonsai 2 27B against Google Gemma on Browser Use tasks on an 8GB MacBook and asks whether Jev should be benchmarked next.

Tested on a real fraud ring, Jev reached 93% recall, ahead of the 90% velocity rules written by a data analyst agent and the 62% of an LLM making the classification decision.

The author benchmarks Jev as a jailbreak and prompt injection classifier: it beats off-the-shelf guardrails including Meta's on newer attack sets, but loses on older sets and at tight false-alarm budgets, and fails on obscured internal attacks.

After looking into Jev, the author concludes it cannot handle signal calculation, position sizing or order execution, but fits as a text event classifier and human-confirmation gate around a trading system, with other uses in support triage, content moderation and agent safety layers.

The author says most people use Jev wrong: its strength is labeling situations or data, not acting as a universal fix.