Jev in Practice: Browser Use and Agent Workflows

A video shows how wasteful it is to make binary judgments through LLM token generation, and explains that Jev returns typed Choice/Score/Boolean outputs in one forward pass at 70ms with no JSON hallucinations.

The author runs his bot's yes/no checks (weather, image, video, untagged messages) through Jev first: 72.5% accuracy, below the big models, but 300ms and 2.5 cents per thousand decisions, handling 88% of ambiguous messages and cutting the judging bill by more than half, at the cost of poor confidence calibration that requires fitting thresholds on labeled data.

Jev acts as a zero-shot classifier that scores stories from a TinyStories-based small model and suggests revisions; after rewriting, coherence rose from 2.0 to 3.0 and stories gained a complete beginning, middle and end.

After a month of tuning strategy, the author opens beta testing for a trading bot that recently added Jev to decide whether a follower would be in profit, with no guaranteed returns and no autotrading yet.

Robotics company Strike Robot announces it is integrating Jev, using confidence-aware routing for fast decisions and escalating to deeper reasoning only when needed, and explains where Jev sits in its Physical AI stack.

Using JEV to decide whether to click or submit on each page and a small DeepSeek model to generate form answers, the author completed 16 questions on an unfamiliar site in 38 seconds with no manual intervention.

Given the same trades and the same 63% hit rate, two bots end at five figures and zero respectively, differing only in position size; Jev sets the stake using a quarter of the Kelly criterion.

TypeSafe released an agent skill for its System One API that lets Claude Code and other coding agents design typed decision workflows such as Choice and Score from natural-language prompts.

Give Jev the current state and goal, define the enabled events it can trigger as Choices, and Jev decides which transition to take next.

The author asks people to submit their Jev projects and experiments to the community awesome-typesafe repo, which has 370 stars and 56 contributors.

The author open-sourced all predictions and charts from a test of Jev on 11,700 decisions, finding that when it says it is 90% sure it is right 81.6% to 99.8% of the time depending on the task.

A walkthrough of Jev's parallel sampler, the caps on its Choice, Score and Noul question types, the shared 64K and 32K token budgets, the 12x savings from batch calls, and the design contract of giving candidates and letting it choose.