Jev 入门指南：面向结构化判断的 System One 模型及其六类工作流

As you may have seen recently, Jev (a generalist System One model) took Twitter by storm. Many misleading examples are everywhere, but they don't really show Jev's real capabilities. Hence, we decided to write this short beginner's guide to Jev. But what is it exactly, and where can it be used in production? Jev is a model built to make focused judgments. You give it information to consider and define the answers it may choose from. Jev returns a structured answer your code can use, along with probabilities that show how uncertain the answer is. TypeSafe AI calls Jev a System One model, meaning it is designed for quick decisions rather than long reasoning or open-ended writing like other models such as Claude Fable 5.1 or GPT-6. This means that your code stays in control of the workflow while Jev handles one narrow judgment at a time. Give Jev context and a decision Imagine handing Jev a support ticket and a short decision sheet. The ticket is the state, which means the information Jev should read. The decision sheet contains the questions you want Jev to answer. For one ticket, you might ask three questions. Which team should handle it? Does it need human review? How severe is the customer impact? Each question has a predictable answer shape. Choice selects one option, such as payments, account, or frontend. Noul returns the probability of yes, such as a 78% chance that human review is needed. Score places the answer on an ordered scale, such as low, medium, or high impact. You can ask several questions about the same state in one request. Jev answers them separately, so one answer doesn't influence another. TypeSafe's primitive guide shows the full request and response format. Try Jev across six workflows We have prepared six use cases you can test Jev on in our new Jev Playground. Start with the clear support ticket. Then try model routing, an agent tool-call guard, and a smart-home request. These examples adapt TypeSafe's official model-routing and LLM-guardrail use cases and its smart-home speculative fan-out demo. You can edit any state before you run it. You can try different use cases in our interactive Jev Playground here: https://academy.dair.ai/resources/introduction-to-jev Read the result A Choice answer includes the selected option, a probability for every option, and a confidence value. A Score answer includes a numeric score, its ordered legend, a distribution, and confidence. A Noul answer contains only the probability of yes. The interface shows yes and no for readability, but it does not invent a separate Noul confidence value. The full distribution matters when the top answers are close, because a single label can hide that uncertainty. The threshold slider turns one Noul probability into an application policy. Raising the threshold triggers that policy less often. Lowering it makes the policy more cautious. Your code makes the final call. It sets the threshold, applies hard rules, and decides whether to allow an action. Where Jev helps Jev fits repeated judgments with bounded answers. Useful examples include model routing, content triage, relevance checks, tool-risk review, and quality gates around an agent loop. Keep exact checks in code. TypeSafe's Jev 1.13 jaggedness guide warns that the model handles arithmetic, counting, dates, literal wording, irrelevant context, adversarial state, and contradictory criteria unevenly. Treat each probability as evidence that your policy weighs, and keep permission decisions in code. Keep experimenting The playground is open again below with four harder cases. Each one hides something that makes the judgment less obvious, such as evidence that points two ways, an instruction planted in the state, or a request that sounds simple but isn't. Pick any case, change anything in the state, and watch how the distributions move. The best way to learn about this model is testing it out. That's why we built the Jev Playground for you to test it out. Start here: https://academy.dair.ai/resources/introduction-to-jev