TypeSafe AI发布Jev：专用于结构化决策的模型及Python指南
Tried Jev, a model that returns typed decisions instead of text. Notes + Python guide
Many LLM calls in our apps are really just decisions. Which tool should the agent use? Is this ticket urgent? Should we escalate? We pay for text generation and then parse the text back into a value.
我们应用中许多 LLM 调用本质上只是决策。代理该使用哪个工具？这张工单是否紧急？是否需要升级处理？我们为文本生成付费，然后再将解析后的文本转换回数值。
Last week TypeSafe AI released Jev, a model built only for this. You define the question and the allowed answers (pick one option, a yes/no probability, or a score on a scale), and it returns a typed answer with a probability for every option. No text, no parsing.
上周 TypeSafe AI 发布了 Jev，这是一个专为这一场景构建的模型。你定义问题和允许的答案（选择一个选项、是/否概率或量表评分），它会返回带有每个选项概率的类型化答案。无需文本，无需解析。
I tested it from Python and wrote a guide on my company blog: https://mljar.com/blog/jev-python/
我从 Python 端测试了它，并在公司博客上写了一篇指南：https://mljar.com/blog/jev-python/
My main takeaways: it can't answer outside your options, but it can still pick the wrong one, so the probabilities are the really useful part. You can send many questions in one request and they run in parallel. On a simple ticket classification, it gave the same answer as GPT-5.4 nano with structured outputs, was a bit faster, and cost about half as much. That was one run, so it's not a benchmark.
我的主要心得：它无法回答超出你设定选项的问题，但仍可能选错，因此概率部分才是真正有用的。你可以在一个请求中发送多个问题，它们会并行运行。在简单的工单分类任务中，它与 GPT-5.4 nano 配合结构化输出给出了相同的答案，速度稍快，成本约为后者的一半。这只是一次运行结果，因此不算基准测试。
Has anyone tried it for agent routing or for judging LLM outputs? I'm curious how it handles messy inputs.
有人用它做过代理路由或评估 LLM 输出吗？我很好奇它如何处理混乱的输入。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力