上手一小时：Jev 无解码循环的决策机制拆解

Jason Goldberg @betashop原推发布于 2026.09.20

在 X 上打开原推

here's how jev works, simplified. an hour in. it is extremely good at making decisions when you give it the right references and context. the actual shift: there's no decoder loop. it isn't sampling tokens one at a time and hoping the last one closes your json. the output space is bounded before inference starts — you declare the answer type, it scores over that set, done. you hand it three things: > the information > your questions > the shape of each answer — a choice from your options, a score on your scale, a probability that something is true what comes back is already typed. no parse step. no retry when it wraps the json in a code fence. no schema drift on the third call. information + questions + allowed answers → a value your code branches on the rest is yours. > route it >> check an agent's work >>>flag it for a human. the hallucination thing is a side effect of the architecture, not a safety feature. there's no paragraph, so there's nowhere to invent a citation. it can still be wrong. it just can't be wrong in prose. and the cost profile changes what you're willing to ask. at a fraction of a cent and sub-second, a decision stops being something you batch overnight and becomes something you put in the hot path. for users with quant agents, we point it at two: 1️⃣ every user interaction. stuck or not, which layer it broke on, does it need a human. 2️⃣ every trade. size within the rule, stop present, same mistake as last tuesday. both are decisions, not essays. neither needs a paragraph. ➡️ take one decision your app already makes, declare its answer type, hand it over.