How to use jev, and the exact tasks where 70ms actually beats a chat model. jev doesn't generate text. you send it unstructured state plus a set of typed questions, it returns a probability-scored answer to every question at once, in parallel, in 70 to 500 milliseconds. a comparable chat model doing the same class of task runs 3 to 329 seconds, because it's generating tokens sequentially, one at a time, even when the actual output you need is just a classification or a decision. that gap is not universal. it only matters for a specific category of task, and knowing which category is the entire skill here. where 70ms genuinely beats a chat model: routing a request before it hits your actual llm call. deciding which of five prompts to use, which tool to invoke, which department a support ticket belongs to. these are typed, bounded decisions, and running them through a full chat completion is paying token generation cost for an answer that was always going to be one of a handful of options. safety and content tagging at volume. flagging a message as needing review, scoring sentiment, checking a field against a compliance rule, anything you're currently doing with a cheap llm call purely to get a categorical answer back, wrapped in unnecessary natural language you then have to parse out again. any high frequency decision sitting in a pipeline where latency compounds. a single 3 second chat call feels fine once. the same call run 10,000 times a day, in sequence, in front of the actual response a user is waiting on, becomes the bottleneck in your whole system. where it doesn't help at all, and where you should still reach for a real chat model: anything requiring the response itself to be natural language. jev never generates a sentence. it can't write your customer's email back to them, it can only decide what kind of email that customer's message needs. anything genuinely open ended, where the set of possible answers isn't something you can define as a fixed schema in advance. jev needs typed questions with typed possible answers. a task where you don't yet know the shape of a good answer isn't a jev task. the honest limit worth knowing. bounded, typed output can't hallucinate text, since there's no text. it can still pick the wrong option within a valid schema. this isn't a replacement for verification, it's a replacement for the specific work of deciding, cheaply and fast, which then feeds into whatever actually needs to happen next. the actual mental model. jev is the classification and routing layer sitting in front of and around your real llm calls, not a competitor to them. use it for every decision in your pipeline that's typed and repeated. keep the chat model for everything that still needs to talk.