用 Jev 做前置路由，把 Claude 的调用量压下来

Kurt @0xkurt原推发布于 2026.10.23

在 X 上打开原推

346,009 JEV REQUESTS. UNDER $10. THIS IS WHY CLAUDE NEEDS A FILTER. Researchers at Bonn, Lamarr and Fraunhofer IAIS report that run across 37 datasets in a public Jev evaluation. The code and raw responses are available. That's a much more interesting starting point than another screenshot of an AI bill. Here's the setup worth stealing: put Jev in front of the expensive model and give it a narrow decision. Which queue? How urgent? Enough evidence to route this automatically? Jev returns structured answers and probabilities. Your code uses them to split the traffic: → Routine case that passes your tested rules: route it. → Ambiguous case: send the original context and Jev's answer to Claude. → Action needing approval: keep it in review. The expensive brain gets a smaller queue with a reason attached to every handoff. It can spend its time investigating the complaint, resolving the contradiction or writing the response. To try this, take 200 past cases with answers you've checked. Use part to choose your routing rules and a separate part to test them. Count the mistakes as well as the cases that skip Claude. Agreement with Claude alone isn't an answer key. Ask the related questions together: topic, urgency and destination can share one state in one Jev request. Then keep sampling the cases you automatically accepted. Confident mistakes are the ones your dashboard can hide most easily. The paper also reports weak spots. Cheap requests don't make every answer reliable, and your queue can behave differently from a benchmark. The animation shows the routing idea, not a live throughput test. The Jev article below explains the decision layer. Start with one repetitive queue and measure how much of it actually needs the expensive brain.

这条是引用着另一条写的：x.com/thegreatest_sv/status/2102802378378621355