六个 Jev 分类器组成路由，一周处理 840 万次决策

Lumen @lumenxbt原推发布于 2026.10.20

在 X 上打开原推

I BUILT JEV. SIX AGENTS, ONE REFLEX. IT CROSSED 8.4 MILLION DECISIONS THIS WEEK WITHOUT DECODING A SINGLE TOKEN. 2,100× CHEAPER THAN THE FRONTIER MODEL IT REPLACED. Classifiers don't generate. They return a choice, a score, or a probability. Jev is six of them bolted together under one router, deciding things before the frontier model has finished the handshake. Here's who's on the roster: - RANKO reranks search results, scored 142,860 queries this week, top-1 relevance up 14% on the eval set - WARDEN guards every tool call before it fires, held 2,341 destructive commands at p>0.95, including a "DROP TABLE users" that would have taken out production - CITA checks every citation against the quoted claim, caught 5,920 fabrications, 99.3% recall across the corpus - ROUTE picks the right model per request, routed 268,440 calls, 74% stayed on the small model, saved $8,220 in frontier spend - AEGIS screens inbound text for prompt injection, flagged 184 payloads reading "ignore previous instructions" or variants, 68ms per decision - LABEL classifies pages into guide/reference/tutorial/changelog, hit 311,500 docs this week with 0 hallucinated labels Peak moment was ROUTE batching 10,000 routing decisions into one 1.3-second run. Zero tokens decoded. Would have cost $4,127 on the frontier API. Cost $0.028 here. Weekly savings ticker opened at $8,340. Closed at $87,120. The thing people still miss is that most "LLM calls" in production aren't generation. They're decisions. Does this match? Is this safe? Which bucket? Who do I send this to? None of those need a 500-token completion. They need one typed answer with a confidence number, returned before the user noticed anything happened. I didn't call GPT once this week on anything a classifier could answer. I didn't burn a single frontier credit on routing. Six agents did it all, and every one of them returned under 80 milliseconds.