The Jev decision system: route before the LLM, ~9× smaller context
ChengOS can now put a decision model in front of your agents. Before any large language model runs, Jev reads the user's latest request, picks one of the routes you declared on the canvas, and hands the request to a focused agent that carries only the tools it needs. Anything unclear still goes to your general agent, so no capability is lost.
We ran the go/no-go evaluation on 157 English and Chinese requests (routing-eval-v2):
| Result | |
|---|---|
| Routing accuracy | 156 / 157 (99.4%) |
| Context on tool-needing turns | ≈ 9× smaller (4,277 → 469 tokens) |
| Routing latency | 0.57 s p50 (chat-model routers: 1.4–2.9 s) |
| End-to-end completion | On par with the general agent, 5–9 above Tool Hub |
A few things worth knowing:
- The router is an optimization, not a gate. Chit-chat, ambiguous or multi-intent requests, low confidence, and outages all fall back to the general agent — the worst case is today's behavior.
- Routing grants no permission. Every tool call still goes through the deterministic approval policy.
- Need data to stay on your hosts? Switch to Laya, which runs on your own hardware and answers in milliseconds, at the cost of falling back more often.
- The residual risk is multi-intent phrasing: every model, Jev included, misrouted the same "email the code review results to the team" request.
📖 Full write-up: The Jev Decision System 📊 Interactive evaluation report: Jev semantic routing evaluation (Chinese) 🛠️ Build one: Semantic Routing with Decision Branch
Share your routing scenarios and measurements in the comments.