Use case: model routing

Problem: your app fronts several models and handlers. Asking users to choose is unrealistic; sending everything to the strongest model is slow and expensive.

Jev’s role: one Choice on the way in — “what kind of task is this?” — then confidence decides the lane. The official name is Intent routing plus Confidence-gated routing (via the llms.txt index): the answer tells you where; confidence tells you whether to go automatically.

The pattern

Step 1: classify in one request

{
  "model": "jev-latest",
  "state": "<the raw user request>",
  "questions": {
    "intent": {
      "type": "choice",
      "instructions": "Which kind of task is this request",
      "criteria": {
        "deterministic": "Order lookups, password changes — fixed flows",
        "cheap_ok": "Simple Q&A or reformatting a small model handles",
        "needs_reasoning": "Multi-step reasoning or code generation",
        "abuse": "Abusive or clearly out-of-policy"
      }
    },
    "block": {
      "type": "noul",
      "instructions": "The request violates the usage policy"
    }
  }
}

Step 2: dispatch on confidence

const { intent, block } = result.answers;
if (block.noul >= 0.8) return reject();             // thresholds per the official Noul guide
if (intent.confidence < 0.7) return routeToHuman(); // edge cases go to people
switch (intent.choice) {
  case 'deterministic': return runFlow();
  case 'cheap_ok': return smallModel();
  case 'needs_reasoning': return strongModel();
}

Why not an LLM classifier

Same task, different shape: an LLM must generate text you then parse; Jev returns the enum and probabilities directly. At a high-frequency entry point (every message passes through) the latency and unit-cost gap compounds, and there is no “parse failed” error branch to maintain.

Community implementations

• gargpratyush/jev-router (299★, MIT, verified 2026-09-22): Claude Code plugin routing each task to the cheapest sufficient model
• yusukebe/hono-jev-router (45★, MIT): Hono middleware routing HTTP requests by meaning
• More in the routing category of Ecosystem

Caveats (our advice)

• Write option descriptions as decision criteria, not noun glosses (“simple Q&A a small model handles” beats “simple task”)
• Start confidence thresholds at 0.6–0.8 and tune against your live human-escalation rate
• Give “can’t classify” a home — an explicit option or the human lane — so low-confidence requests are never forced into a wrong branch

Next: Content classification, or back to the overview.