jevsdev — Jev Decision Router

Companies are paying LLMs to generate text for decisions that only need a label. Jev (TypeSafe) answers typed questions with choices, scores, and probabilities — cheap, fast, no prose. This service sits in front of it and adds the part that makes it safe to automate: calibrated confidence routing.

How it works

Send a state plus typed questions. Each answer comes back with a confidence. Above your threshold it routes automatically; below it, the verdict is marked human_review — the cheap classifier decides, the expensive fallback only fires when confidence is low.

POST /decide
{
  "state": { "any": "structured input" },
  "questions": {
    "q1": { "type": "choice",
            "instructions": "Route this request",
            "criteria": { "auto": "clear and low-risk",
                          "human": "ambiguous or high-risk" } }
  },
  "threshold": 0.85
}

Endpoints

• GET /health — service status (key_loaded, endpoint, model, threshold)
• POST /decide — batch typed questions over one state; per-question decision + action
• POST /route — single-choice routing shortcut (auto / human / unknown)

Why confidence matters

Independent benchmarks show Jev matches flash-tier LLM judges at a fraction of the cost and latency — and its confidence score predicts its own errors. A cascade that defers low-confidence verdicts is the documented best use; this router implements it by default.

Notes

• Jev is weak at counting, arithmetic, date comparison, and multi-step questions — keep questions single-step.
• Prompt injection in untrusted input can sway verdicts — sanitize or route such cases to review.