What a decision engine is
A decision engine answers a bounded question instead of producing open-ended text. You supply the material to judge — a message, a record, a retrieved passage — and a question with a defined answer space. The engine returns a label, a score or a probability. The official Jev introduction describes this as a System One model: fast, typed judgment at a decision point, with generation left to other components.
That definition is narrower than "AI that decides." A planning loop, a rules table and a language model can all influence a decision. The useful distinction is what a component returns and who owns the consequences: a decision engine returns a checkable answer, while your application still owns permissions, side effects and review.
Decision engine vs rules, classifiers and chat
| Component | Returns | Best at | Watch for |
|---|---|---|---|
| Rules engine | Deterministic branches from explicit conditions. | Exact, auditable policy that rarely changes. | Brittle on language no rule anticipated. |
| ML classifier | A label and score from a trained model. | High-volume, stable label sets you can train on. | Label drift when the task or taxonomy changes. |
| Chat or agent model | Free-form text, tool calls or plans. | Prose, code and multi-step reasoning. | Output shape and reasoning are not guaranteed. |
| Decision model | A typed choice, score or yes/no probability for a supplied question. | Judgment over messy text with a stable answer type. | Confidence is not correctness; keep permissions in code. |
These are complements, not substitutes. A common pattern uses a decision model to route or screen, ordinary code to enforce permissions, and a generative model only when the required output is prose.
How Jev exposes a decision
Jev accepts text or JSON state and one or more questions. The documented primitives are Choice (one option from a set), Score (an ordered rubric) and Noul (an independent yes/no probability). Choice and Score carry option probabilities and a confidence value; Noul returns a probability from 0 to 1.
Two properties matter in production. All questions in one request evaluate the same state independently, so a decision that depends on a previous answer needs a later request. And constrained output prevents an unexpected label; it does not make the judgment correct. See What is Jev? for the full request shape and the API and SDK guide for the endpoint.
Open and local decision engines
"Decision engine" is also a product category beyond Jev. Laya documents a non-autoregressive decision engine returning typed choice, score and yes/no outputs; Kev and von document open, local decision models; decider, NanoJev and ollaya cover further variants; and AnyJev documents wrapping an existing LLM in a Jev-style interface. Treat each as an independent project with its own accuracy and hardware evidence; the local alternatives guide compares them.
How to evaluate a decision engine
Measure on your own labeled examples, not on a leaderboard. Build a small set that includes ambiguous, out-of-scope and adversarial inputs, then record the input, the model version, the answer and the eventually correct outcome. Track accuracy, calibration, how often a human had to review, latency and cost together, because improving one usually moves another.
Read the model's documented failure notes rather than assuming uniform competence. Jev's 1.13 guidance recommends keeping arithmetic and date comparisons in code and testing irrelevant context. The evaluation guide explains how to read benchmarks without mixing datasets or versions.
Where a decision engine belongs
- Routing: assign a ticket, message or request to the right queue or handler.
- Screening: decide whether a retrieved passage, document or lead is relevant.
- Tagging: answer independent yes/no questions when several labels can apply together.
- Guardrails: score risk before a deterministic rule decides whether an action is allowed.
Keep the boundary explicit: the decision engine proposes, and ordinary code enforces. A model probability alone should never grant access or move money.
Next steps
Start with one decision and a defined answer space. Read What is Jev?, wire a single question through the API or SDK, and evaluate it against your own examples before expanding. If your input is an image, audio or video, convert it to text or structured fields first: the documented Jev API accepts text, not raw media.