INTRODUCTION
Introduction
Jev evaluates state against typed questions. Each result has a defined shape, so your application can use it in a branch, a filter or a ranking.
A language model can turn a question into prose, but a software decision needs an answer with a known form. The application has to handle the result, choose a branch and explain what happens next. This reference uses small, named questions to make that boundary visible. Choose the answer type before writing the request.
A System One model evaluates questions against the state you provide. Your program keeps control of the next action: it can route an item, rank a set or ask a person to review an uncertain answer. Treat confidence as separate data. A typed value helps the interface explain the result without turning it into an unsupported guarantee.
TypeSafe primitives
The three public question types describe different answer spaces. Choose a category for a routing decision, a rubric for a score or a statement for a truth estimate. These local descriptions explain the table; the public documentation defines the production API.
| Question type | Goal | Returns |
|---|---|---|
| Choice | Choose an option | Category and confidence |
| Score | Evaluate an ordered rubric | Numeric score and distribution |
| Noul | Evaluate a statement | Truth estimate |
Atomic questions, composed in code
Keep each question narrow enough that the returned value has one clear use. If a request combines several decisions, define separate questions and combine the results in application code.
A support workflow could check urgency, choose a team and identify whether required details are missing. The application can then decide which branch is safe to run. None of the controls on this reference page sends a model request.
Next steps
- Quick start — follow the local setup sequence.
- Score — inspect a rubric and its data states.
- Function calling — connect a typed result to an application branch.
- Python client — inspect the reference member layout.