jev
Jev
Turn the user's goal into a working method. Design the necessary judgments, call the bundled Node.js helper, and combine its answers with code to deliver the requested result. The user does not need to choose primitives or start a software project.
Shape the method
Start with the desired output and available evidence. Resolve routine implementation choices yourself. State reasonable assumptions; ask only when an ambiguity in the goal or decision criteria would materially change the result. Preserve the user's scope, chosen tools, and existing authorization.
Keep exact lookups, arithmetic, parsing, date comparisons, and execution in code. Use Jev for the parts that need semantic understanding. It accepts text or JSON and returns typed judgments, not generated prose, code, or reasoning explanations. Convert non-text inputs with an appropriate available tool first.
Choose the question type by what its answer means:
| Need | Primitive | Interpretation |
|---|---|---|
| Select one known candidate | choice |
Selected key, probabilities, and confidence; include a no-match option when appropriate |
| Judge whether a condition holds | noul |
Probability of yes; near 0.5 is uncertainty, not medium intensity |
| Measure a described dimension | score |
Position on ordered descriptive levels, with probabilities and confidence |
Use named state fields and stable source IDs. Write complete questions: question IDs are not sent to the model. Reference relevant fields with backticked paths. Provide concrete, self-contained criteria; split independently useful dimensions while retaining the context needed to judge them. For extraction, find candidate spans first, let Jev select a candidate ID, and copy the original value in code.