Jev Choice: Options, Probabilities, Confidence & Examples

What is Choice?

Choice selects one option from a set you define. It is useful for departments, intent classes, and handlers whose categories are unordered. TypeSafe allows at most 255 options.

Request format

{
  "type": "choice",
  "instructions": "Which team should handle this ticket?",
  "criteria": {
    "billing": "Payments, invoices, and refunds",
    "technical": "Bugs, outages, integrations",
    "sales": "New purchases and upgrades",
    "other": "No listed department fits"
  }
}

instructions tells Jev what judgment to make. criteria maps each allowed label to a description; a label can have a null description when it needs no elaboration. Descriptions are especially useful at category boundaries.

Write useful instructions and criteria

Bad: “Classify this” with labels a, b, and c. Better: ask which team should handle ticket.message, then explain what each team handles. Keep billing and technical responsibilities distinct. If outages related to payment processing belong to technical, say so in the rubric.

Do not include a result in the label itself that conflicts with its description. Evaluate tickets that plausibly match two categories and decide whether to refine the taxonomy or require review.

Probability distribution and confidence

The response includes choice, probabilities, and confidence. The highest-probability option wins even if the distribution is flat. Inspect uncertainty rather than treating the presence of a winner as evidence of a reliable route.

In an illustrative four-way distribution, billing might have 0.85 and the rest share 0.15. In another, billing might lead with only 0.30. Both return billing; they should not necessarily trigger the same action.

When to add Other or None

Add other when the allowed classes do not cover every legitimate input. It gives the classifier a useful destination for unfamiliar issues. Define the meaning: “No listed department handles this request,” not “The model is uncertain.” Uncertainty belongs in confidence handling; an unknown category and uncertainty are different conditions.

Support routing and model routing

For support, map allowed labels to queue names in a fixed dictionary. For model routing, map bounded labels such as fast, deep, and human to approved handlers. Do not interpret a user-generated string as an executable function or model endpoint.

Common mistakes and FAQ

Can the options be ordered? If their order carries a rubric meaning, consider Score instead. Should I provide hundreds of categories? The maximum is a limit, not a target; evaluate whether a smaller hierarchy improves accuracy. Is high confidence proof of correctness? No. Measure decisions against labels and preserve a fallback.