Jev Question Types: Choice, Score, and Noul Explained

TypeSafe AI's Jev answers exactly three kinds of questions: Choice (categorical), Score (numeric/rubric), and Noul (yes/no probability). Every schema you define for Jev has to fit one of these three shapes — there's no free-text output type. Understanding what each one is for is most of what you need to design a Jev integration.

What is a Choice question in Jev?

Choice is categorical classification: you give Jev a fixed list of possible labels and it returns one (plus, per TypeSafe's positioning around calibrated probabilities, likely a confidence signal on that pick — check the official docs for the exact response fields, since we haven't been given the precise schema shape). Example: routing a support ticket into one of {billing, technical, account, other}, or classifying a log line's severity into one of a defined incident taxonomy. The category list can be as small as two options or as large as 255 — that's the hard cardinality cap on any Choice schema.

// Illustrative only — confirm exact field names against TypeSafe's official docs
{
  "type": "choice",
  "options": ["billing", "technical", "account", "other"],
  "state": "Customer says their invoice charged twice this month..."
}

What is a Score question in Jev?

Score returns a numeric or rubric-based value rather than a category — useful when the answer is a magnitude, not a label. Example: scoring an invoice 0–100 for fraud risk, or rating a customer service transcript against a defined rubric. Score questions are where Jev's calibration training (see RLCD explained) matters most, since a score is only useful if it's honestly calibrated to the underlying likelihood or magnitude rather than just monotonically ordered.

What is a Noul question in Jev?

Noul is Jev's yes/no question type, and the name signals what makes it different from a plain boolean: instead of a hard true/false, TypeSafe positions it as returning a calibrated probability that the answer is "yes." Example: "is this transaction fraudulent?" as a Noul question should return something like 0.83, not just true — a probability you can threshold, log, and check for calibration over time. See calibrated confidence scores for what makes a 0.83 trustworthy versus decorative.

Worked example: routing three questions through Jev

A single customer-service message might generate all three question types against the same input state:

Input state: "Customer's third message this week about the same
             shipping delay, tone is frustrated, order value $340."

Choice  -> category: "escalate_to_human" (from a fixed routing taxonomy)
Score   -> urgency: 82 (0-100 rubric)
Noul    -> is_at_risk_of_churn: 0.71 (calibrated probability)

Three typed answers, in one low-latency call (70–500ms end to end, per TypeSafe's vendor benchmark), instead of parsing a paragraph of generated text into three separate values.

Why does the 255 cardinality cap matter?

If your categorical list can grow past 255 options — a large product catalog, a fine-grained taxonomy — a single Choice question can't represent it, and you'd need to restructure the problem (e.g., a hierarchy of narrower Choice questions, or a pre-filtering step) rather than assume Jev scales to arbitrary category counts. This is a hard architectural limit, not a soft default you can raise via configuration, as far as the facts we have confirm.

Common pitfalls

A few mistakes come up repeatedly when designing a Jev schema:

• Trying to squeeze free-text generation into a Choice question by making the "options" open-ended — that defeats the purpose and likely isn't supported.
• Treating a Noul probability as a hard boolean without picking and documenting a threshold.
• Assuming Score questions are always 0–100 — the scale should be whatever your rubric defines; confirm the exact response shape in TypeSafe's docs before building against it.
• Hitting the 255-category ceiling on a large taxonomy and not finding out until production.