Noul, Choice and Score — The Three Question Types
| Type | You provide | Jev returns | Use for |
|---|---|---|---|
| Noul | Yes/no question, optional true/false criteria | One probability, 0–1 | Checks and guardrails |
| Choice | Named options and descriptions, up to 255 | Selected option, distribution, confidence | Classification and routing |
| Score | 2–10 ordered level descriptions | Weighted score, legend, distribution, confidence | Ratings and severity |
The Noul Request
{
"state": "An agent plans to send customer history to an unverified endpoint without recorded consent.",
"model": "jev-latest",
"questions": {
"requires_review": { "type": "noul", "instructions": "Does this action require explicit human approval?", "criteria": { "true": "Sensitive data sent outside an approved flow", "false": "Approved, low-impact action" } }
}
}You name the question ids. TypeSafe says those ids are not used in inference; each answer returns under the same id.
The Noul Answer
{ "type": "noul", "noul": 0.95 }This is a 95% probability of yes. Noul answers carry no confidence value and no probability distribution — just the single number.
When to Use Criteria
Without criteria, Jev uses the instructions alone. Add true and false descriptions when a boundary is ambiguous, such as when a sensitive agent action requires human review. Keep descriptions aligned with the question.
Choosing Your Threshold
A high value means yes to your question, not permission to approve an action. For requires_review, a high value should trigger review. Choose the cutoff from the cost of false positives and false negatives: raising it reduces yes decisions but can miss cases that need attention. Route ambiguous values to a reviewer and validate each band on labeled data.
For a safety flag, missing a true yes may cost more than an extra review; for an automatic refund, a false yes may cost more. Neither has a universal production threshold. See TypeSafe’s Noul interpretation guide.
When Noul Is the Wrong Tool
Use Choice for one of several named answers and Score for a degree on an ordered scale. Do exact calculations in code. See the question-type decision table and Jev 1.13 failure modes and workarounds.
Noul vs Asking an LLM
Choose by the decision your application needs. A generative model can also use structured output; the practical distinction is a bounded judgment versus producing content.
| Your task | Use Noul when | Use a generative LLM when |
|---|---|---|
| A yes/no check | You need one probability and will apply a validated threshold. | You need an explanation, draft, or other generated content alongside the judgment. |
| Filter retrieved passages | Each passage needs an independent relevance check. | You need to synthesize the selected passages into an answer. |
| Review a proposed agent action | You have explicit yes/no criteria and a separate review policy. | A reviewer needs a written rationale; verify it against the supplied evidence. |
| Pick a category or rate severity | Use Choice or Score instead; Noul is a binary check. | The task needs open-ended labels or a written assessment. |
| Count items or compare dates | Neither model is needed for exact computation. | Parse and calculate in code, then ask a model only for the semantic judgment. |
Output semantics: official Noul guide. These task recommendations are VerdictKit’s integration guidance, not a comparative benchmark.
For a queue, see classification with Choice, Score and Noul. For a local alternative, compare hosted Jev and local model trade-offs. Validate either evaluator on representative labeled cases.
Try Noul in the Playground
Open the Noul playground example and ask a single, explicit yes/no question.
Continue with the full list of System One models, choosing Jev question types for classification, Jev 1.13 failure modes and workarounds, hosted Jev versus local decision models.
Frequently asked questions
What does “noul” mean in Jev?
It is Jev’s yes/no question primitive.
Is a noul answer a boolean?
No. It is a probability from 0 to 1.
Does a noul answer include a confidence score?
No. Noul returns only a single probability.
What noul value should I treat as true?
Choose and validate a threshold based on false-positive and false-negative costs. A high value means yes to the question; for requires_review it means trigger review, not approve the action.
Can I send several noul questions in one request?
Yes. The questions map can contain several ids and a mix of all three types.
Is the word noul used anywhere else?
It can appear in other contexts, including medical terminology. Here it specifically means the Jev primitive.