Reference · independent
Choice, Score and Noul
Every decision model quoted on this site answers variations of three question shapes. Choice selects from a closed set, Score returns a level on a scale, and Noul is a two-option gate. The caller declares the legal answers before the call, which is why the response can be a typed value rather than text.
The names come from the Jev / System One request shape and are reused — sometimes with small schema differences — by Clef, Laya, and other entries on decision-model gateways.
Why three shapes
A decision is only constrained if the caller can say what the legal answers are. Three shapes cover most of what applications actually branch on: a label, a number, or a boolean. Anything outside those — a paragraph, a list of unknown length, an explanation — belongs to a generative model.
Defining questions at request time is the part that separates this from a classic classifier. You do not ship a trained artifact per taxonomy; you change the option IDs in the request.
The primitives
| Primitive | You supply | Returns | Typical use |
|---|---|---|---|
| Choice | A closed set of option IDs, usually with criteria per option | One selected option plus probabilities | Intent, routing, tool selection, classification |
| Score | An ordered scale or a numeric range | A level on that scale | Urgency, relevance, quality, similarity |
| Noul | A condition phrased as a yes/no question | A two-option value with a probability | Approval gates, escalation flags, circuit breakers |
Cloudflare documents 2–255 options for Choice and 2–10 levels for Score on Clef. Other surfaces publish their own limits; check the schema before sending large option sets.
One state, several questions
All three share the same state, so one evaluation produces three answers your code can branch on independently. This is where the cost and latency advantage over a chain of chat completions is most visible.
{
"state": "Customer says the login page keeps timing out after an upgrade.",
"questions": {
"intent": { "type": "choice", "options": ["bug", "billing", "how_to", "other"] },
"needs_human": { "type": "noul", "instructions": "Should this be escalated?" },
"urgency": { "type": "score", "range": [1, 5] }
}
}A naming note: TypeSafe, typesafe, and type safety
Queries such as "typesafe jev ai" usually refer to TypeSafe AI, the company behind System One — not to TypeScript or static type safety. That said, the name is apt: the answer is typed in the ordinary sense that your code can branch on it without parsing prose.
This directory is an independent third-party catalog and is not affiliated with TypeSafe AI. Everything numeric quoted here is attributed to a first-party or named source.
Frequently asked questions
What does "noul" mean?
+
In the Jev / System One request shape it is the two-option question type — a yes/no style gate. This directory uses it for decisions such as "needs human" or "block this request". Other vendors reuse the name under similar schemas; check their docs for exact field names.
How many options can a Choice question have?
+
It depends on the surface. Cloudflare documents 2–255 options for Clef. The TypeSafe-side limits are in the official API docs, and large option sets have known accuracy effects: Laya reports a sharp drop on Banking77 because each candidate shares a fixed token budget.
Can I ask Choice, Score and Noul together?
+
Yes. Several questions of different types can be asked against one state in a single request. Cloudflare documents 1 to 64 questions per request on Workers AI.
Is this the same as function calling or tool schemas?
+
No. Function calling asks a generative model to select a tool and produce arguments as text shaped by a schema. A typed decision call scores a fixed answer space and returns a value with a probability. It does not produce arguments, prose, or a token stream.
Is "typesafe" here about TypeScript?
+
No. It refers to TypeSafe AI, the maker of System One / Jev. The typed part of the request shape is a coincidence of naming that happens to describe the output well.
Keep reading
Sources
- TypeSafe AI — documentation ↗
- Cloudflare — Clef question types and limits ↗
- Laya on PyPI — Choice, Score and Noul primitives ↗
Reviewed 2026-11-13. Vendor- and project-reported figures are labelled as such; none were independently reproduced by this site.