What is a System One decision model?

Key takeaways

• A System One decision model answers typed questions with probabilities, not chat paragraphs.
• Three common question types: Choice, Score, Noul (yes/no).
• Built for routing, moderation, reranking, guardrails, metadata tagging inside software.
• Not a replacement for ChatGPT, Claude, or other generative models.
• Directory examples: Jev, Laya, Cloudflare Clef.

Table of contents

1. Simple definition
2. What people mix up
3. How a request looks
4. Where they fit in a stack
5. Who should care
6. FAQ

---

Simple definition

The term is already overloaded. Some people hear “decision model” and think of any classifier. Others think it is a tiny LLM.

For builders, the useful definition is the simplest one:

A System One decision model reads application state and a schema of typed questions, then returns a probability for every allowed answer so your code can branch without parsing free-form text.

What that means in practice:

• What it is: a fast judgment engine with a closed answer set.
• What keeps it useful: question design and thresholds you choose in code.
• What it is not: a chatbot, a summarizer, or a general reasoning replacement.

A simple way to think about it: the LLM proposes; the decision model adjudicates.

Last verified: October 2026.

What people mix up

Thing	Job	Output
Generative LLM	Write, plan, tool-call	Free-form text / tools
Classic classifier	Fixed labels you trained	Class id / score
System One decision model	Bounded questions over arbitrary state	Probabilities per option
Always-on agent (Dots, Muse)	Act across apps over time	Actions + messages

Classic classifiers need retraining when labels change. Decision models in this family are marketed as handling new Choice sets without a fine-tune for every category. Validate that claim on your labels.

How a request looks

Pattern shared across Jev-compatible APIs:

1. Send state (ticket text, JSON record, conversation snippet; some models also accept images).
2. Ask one or more typed questions (often up to dozens per request).
3. Get back answers + probabilities (and often confidence) keyed by question id.
4. Branch in code: route, block, escalate, rank, tag.

You still own thresholds. A 0.62 “yes” is not a policy. Your product decides whether 0.62 is enough.

Where they fit in a stack

Typical placements:

• Before an expensive LLM call (should we even spend tokens?)
• After retrieval (which chunk is relevant enough?)
• Inside agent loops (allow / ask / block a tool call)
• Beside moderation (score severity, pick a queue)

Wrong placement: asking a decision model to “write the customer email.” That is generative work.

Product pages in this cluster:

• Jev (TypeSafe, hosted)
• Laya (open-source lineage, hosted Studio + MCP)
• Cloudflare Clef (Workers AI + open weights, vision)

Who should care

Yes if:

• You ship routing or guardrails in production code
• LLM JSON parsing is a reliability tax you want to delete
• Cost and latency of frontier chat models hurt hot paths

No if:

• You only need occasional chat answers
• Your team will not design questions or thresholds
• You need open-ended explanations as the product output

Next reads: Jev review, Laya review, Cloudflare Clef review, Jev vs Laya vs Clef.

FAQ

What is a System One decision model?

A System One decision model takes application state plus typed questions (yes/no, choice, or score) and returns probabilities for each allowed answer. Your code branches on those probabilities. It does not generate free-form text.

How is a decision model different from an LLM?

LLMs generate open-ended text and tool calls. Decision models answer bounded questions with calibrated probabilities. Use decision models for routing, scoring, and guardrails; use LLMs when you need writing, planning, or open-ended reasoning.

What are Choice, Score, and Noul?

They are the three common question types. Choice picks one option from a set. Score places an item on an ordered rubric. Noul is a yes/no (or true/false) probability. Names come from the System One / Jev-style API.

Is Jev the only decision model?

No. TypeSafe’s Jev popularized the pattern. Compatible or similar options include Laya and Cloudflare Clef (plus other open-weight hosts). Compatibility of wire formats varies; check docs before swapping.

When should I not use a decision model?

Skip them for open-ended writing, creative generation, or tasks where the answer set cannot be bounded. Also skip if you need rich chain-of-thought explanations as the primary output.