Explainer
Teams say “we use GPT for routing” when they mean “we ask GPT for a label and hope the JSON parses.” That works in a demo. Under load it is latency, spend, and silent schema drift.
For builders, the useful split is:
If the program needs a paragraph, use an LLM. If the program needs a branch, use a decision model.
Category primer: what is a System One decision model.
| Job | Prefer LLM | Prefer decision model |
|---|---|---|
| Write the user reply | Yes | No |
| Pick support queue from 8 teams | Overkill | Yes |
| Multi-hop research then choose | Yes (then maybe decide) | Alone, weak |
| p(destructive) before shell | Expensive | Yes |
| Explain why it routed | Yes | No (no prose) |
| Score urgency 1 to 5 | Possible | Better fit |
| Classify screenshot UI state | Vision LLM or Clef | Clef if typed |
Support desk: Laya or Jev triage → human or LLM draft → send.
Coding agent: LLM plans and edits → Jev model route + tool Choice + risk Noul each step. Details: Jev for coding agents.
Edge trust & safety: Clef-flash Score on Workers → block or queue → LLM only for appeals copy.
Growth inbound: decision-model lead Score → CRM; LLM writes the first email only for high-fit leads.
| Guide | Why |
|---|---|
| Top 10 Jev use cases | Concrete Jev patterns |
| Top Laya use cases | Open-weight patterns |
| Top Cloudflare Clef use cases | Multimodal patterns |
| Jev vs Laya vs Clef | Vendor pick |
See frontmatter FAQ for schema-ready Q&A (also rendered on the live guide page).
3 curated tools below.

Jev is TypeSafe's flagship System One decision model: send application state plus typed Choice, Score, and Noul questions, and get calibrated probabilities your code can branch on, without generating free-form text.
Laya Studio hosts the open-source Laya System One decision model behind a Jev-compatible API: send state and typed questions, get answers with probabilities in one forward pass, plus optional MCP tools and a guard endpoint.

Cloudflare Clef and Clef-flash are open-source (Apache 2.0) System One decision models on Workers AI: multimodal state plus typed noul/choice/score questions return probabilities, with a Jev-compatible API shape and optional fine-tuning path.
| Tool | Best for | Pricing | Billing note |
|---|---|---|---|
| Jev | Productivity | Paid | Paid Service |
| Laya | AI Agent | Freemium | Free Trial |
| Cloudflare Clef | AI Agent | Freemium | Free Trial |
Should I route with an LLM or a decision model?
Use a decision model when the answer set is closed and you need calibrated probabilities your code can trust for format. Use an LLM when routing depends on open-ended reasoning, tool-using research, or generating the next message.
Is structured output from an LLM the same as Jev?
No. LLMs can emit JSON, but probabilities in that JSON are still generated text estimates, and format can still drift. System One models return native typed distributions and cannot emit an invalid choice key.
Can I use both in one stack?
Yes. The usual pattern is LLM for plan and prose, decision model for triage, tool gates, and model-tier selection. Many production agents already look like this even if the gate is a brittle classifier today.
Which decision model should I start with?
Start with Jev if you want a managed TypeSafe API. Choose Laya for open weights or self-host. Choose Cloudflare Clef for Workers AI and image state. See our compare guide.
When is a classic classifier enough?
If labels are stable, English-only, and you have plenty of training data, a small fine-tuned classifier can beat a general decision model on cost. Reach for System One when labels change often or you want natural-language criteria without retraining.