Decision Models

When to use one

Use bounded outputs for routing a ticket, classifying a record, scoring a rubric, checking a policy gate or verifying a proposed action. In an agent hot path, your code can branch on a probability without parsing an essay. Unlike a fixed classifier, these systems let you describe the decision choices in the request; that does not remove the need to evaluate each new task.

Current models

Four profiles, grouped by deployment path—not a performance ranking.

Deployment choices

• Hosted: TypeSafe operates Jev inference; verify account limits and data handling.
• Workers AI: Cloudflare operates Clef inference through its platform.
• Open weights: operate AutoTrust, Clef or Perplexity inference; include hardware, upgrades and monitoring in the cost.
• Hybrid: keep a bounded decision step, then escalate ambiguous cases to a person or a separate reasoning model.

Output contract

Choice selects from named options with a distribution. Score uses an ordered rubric. Noul expresses a yes/no probability. Check each provider’s field names and runtime; an analogous primitive is not proof of wire-level API compatibility.

Define thresholds using held-out examples, including uncertain and out-of-domain inputs. A probability is not permission: enforce authorization and deterministic safety checks outside the model.

What not to use them for

Use a general chat or generative model for essays, creative assets and open-ended dialogue. A decision interface is for a bounded output contract, even when its underlying model also supports generation. It is a component, not a complete automation agent or coding environment.

Benchmark caveat

Vendor evaluations differ in datasets, prompts, hardware and aggregation. They cannot be combined into an AI Tool Finder ranking. We have not independently reproduced them; compare deployment fit first, then test the same workload and failure costs.

Automation & Agents · Agent workflow guide