A field guide to System One models.

A model that returns a decision

TypeSafe AI uses “System One Models” for models built to return typed decisions inside software. Jev is its model in this category. Laya is a separate open-source project that describes itself as a System 1 decision engine.

For this atlas, the practical question is simple: can a model turn an input into a useful decision with an output your application already knows how to handle?

Three useful outputs

These are example workflows, not measured results. A typed output can make integration simpler, but a valid label can still be the wrong label.

Laya and Jev are separate projects

Laya offers a local, open-source path. Jev is offered by TypeSafe AI through hosted access. Similar goals do not establish identical architectures, compatible APIs, or equivalent performance.

Start with a bounded decision

A useful first experiment has a small output space, examples you can label, and a clear consequence for a mistake. Ticket routing is often easier to specify than an open-ended instruction such as “decide what this customer needs.”

• Define the labels. Make the boundaries between categories explicit.
• Include exceptions. Decide where ambiguous, empty, or unrelated inputs should go.
• Keep a fallback. Give uncertain or costly decisions a review path.

If your main task is writing a reply, creating a report, or synthesizing a long explanation, you still need a component designed to generate that content.

Structured does not mean correct

Evaluate wrong decisions, not just malformed responses. Measure minority classes, ambiguous cases, and the cost of sending work to the wrong place.

A confidence value is useful only when its behavior is understood on your data. Do not adopt an arbitrary threshold because a demo looks convincing. Check calibration and tune an escalation policy on a separate validation set.

Choose your next step

Published evidence & practical details