Jev - A Decision Model

Read

• Introducing System One Models & Jev

  Diogo Almeida’s launch post, 15 September 2026. The names, the training claim (RLCD), the speed and price figures, and the comparison with chat models.

• Jev 1.14

  14 October 2026. Jev 1.14.0, under the alias jev-preview. Images, a Box type, and a longer context, each behind a danger-zone header. The price is unchanged.

• Introduction

  The short version: state and typed questions in, structured answers out.

• System One

  What the model class is for, and how it differs from a language model.

• Primitives

  How to choose Noul, Choice, Score, or Box, and how to ask several at once. The pages for Noul, Choice, Score, and Box have the response fields.

• Confidence

  How confidence is derived from a distribution, and how to use it as the line between acting and escalating.

• State

  What to send as context, and how a question points at a nested field.

• Images

  An image is listed in assets, and a reference to it can sit in the state, in the instructions, or in the criteria. Send TypeSafe-Danger-Zone-Images: true. Jev 1.13 and Jev 1.14 both accept images. A Box question needs 1.14 and the same header.

• How to build with System One

  Keep the workflow in code. Give the model narrow judgments.

• Jev 1.13 jaggedness

  TypeSafe’s own list of rough edges on Jev 1.13. Worth reading before you trust a new question shape. Jev 1.14 is the newer release. This page is still about 1.13.

• Models

  jev-1.14.0 is the 14 October 2026 release. The alias jev-preview points at it. jev-latest still points at jev-1.13.0. Input is $42 per Btok, which is $0.042 per Mtok. Output tokens are free. The default context is 64k tokens per request and 32k per question. The header TypeSafe-Danger-Zone-Long-Context: true raises both to 262k on 1.13 and 1.14. TypeSafe says that path is slower and less accurate.

Patterns

These four are the ones that change how you design the call. The patterns index and the cookbooks go further.

• Speculative fan-out. Ask every question you might need, including ones some inputs will ignore.
• Confidence-gated routing. The answer says what. Confidence says whether to act.
• Composite scoring. Split one vague rating into atomic scores and weight them yourself.
• Intent routing. Send a request to code, to a specialist model, or to a person.

Call it

• Laravel AI SDK

  What this site uses. Boolean, choice, and score classifications. A boolean is the Noul question under Laravel’s name.

• TypeSafe JavaScript SDK

  The official client. noul(), choice(), and score(). The snippets on the Demo page follow it.

• TypeSafe Python SDK

  The same questions as Python objects, synchronous or asynchronous.

• HTTP API

  POST /v1/systemone. The model field selects Jev. The docs’ examples use jev-latest, which still names jev-1.13.0. Box examples use jev-preview or jev-1.14.0. Images need TypeSafe-Danger-Zone-Images: true. A longer context needs TypeSafe-Danger-Zone-Long-Context: true.

• Jev on OpenRouter

  How to call ~typesafe/jev-latest with an OpenRouter key. That is the path this site takes when a key is set. OpenRouter bills the account.

This site

• Demo

  Three questions, recorded answers on the public site, live answers when you bring a key.

• github.com/tibor-src/jev-site

  The Laravel app behind these pages.

• tibor.io

  The person who published the site.