Typed Decisions, Not Generated Text: Jev 1.13 on RouterHub

TypeSafe Jev 1.13 is now available on RouterHub as typesafe/jev-1.13 through the /v1/systemone endpoint. Jev is a decision model: you send a state and one or more typed questions, and receive structured answers your application can use for routing, classification, and scoring. It is designed for a different job from a chat model that writes prose.

From an open-ended prompt to a bounded decision

Many AI workflows ultimately need a small, actionable answer. Which team should receive this ticket? Does this message suggest an active incident? How severe is the issue against a defined rubric? A generative model can answer those questions in text, but the application then has to constrain and interpret that text before it can act.

Jev starts with the decision shape. Your request supplies the relevant context as state and defines each question with a type, instructions, and, where needed, criteria. The response returns values under your question IDs. That makes it easier to connect a model judgment to explicit application logic: branch on a category, rank by a score, or send an uncertain case for human review.

Three question types for different decisions

Question type	What it asks	Useful example
Choice	Select one option from a defined set	Route a support request to the right queue
Score	Rate an item against ordered rubric levels	Assess incident severity or lead fit
Noul	Evaluate whether a statement is true, returning a value from 0 to 1	Check whether a message indicates a refund request

Choice and Score also return probability distributions and confidence information. These outputs can help an application choose when to act automatically and when to escalate. The threshold is a product decision: test it against your own examples and the cost of a wrong decision before putting it into production.

The question types can be combined in one request against the same state. A support workflow, for example, could check whether a message describes an outage, choose an owner, and score urgency. Your code can then apply the organization’s own rules to those answers. This keeps the model focused on bounded judgments while the application owns the resulting action.

Where Jev fits—and where it does not

Jev is a useful candidate when the answer space is known or can be expressed as a rubric: triage, classification, prioritization, and other decisions that must feed software logic. It is not a general chat or writing model. If the task is to draft a reply, summarize a document in prose, or produce new content, use a generative model for that step.

Good evaluation starts with precise questions. Define each option clearly, include relevant state, and test ambiguous and adversarial examples. TypeSafe’s Jev 1.13 guidance also calls out limits with counting, arithmetic, date comparisons, indirect instructions, and large amounts of irrelevant context. Use ordinary code for deterministic calculations and keep the model question narrow enough that its answer can be checked.

A new decision surface in your RouterHub workflow

RouterHub exposes Jev through a dedicated System One API with the model ID typesafe/jev-1.13. The request uses state and a map of questions rather than chat messages. For teams already using RouterHub for model access, Jev adds a distinct tool for decisions that need typed outputs. Start with a single, well-scoped workflow, compare results with labeled examples, and set explicit handling for low-confidence or unexpected cases before expanding its role.