Domain Signal

Overview

domains routes on the request's subject area, one of 14 that the decision model chooses. Each rule names the subject areas it matches. It asks the domain preset: "Which subject area is this request about?", over biology, business, chemistry, computer science, economics, engineering, health, history, law, math, other, philosophy, physics and psychology.

Key Advantages

• Routes by topic without a keyword list per topic.
• One question per request answers every domain rule.
• Groups several subject areas under one route name.
• Feeds static model scores per domain.

What Problem Does It Solve?

Routes are often organized around subject areas: a math model, a legal reviewer, a coding assistant. Keyword lists for each are brittle and miss paraphrases; the decision model reads the request and names its subject.

When to Use

Use domains when:

• routes are organized around subject areas;
• lexical matching is too brittle for the topic boundary;
• the same topic boundary should feed several decisions.

For subjects outside the 14, write a systemone choice of your own.

Configuration

recipes:
  - name: default
    signals:
      domains:
        - name: math
          description: Mathematics and quantitative reasoning.
        - name: stem
          description: Science and engineering.
          labels: [physics, chemistry, biology, engineering]
        - name: computing
          labels: [computer science]
          predicate: {gte: 0.6}
    decisions:
      - name: stem-route
        priority: 100
        rules:
          operator: OR
          conditions:
            - {type: domain, name: math}
            - {type: domain, name: stem}
        modelRefs:
          - model: science-model

Field	Meaning
name	The rule's name, which {type: domain, name: ...} conditions use
labels	The subject areas the rule matches; by default [name], so a rule named after a subject area needs none
predicate	The least probability of the chosen subject area; by default the model's calibrated threshold (0.41 on vela3-fast), else 0.5
model_scores	Per-candidate scores the static selector ranks by for requests in this domain
deployment, timeout_ms	The deployment that answers, and the answer's deadline

A rule matches when the subject area the model chooses is one of its labels and its probability meets the predicate. A rule whose name is not a subject area must set labels; the configuration is refused otherwise, with the closest subject area suggested. A request about none of your rules' subject areas matches no domain rule: give a rule labels: [other] to catch the requests the model files under other.

The model that answers

The decision model answers every domain rule, unless the domain binding or the rule's deployment names another; see Choose a deployment. vela3-fast, the Vela 3.0 d3 sizes and Vela 2.0 have trained forms of the question; every other System One model answers it zero-shot. The question reads the current user turn, truncated to the model's input.

Through the Router, the routing preview (POST /api/v1/routing/preview) reports the matched domain rules. See a complete example: config/fragments/signal/domain/mmlu.yaml.