Refusal Signal

Overview

refusal asks the decision model whether the model's answer declines the request. It asks the refusal preset, "Did the assistant refuse?", with the options compliance and refusal, at the response stage. It reads the answer, so a refusal rule exists only once the model has answered: a response plugin reads it, never a decision's rules.

Key Advantages

• Detects a declined answer without a separate classifier.
• Uses the same decision model, and the same call, as the other response-stage signals.
• A predicate sets the operating point when the default does not fit.

What Problem Does It Solve?

An answer that refuses a benign request is a failed answer even when the request routed correctly. refusal names that outcome so a response plugin can act on it. escalate asks the same question to move on from a candidate that refused (accept.refusal: false).

When to Use

Use refusal when:

• a response_guard should annotate or block answers that decline the request;
• you measure how often a route's model refuses.

Configuration

recipes:
  - name: default
    signals:
      refusal:
        - name: declined
          description: The answer declines the request or refuses to help.
          predicate:
            gte: 0.5
    decisions:
      - name: assistant
        priority: 100
        rules:
          operator: AND
          conditions: []
        modelRefs:
          - model: answer-model
        plugins:
          - type: response_guard
            configuration:
              rules:
                - when: {type: refusal, name: declined}
                  action: annotate

Field	Meaning
name	The rule's name, which guard rules use as {type: refusal, name: ...}
description	What the rule detects
predicate	The least probability of refusal; by default the threshold calibrated for the answering model, else 0.5
deployment	The deployment that answers the rule; else the refusal binding, else the decision model
timeout_ms	The answer's deadline

The question sends the request and the answer to the deployment. vela3-fast and the Vela 3.0 d3 sizes have a trained form of it; every other model answers it zero-shot. See a complete example: config/fragments/signal/refusal/refusal.yaml.