Decisions and confidence

All decisions in a workflow run in one provider call. Each result keeps both the normalized answer and the raw answer.

Decision primitives

Primitive	Definition	Value
one_of	Two or more named options with descriptions	Chosen option name
whether	A yes/no question	true when P(true) ≥ 0.5
scale	Two or more ordered levels	Expected position, zero-based, can be fractional

decisions:
  urgent:
    decide: whether
    question: Does this ticket need immediate attention?
  severity:
    decide: scale
    question: How severe is the customer impact?
    levels:
      - negligible
      - minor
      - significant
      - critical

With four levels, severity.value is anywhere from 0 to 3, such as 1.7.

Metadata

Rules can read value, probability, confidence and accepted where available.

• whether.probability is always P(true), even when the value is false.
• scale has no probability.
• Laya’s whether has no confidence; use probability.

Reading a field the backend didn’t supply raises an error.

Confidence policies

decisions:
  department:
    decide: one_of
    question: Which team should handle this ticket?
    options:
      billing: Payments and invoices
      support: Product problems
    confidence:
      minimum: 0.75
flow:
  - when: department.accepted == false
    do: return
    value:
      route: human_review
  - else:
      do: return
      value:
        route: "{{ department.value }}"

accepted is true when confidence ≥ minimum. Missing confidence fails the policy. Without a policy, accepted is always true.

A policy doesn’t retry or reroute; your rules decide what happens. Pick thresholds from representative data.

Model limits

Keep inputs and option lists short. Laya’s English checkpoint has a 512-token window, including question headers. See the Laya project.