Decisions and confidence

Laya evaluates all decisions in a workflow together in a single model call. Each result preserves the normalized answer and the raw backend answer.

Decision primitives

Primitive	Definition	Normalized value
one_of	At least two named options with descriptions	Selected option name
whether	A yes/no question	Boolean; true when P(true) ≥ 0.5
scale	At least two ordered levels	Fractional, zero-based expected rubric position

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

A scale result can fall between levels. For the four levels above, its value lies between 0 and 3; it is not necessarily an integer label.

Metadata

Rules can read value, probability, confidence, and accepted where supported. Probability distributions and raw answers are also preserved in execution results.

For whether, probability is P(true), including when the boolean value is false. scale.probability is unavailable. Laya’s whether answers do not provide confidence: compare their probability directly instead.

Reading unavailable metadata raises an error. Do not assume all backends provide every field.

Confidence policies

Add an inclusive threshold to a decision:

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 meets the configured minimum. Missing confidence fails the policy. Without a policy, accepted is true.

A policy does not retry inference, make a second call, or automatically route the result. Your flow decides what happens after a failed policy. Choose thresholds using representative inputs; they are not universal quality guarantees.

Model limits

Keep inputs and option lists short. Laya’s English checkpoint has a 512-token context window including question headers, and long input may be truncated. Review the Laya project for model details.