Skip to main content

​
Top-level policy shape

Policies are YAML files. A common layout: 
name: my_policy
version: "1.0.0"
policy:
  environment: prod   # optional context, depending on your setup
  rules:
    - priority: 1
      name: rule_name_snake_case
      when:
        component: retrieval    # optional
        metric: context_relevance
        operator: ">="
        threshold: 0.8
      then:
        action: pass

    - priority: 2
      name: another_rule
      when:
        metric: human_reviewed
        operator: presence
      then:
        action: require_approval
        reason: "Reviewer must acknowledge"

​
Rule fields

FieldNotes
priorityInteger; unique in this policy. Lower = higher precedence among matches.
nameStable identifier for logs and artifacts.
whenCondition on signals: at minimum metric; operator; threshold unless operator is presence.
then.actionpass, block, or require_approval.
then.reasonOptional human-readable explanation.

​
Operators

  • Comparison: >, <, >=, <=, == — requires a numeric signal value for that metric (and component if specified).
  • presence — true if a matching signal exists; value optional.

​
Priorities and “winning” rule

All rules that match are considered. The winning rule is the one with the smallest priority value among matches. That rule’s action drives the policy outcome before cross-policy merge.

​
Validation

Duplicate priorities in one policy file are rejected when the policy is loaded.

​
Next

Signals and rules · validate-contract