Label policy

aegis.policy turns labels into verdicts. Labelling nodes — classification or a model-backed labeler — only describe a request; the policy is where you say what each label means on this route: block it, pause it for a reviewer, or let it through.

guardrails:
  classify:
    pack: aegis.classification
  content_policy:
    pack: aegis.policy
    rules:
      - when: {label: classification, in: [secret]}
        verdict: block
        reason: credentials must not be sent to a model
      - when: {label: classification, in: [pii, financial]}
        verdict: require_approval
        reason: personal data needs sign-off

pipeline:
  ingress: [classify, content_policy]   # label first, then decide

Rules are checked in order and the first match wins; a request no rule matches is allowed. Ships in aegis-gateway-pack-classification.

Rules

Field	Required	Notes
when.label	yes	Key in state.labels — classification for the classification pack.
when.in	yes	Label values that match (a list, or one string).
when.min_confidence	no	0–1. Applies when the labeler also records "<label>.confidence"; labels without a confidence (regex classification) always pass it.
verdict	yes	block, require_approval or allow. An allow rule stops later rules from matching — use it for exceptions.
reason	no	Shown in the verdict and the evidence ledger. Default: policy: <label> is '<value>'.

A missing or malformed rule fails at startup with AEG-CFG, not on the first request.

Same labels, different routes

The labels don't change between routes; the policy does. An internal analytics route might allow financial content that a public chatbot route must block:

guardrails:
  classify:
    pack: aegis.classification
  public_policy:
    pack: aegis.policy
    rules:
      - when: {label: classification, in: [secret, financial, pii]}
        verdict: block
  analyst_policy:
    pack: aegis.policy
    rules:
      - when: {label: classification, in: [secret]}
        verdict: block

providers:
  llm:
    type: fake

routes:
  chatbot:
    provider: llm
    pipeline: {ingress: [classify, public_policy]}
  analytics:
    provider: llm
    pipeline: {ingress: [classify, analyst_policy]}