Residency

aegis.residency enforces where a route's model endpoint is allowed to be. It is fail-closed: an endpoint without a declared region is blocked, and a declared region outside the allow-list is blocked or paused for approval.

Configure

providers:
  us_llm:
    type: openai_compatible
    base_url: https://api.example.com/v1
    model: example-large
    residency:                       # documentation of where the endpoint runs
      region: us-east-1
      jurisdiction: US
      source_url: https://example.com/trust/data-residency

guardrails:
  residency_ca:
    pack: aegis.residency
    region: us-east-1                # region of the endpoint this guard protects (required)
    jurisdiction: US                 # required
    allowed_regions: [ca-central-1]  # default: [region]
    require_approval: true           # default: false → block

pipeline:
  ingress: [residency_ca]

routes:
  underwriting:
    provider: us_llm

Region comparison is case-insensitive. The guard's region applies to every route it is attached to — attach different residency guards via per-route pipeline: blocks when routes use endpoints in different regions.

What can and can't be verified

You cannot reliably detect where inference happens: geolocating an API hostname finds an edge node, not the GPUs. Aegis enforces what is declared, and checks what is verifiable:

• Some endpoints encode their region — Azure OpenAI (<resource>.<region>.cognitiveservices.azure.com), Bedrock (bedrock-runtime.<region>.amazonaws.com) and Vertex (<region>-aiplatform.googleapis.com). aegis policy lint compares a provider's base_url against its declared residency.region and reports a mismatch as AEG-POL-005.
• Every run's ledger record includes the config digest, so the declared regions in force for any request are provable after the fact.

Hard guarantees ("data never leaves Canada") belong at the network layer. Pair this pack with egress allow-listing in your firewall or DNS.