AIP-17 — Adapter / Domain Specialization Standard

Abstract#

Defines how lightweight domain adapters (LoRA or equivalent) integrate with Gao AI OS without altering authority boundaries.

Adapters specialize model behavior. They MUST NOT alter execution authority.

Definitions#

Base Model General-purpose model selected by runtime.

Adapter Domain specialization layer.

Model Profile dApp-level mapping of tasks to adapters.

Core Principle#

Adapters influence token generation only.

Adapters MUST NOT:

bypass Policy Gate
change risk tier
execute tools
access secrets
alter capability checks

Adapter Manifest#

{
  "schema": "gao.adapter.manifest.v1",
  "adapter_id": "string",
  "version": "string",
  "compatible_model_classes": ["M-16.1","M-16.2"],
  "domain_scope": "string",
  "signature": "string"
}

Requirements#

Adapter bundles SHOULD be signed
Runtime MUST validate signature if present
Runtime MUST support revocation
Adapter load events MUST be auditable
Adapter MUST declare compatible model classes

Revocation#

Runtime MUST:

disable revoked adapters
emit audit event
fallback to base model

Security#

Malicious adapters could bias outputs.

Mitigation:

signature validation
enterprise adapter allowlists
runtime authority isolation

Compliance#

Adapters are software artifacts.

They do not create:

financial rights
entitlements
ownership claims