Skip to main content

​
Cross-role governance enforcement

Experimental-surface notice. The implementation may merge during 3.2 development. The beta-to-GA period provides the six-week notice window before production availability. These changes are part of the existing governance.campaign experimental feature, including the new adcp.governance_enforcement capability.
AdCP 3.2 makes buyer governance enforceable across service roles without disclosing plan_id to those services.

​
Provider changes

  • Resolve transport credentials to an authenticated agent URL on every governance check. Require caller to match and never grant plan access or authorization from the request assertion alone.
  • Persist the authenticated sync_plans caller as the immutable plan owner. Treat absent delegations as owner-only; delegations add scoped non-owner authority and never displace the owner.
  • Emit the modern check_type response shape. The absent-check_type legacy response arm remains accepted during 3.x migration.
  • Return governance_context only for approved; return non-authorizing consultation_context only for intent conditions.
  • Require task-neutral proposed_commitment for update_media_buy and on intent checks whose task payload carries pricing indirectly; amount 0 explicitly represents a verified no-cost or non-increasing action.
  • Allow purchase execution checks before a durable media_buy_id exists; require the ID for modification and delivery.
  • Bind intent approval to the target service. Require execution checks to authenticate as that preserved audience, keep the intent currency, and narrow rather than widen its authorized amount.
  • Send the exact downstream service URL as top-level target_agent on every intent check. Keep routing metadata outside payload, which must remain the same JSON task-arguments value sent downstream. The JCS hash is serialization-independent; the service adds governance_context and may add context, then removes those two metadata fields before recomputing it.
  • For update_media_buy, have the buyer propose a positive-delta ceiling without exposing seller state to governance. The seller computes the actual positive delta from authoritative current state atomically with its revision check and write, enforces the signed ceiling, and sends execution_commitment on an online execution check.
  • Settle each opaque action binding once, even though intent and execution checks have distinct IDs. Require the outcome reporter to authenticate as the original buyer caller and preserve purchase_type. Derive the ledger amount from governance-owned intent or purchase-execution state, validate every reported monetary value, and treat caller amounts as reconciliation evidence only.
  • Cache identical report_plan_outcome retries by idempotency_key; reject key reuse with another payload.

​
Service changes

  • Advertise task-scoped enforcement only through adcp.governance_enforcement and only for behavior actually implemented.
  • Resolve the account before deciding whether governance applies; return ACCOUNT_REQUIRED when applicability cannot be determined.
  • Verify inbound intent tokens. Media services claiming online_execution_check additionally perform prepare → check → commit.
  • Require critical authorized_task and authorized_payload_hash claims, and recompute the RFC 8785 JCS/SHA-256 payload binding before every governed side effect.
  • Keep cancellations, pauses, deactivations, and pure decreases available; any commitment-increasing part of a mixed update wins over an exemption.

​
SDK changes

Implementation trackers: Regenerate request/response types and update helpers for:
  • typed governance_context on governed tasks;
  • adcp.governance_enforcement.tasks[];
  • SDK-normalized discriminated result types or type guards for modern responses, plus the legacy compatibility arm (the wire schema uses conditional validation and does not itself guarantee a generated union);
  • proposed_commitment;
  • top-level intent target_agent and whole-request-rooted condition paths;
  • execution_commitment for seller-side update_media_buy execution checks;
  • opaque JWS sub continuity rather than sub == plan_id;
  • signed authorized_commitment verification against the service-computed price;
  • critical protected-header recognition for authorized_commitment, authorized_task, and authorized_payload_hash;
  • phase-conditional planned_delivery.media_buy_id;
  • exact outcome tuple and one-settlement semantics.
Validate generated SDK signers and verifiers against the published governance-authorization.json vectors. They cover bit-exact payload hashing and 27 deterministic compact-JWS cases for task, payload, audience, authenticated caller, monetary ceiling, currency, time, replay, signature, and critical-marker behavior. The bundled private component is public test material and MUST NOT be installed as a production trust anchor. Until an SDK release contains this schema bundle, its generated check_governance validators are stale. The training agent deliberately uses passthrough MCP tool schemas with framework request/response validation disabled for this experimental tenant, then validates in the source-aligned handler. Production integrations MUST regenerate or pin a compatible SDK rather than copying that compatibility overlay. Pin experimental integrations to the matching 3.2 schema bundle until the surface graduates.

​
Conformance coverage

The cross-role governance index connects this contract to the executable proof for each service role. It is an index, not one universal workflow: each service storyboard runs only when the agent advertises the exact task and mode in adcp.governance_enforcement.tasks[]. Governance providers and enforcing services remain separate roles. Provider responses prove verdict semantics; service storyboards prove that signed intent is checked before a role-specific side effect. Merely returning or echoing governance_context is not enforcement evidence.