Skip to main content
AdCP 3.3 adds the cache_namespace conformance contract. The experimental breaking/security notice published September 13, 2026 had an October 25, 2026 notice floor, which has passed. The contract did not ship in 3.2 and is not retroactive conformance for 3.1 or 3.2. Published release artifacts remain unchanged.
Version 3.1.22 already shipped two emergency privacy/security corrections:
  • #7396 prohibited Context Match embeddings derived from non-public content attributable to one user or session, including through an intermediate summary. Keywords and summaries require documented privacy reduction.
  • #7397 replaced placement-only response caching with {provider_id, context_hash}.
Those emergency protections remain in force. The new evaluation namespace extends the request partition rather than replacing it.

Immediate safe operation

An affected implementation MUST bypass or disable cache lookup, in-flight request coalescing, and insertion until both of these controls are implemented:
  1. Request partitioning by the RFC 8785 JCS/SHA-256 context_hash.
  2. A trusted cache_namespace for result-affecting security and provider configuration contexts, with the lifecycle and warm-hit authorization controls in the TMP caching contract.
The router-owned inputs include authenticated principal or tenant, authorization and entitlement revision, active-package generation, endpoint replacement, provider configuration, and targeting/rules generations. Bypass caching when the required trusted context or generation cannot be established, or generation change and invalidation cannot be coordinated. Never accept a namespace supplied by the caller.

Cache-key migration

For AdCP 3.3, partition each provider’s cache by:
Hash the validated provider-forwarded request after removing only $schema and request_id; preserve array order and every response-affecting input. Use RFC 8785 JCS and SHA-256. provider_id stays the stable registration identity, not a rotating generation. Select the namespace from trusted authentication, authorization, deployment, and provider-configuration state. It must not contain raw principal or tenant identifiers, derive from credentials, or incorporate viewer identity, user tokens, or Identity Match state. Treat namespace mappings, hashes, and cached responses as sensitive request data; keep them out of logs, metric labels, and traces. Capture one evaluation context for lookup, coalescing, outbound endpoint and authentication selection, provider evaluation, and insertion. Do not relabel an old in-flight result into a new namespace, insert it after its context changes, or reuse a generation while old entries or work survive. Authenticate and authorize the current publisher principal and property before lookup and again before returning a warm hit.

Provider-owned model changes

A model change that can affect output is a targeting-configuration change under the existing cache_ttl: 0 SHOULD. Providers SHOULD lower their advertised cache_ttl ahead of a planned model deployment that affects output. Routers are not required to observe internal provider deployments. A warm entry cached before the change may still be served until its TTL expires: at most the cache_ttl advertised for that entry, 5 minutes by default and at most 86400 seconds. A later cache_ttl: 0 cannot evict an already-cached entry. Lowering TTL ahead of deployment keeps this stale window short; it does not retroactively shorten an older entry’s TTL.

Validation cases

These cases are router conformance requirements, not a claim that this specification repository implements a router cache.

Single-user context migration

Do not forward embeddings derived directly or indirectly from non-public single-user content. Apply aggregation, thresholding or another documented privacy-reduction method to keywords and summaries, and remove or reprocess stored single-user derivatives before using them in new requests. Neither namespace isolation nor hashing makes prohibited private context safe.