The Scope Creep Automated reviews of new AI-related working group submissions

Update review · November 21, 2026

DAWN Requirements -02, 43 Requirements Cut to 17

draft-king-dawn-requirements -02View on Datatracker →

What changed in -02?

The scope. DAWN now stands for "Discovery of Agents With Names", and workloads are gone from the title, the abstract and the term list.

The requirements went from 43 to 17. The entity classification section was deleted outright. The -01 review said the requirements were so broad that any mechanism could claim compliance. This revision is a direct answer to that.

Discovery is now defined as one narrow step: getting the public Minimum Discoverable Information (MDI) about an entity (Section 1.1). Search, ranking, selection, capability catalogues and everything after first contact are out of scope.

What is new

  • A defined minimum (REQ-PROP-1). The MDI MUST include the entity's type, the services it offers, reachability information, communication protocols and any security parameters. Everything else is optional.
  • By value or by reference (REQ-PROP-2). The MDI MUST be available by value. Anything beyond it MAY be a reference. A mechanism MUST set a maximum depth of indirection and say what happens when it is exceeded.
  • Failure handling for references (REQ-PROP-3). A mechanism must distinguish a reference that does not resolve, one blocked by the client's own retrieval policy, and one that resolves to nothing. An incomplete result MUST NOT be reported as absence.
  • Discovery scope (REQ-DISC-4). Discovery starts from a name or scope the client already knows. The mechanism SHOULD say whether the scope was fully searched or the results were truncated.
  • Three deployment contexts (REQ-DISC-2). Local network, within an organisation, and between cooperating parties over the public Internet.
  • Replay joined poisoning and corruption in REQ-SEC-1.
  • Operator control (REQ-SEC-2). An operator can control which entities are discoverable in a given context.
  • Two honest caveats. Integrity of discovery information does not authenticate the live peer in a later exchange (Section 5). Unless a mechanism says it resists enumeration, the set of discoverable entities should be assumed public (Section 6).

What was removed

  • All five classification requirements. Type survives inside the MDI. Geography, jurisdiction, owning organisation and provenance do not.
  • The responsible party (REQ-PROP-6 in -01). Nothing in the MDI says who operates an entity.
  • Attestation, provenance, policy and risk information (REQ-PROP-12 in -01).
  • No single global trust anchor (REQ-SEC-5 in -01).
  • Auditability (REQ-SEC-7 in -01).
  • Relays, proxies and rendezvous for entities without a routable path (REQ-ARCH-6 in -01).
  • Capacity, version information and schema lifecycle requirements.
  • Human operators as discovery initiators.

Is it still solution-neutral?

Less than it claims. REQ-PROTO-1 now names DNS and mDNS as the protocols to build on.

REQ-PROP-3 reads like DNS response handling. No such name, no data, a truncated answer and a resolver policy block are exactly the cases it separates. The introduction still asks for a mechanism that "benefits from an established trust model".

This is not a problem by itself. The DAWN solution drafts are nearly all DNS-based. The document should say so plainly.

Was the previous feedback addressed?

Addressed:

  • Length and repetition. 43 requirements became 17, and the merged ones read better. REQ-ARCH-1 now covers what took three requirements in -01.
  • Auditability. Removed.
  • Cloud, edge and enterprise support. Removed as a requirement. One sentence survives in Operational Considerations.
  • SHOULD-level classification requirements. Removed with the rest of the section.
  • Capability cards. -01 required discovery of them with no format. -02 only asks that a mechanism SHOULD allow a reference to one, and says the format is out of scope (REQ-PROP-4).

Partly addressed:

  • Requirements too broad to constrain a solution. The MDI and the reference rules are specific and testable. The catch is REQ-PROP-1, which says each discovery mechanism defines its own MDI. Two compliant mechanisms can still disagree on what "type" and "services" look like.

Not addressed:

  • The predictable entry point. Now REQ-PROTO-2, with the same wording. Still no hint whether that means a DNS name, a well-known URI or something else.
  • Trust across administrative domains. The requirement was deleted, not answered. REQ-SEC-1 still requires verifying the origin of discovery information. Trust management is out of scope, so the draft never says what the origin is verified against.
  • The normative dependency on the terminology draft. It got deeper. Four more terms are imported, including MDI. The reference points at terminology -04, which still carries the old DAWN name and an Entity definition that includes compute workloads. Terminology -05 came out the same week, and -07 is current.

New problems

REQ-DISC-2 contradicts itself. A mechanism MUST apply to all three deployment contexts, and then different mechanisms MAY be specified for different contexts.

The use cases draft says the mapping from use cases to requirements "is maintained in the requirements document". It is not here. This draft cites use cases -01 once, in the introduction.

There is no reference to draft-cui-dawn-mdi-model, which defines an MDI. A requirements document that makes the MDI its central concept should point at the draft trying to specify it.

Background of the authors

No change. Daniel King and Adrian Farrel, both of Old Dog Consulting, remain the only authors. Five names were added to the acknowledgements.

Feedback for the authors

  1. Fix REQ-DISC-2. Say whether one mechanism must cover all three contexts or whether the set of DAWN mechanisms must.
  2. Say who defines the MDI. One MDI across all mechanisms is an interoperability requirement. One per mechanism is not.
  3. Say what origin verification is anchored to. If the answer is DNSSEC or the Web PKI in practice, state it. If it differs by deployment context, list it per context, as Section 5 already asks solution authors to do.
  4. Decide on the responsible party. Cross-organisation discovery with no indication of who operates the entity is a gap. Either restore it as an optional property or say why it was dropped.
  5. Update the terminology reference to the renamed version, and add the requirements mapping the use cases draft says lives here.
  6. Drop "solution-neutral" or earn it. Naming DNS and mDNS is fine. Claiming neutrality in the same document is not.