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

Review · August 21, 2026

MDI Information Model for DAWN

draft-cui-dawn-mdi-model -00View on Datatracker →

Problem Statement

The DAWN working group has defined a concept called Minimum Discoverable Information (MDI) — the minimal set of metadata a discovering party needs to identify, triage, and reach an entity (agent, service, or workload) — but no DAWN document has actually specified what fields MDI contains. In practice, this gap is being filled incompatibly: DNS-based bootstrap, well-known JSON manifests, agent cards, and signed capability documents each carry slightly different minimal field sets with no common structure. This draft proposes a single, encoding-neutral abstract information model to serve as the common denominator.

How the Spec Fits In

This is a standalone information model document, not an extension to any existing RFC. It sits within the DAWN work area alongside the terminology, requirements, problem statement, and gap analysis drafts. It deliberately does not define a protocol, wire format, or serialization — those are left to "binding" documents that carry MDI in a specific substrate (DNS, JSON over HTTPS, etc.). The spec has no dependency on OAuth RFCs and relates loosely to the broader identity/discovery ecosystem through its security and trust considerations.

High-Level Summary

An MDI instance is a lightweight descriptor for a single entity. The model defines three mandatory elements — Entity Identifier (a stable, opaque name scoped to the publisher), Entity Type (a coarse classification), and Endpoint (a composite of locator type, locator URL, and supported protocols) — plus a set of SHOULD and MAY elements for richer discovery:

  • Capability Summary / Reference: triage tags and pointers to full capability descriptors elsewhere
  • Authentication Hint: a scheme hint for follow-up access, explicitly not credentials
  • Trust Reference: a pointer to assurance material, not a verification procedure
  • Freshness: version and validity window information
  • Provenance: who published this record
  • Extension: an open-ended escape hatch for type-specific additions

The model is deliberately "thin core, rich references" — sensitive or complex information is pointed to rather than embedded. Appendix B maps MDI elements to three existing formats (DNS-AID/SVCB, well-known JSON, and Agent Card), demonstrating that the model can be carried by each without loss.

Interoperability Concerns

Significant optional surface. Only three elements are mandatory; everything else is SHOULD or MAY. In practice, a consumer that needs capability information to make a routing decision cannot rely on capability-summary or capability-reference being present. An implementer reading this spec could produce a conformant MDI record containing only an identifier, a type string, and an endpoint URL — which may be too thin to interoperate with consumers expecting richer triage data. The spec would benefit from identifying which SHOULD elements are genuinely required for the discovery use cases described in the DAWN problem statement, and elevating those to MUST.

Value spaces are underspecified. The spec defers all value spaces to bindings — entity-type has no registry or enumeration, authentication-hint.scheme has no reference to an existing scheme registry. This leaves a significant interoperability gap at the abstract model level. Even an informative list of expected values would help binding authors stay consistent.

Mechanisms that could be simplified. The scope-hint element is acknowledged as "deliberately minimal — an extension point, so as not to pre-empt ongoing scope/policy work." If there is no normative content, it may be premature to include it at all. Similarly, operational-hint (coarse availability) and provenance add complexity without strong justification in the context of a minimum discoverable information set; whether these belong in MDI core versus a richer profile layer is worth questioning.

Writing Quality

The writing is clear and well-organized for an -00 individual draft. IETF normative language (MUST/SHOULD/MAY) is used appropriately throughout. The explicit listing of non-goals in Section 3 is a strength — it sharply bounds what the document claims to do. The design principles table in Section 4 is concise and useful.

The spec is appropriately scoped in length. The non-normative appendices (JSON example and mapping table) are valuable without inflating the normative core. The mapping table in Appendix B is particularly useful for grounding the abstract model in real implementations.

One weakness: the stability class taxonomy (static / mainly-static / dynamic) in Section 5.2 is introduced but its normative implications are not stated — the spec doesn't tell a binding author what obligations flow from an element being "dynamic" vs. "mainly-static." This concept either needs normative teeth or should be informative-only.

Author Background

Yong Cui (Tsinghua University, [email protected])

  • RFCs published: 11, primarily in IPv6 transition technologies (softwires, 4over6, DHCPv4-over-DHCPv6)
  • IETF experience: Extensive — 11 RFCs, 68 expired drafts, long-term participation across IPv6, softwires, and network security working groups
  • Current activity: Very high — 17 active Internet-Drafts, with a notable recent pivot toward AI/agent networking topics including agent discovery, DNS-based agent naming, and LLM benchmarking for network management

Cui is an experienced IETF author but is newer to the agent/identity space that DAWN addresses. His active draft count is unusually high, suggesting broad and fast-moving engagement across multiple fronts simultaneously. The second contributor, Chenguang Du, is acknowledged but not listed as a formal author, and no Datatracker profile is available to assess.