How it fits the quilt
Agent discovery is a quilt of registries: the NANDA Index and AgentFacts alongside the Agent Name Service (ANS), DNS-AID, DNS-anchored discovery, and any registry implementing the registry-assisted resolution profile — so an agent can be found across many sources rather than one. That breadth is the strength, and the exposure: the more sources point at an agent, the more places a wrong answer can enter, and no single record proves the others agree.
Corroborated Discovery spans that quilt and proves it. Each registry is reached as a source
(sm-bridge for the substrates it speaks; the known-subject resolver in
sm-divergence for any POST /resolve registry), and corroboration asks
them all the same exact-lookup question and checks that their answers hold together. It does not
replace discovery; it makes the quilt self-checking.
The internet of AI agents
│
├─ Resolution Registries · NANDA Index · ANS · DNS-AID · direct DNS the quilt of registries
│ each resolves ONE known subject to ONE terminal target "where is this agent — per source"
│
└─ Corroborated Discovery the proof layer
"do the sources agree — or is one lying,
or merely behind?"
The registry points; corroboration proves. Each registry resolves a known subject to a terminal target; corroboration decides whether to believe the set of them.
Resolution and corroboration
Registry-assisted resolution of a known subject — the DAWN-aligned profile — is
narrow on purpose. A discovering entity that already knows a Subject Identity (an agent name, an
email, a DID) and has no usable DNS discovery anchor for it asks one registry it selected
by local policy, and gets back the minimum needed to continue: a terminal target (an Agent Card, a
descriptor, a subject-authorized gateway — never another registry), subject-specific authority
evidence separate from the registry's own provenance, and a renewable binding with a
sequence, an expiresAt, and revocation information. No search, no
ranking, no global root, no registry-to-registry forwarding, no trust evaluation.
That profile leaves one thing open — what a client does when it holds bindings from several registries — and names a transparency mechanism for conflicting bindings as an open question. The Quilt is that mechanism, and the two compose without either widening its scope:
Resolution is a claim; corroboration is the diff
A Resolution Registry is a source. Each binding it returns is one classified claim. A Corroboration Record is what a client — or an auditor acting for many clients — produces from two or more of them.
The registries never talk
Corroboration is done by the party holding the claims. The profile's ban on inter-registry forwarding is preserved exactly: it is the sweeper that queries each configured registry directly, one exact lookup per source per subject, and enumerates nothing.
Lag is attributed, not accused
The binding's sequence, counted within its authority scope, is read by the diff. A registry one step behind is stale_binding; its superseded target is withheld from the field comparison. Equal sequence, different target: that is the real disagreement, and it still fires.
Authority evidence does not move the verdict
The profile separates registry provenance from subject authority and leaves acceptability to local policy. The Quilt carries the issuer as the scope a sequence counts in and as provenance for the record. A corroborator that preferred the "more authoritative" claim would be doing the trust evaluation DAWN excludes.
The downgrade detector
Where a validated DNS-anchored path exists, the profile says it stays direct and a resolver must not silently downgrade to a weaker registry result. Treat the direct path as one source beside the registries, and a registry that diverges from it is a finding attributing the registry — the detector that rule otherwise lacks.
Specified in -02. The relationship is stated normatively in
Section 12 of draft-chandra-agent-registry-corroboration-02; the supersession rule is
Section 6.4, implemented in sm-resolver 0.7.0; the known-subject resolver and the
digest field are sm-divergence 0.12.0.
The drafts →
The method
A sweep takes a watch set of subjects and, for each, queries every configured registry the same way, from every configured vantage. Each answer is classified — the source serves the record, positively denies it, or fails to answer (which is no claim, never a denial). The classified claims are then compared:
Omission
A subject one source serves and another confirms absent. An unreachable source is excluded from the comparison — a timeout is not a claim that a record is gone.
Endpoint / identity divergence
Sources agree a subject exists but disagree on a field — the endpoint it serves, the integrity digest of what that endpoint serves, or the key identity (DID) a valid attestation binds. Tampering, host substitution, and equivocation surface here.
Stale binding
Sources agree a subject exists, but one serves a binding whose lease sequence is behind the current one in the same authority scope. It is attributed as behind, and its superseded target is withheld from the field comparison — lag is not a lie.
Source equivocation
One source's own answers, seen from different vantages, contradict each other — the tell that a registry is giving different callers different answers.
Declared divergence
Not every difference between two registries is a lie. A registry under a data-residency obligation serves a different endpoint in one region than another, by design and often by law. A procedure that reports every such difference as equivocation fires during normal operation, and a detector that fires during normal operation gets switched off.
So a source may publish an Answer Scope: a signed statement, made in advance, of which fields it varies and along which context. Three properties keep that from becoming an exemption.
What can be declared
“Endpoint varies by region” is checkable: a verifier observes from another region and compares. “EU callers get this address” would be a promise about values, made by the party doing the varying, and adds a label to a finding without adding any way to test it.
What cannot
The context must be one a third party can construct and observe from. Region and network qualify. Caller identity, tenant and subscription tier do not — a declaration about one could only be believed, and a corroborator crediting it would be repeating the source's own assertion back as evidence.
What it changes
The finding is still raised and the verdict is unchanged. The declaration marks it, and records which revision was matched. Whether a declared difference is acceptable is the consumer's judgement, not the corroborator's.
Why it never removes a comparison. A source that disagrees with itself normally contributes nothing to the cross-source comparison — a source that contradicts itself cannot corroborate. Suppressing its finding while leaving that exclusion in place would let a registry declare variation, serve one region a forged address, and be reported as agreement, because nothing would remain to compare it against. A declaration therefore redirects the comparison into that context rather than removing it: the EU answer is checked against other sources' EU answers.
What is specified and not yet built. A source that varies contrary to a
scope it published itself has broken a rule it set, which is a stronger finding than an
undeclared difference. -01 specifies it and also forbids raising it without
evidence that the scope predates the observation — otherwise a narrower scope published
after the fact manufactures an accusation against an honest source. Establishing that order
needs a witnessed record of what each source published and when. Nothing in the reference
implementation provides one yet, so the finding is defined and not emitted.
The atomic object: the Corroboration Record
A sweep produces one signed, content-addressed Corroboration Record per subject — a self-contained verdict a third party re-verifies offline, without trusting the party that swept. It is signed with Ed25519 over the JCS-canonical body, the same canonical path the records it compares already use.
corroboration_record {
subject
observed_at
claims: [ per source, per vantage: present | absent | error
+ outcome, + binding {sequence, scope} when leased ]
verdict: AGREE | DIVERGENT | INSUFFICIENT
findings: [ omission | endpoint | digest | did | source_equivocation
| stale_binding | scope_violation | stale_description ]
staleness_window
sweeping_party_identity
sig (Ed25519 over JCS(record sans sig))
}
A finding is suspected when first seen and confirmed once re-observed past the staleness window — so legitimate propagation delay is not mistaken for a lie. Two records from one sweeping party that disagree about the same subject at the same instant are themselves evidence.
What it consumes, produces, and does not do
Registry answers
What each configured registry and name service claims about the watch set — a POST /resolve target object, a by-id record, a DNS answer — attestations and lease positions included.
The Corroboration Record
One signed, content-addressed verdict per subject, re-verifiable offline — the artifact that turns an assumed answer into a corroborated one.
Merit, capability, behavior
It does not rank agents, judge capabilities, or vouch for conduct. It attests only whether the sources describing a subject agree.
Ecosystem positioning: Corroborated Discovery is the proof layer for the quilt — it makes the registries self-checking, complementing the discovery and identity that NANDA, ANS, DNS-anchored discovery and any resolution registry already provide. A layer above the registries, not a competitor to any of them, and inside the DAWN boundary: integrity of the discovery exchange, never trust evaluation of the agent.