TheQuilt Working draft

Documentation

FAQ

What Corroborated Discovery is, what it is not, and how the pieces fit.

What is Corroborated Discovery, in one sentence?+

A way to trust a registry answer without trusting the registry: ask several independent registries the same question, compare their answers, and produce one signed Corroboration Record that anyone can re-verify offline.

Why isn't a signed registry record enough?+

A signature proves a record was not altered — it says nothing about what a registry chose not to serve, or whether it served the same answer to everyone. A registry can hide a record, forge an endpoint whose signature checks against a key it controls, or answer two callers differently. None of that breaks a signature; only comparison across sources catches it.

What does it catch?+

The three ways a registry lies: omission (a record one source serves and another confirms absent), tampering (sources claim different endpoints for the same agent), and equivocation (a source serves different valid identities to different callers or vantages).

What is the Corroboration Record?+

The atomic object: a signed, content-addressed verdict about one subject, carrying the per-source per-vantage claims, the verdict (AGREE / DIVERGENT / INSUFFICIENT), the findings, a staleness window, and the sweeping party's identity — Ed25519 over the JCS-canonical body, re-verifiable offline without trusting the party that swept. Reference implementation: sm-divergence.

Which discovery systems does it work across?+

sm-bridge is the on-ramp: it speaks the NANDA Index and AgentFacts, the Agent Name Service (ANS), and DNS-AID agent-identity discovery. A sweep reaches every substrate the bridge speaks and corroborates their answers against each other.

How does it relate to registry-assisted resolution and DAWN?+

Registry-assisted resolution of a known subject is one registry's answer: a terminal target, authority evidence, a renewable binding. Corroboration is what a verifier does when it holds two of them. The profile forbids inter-registry forwarding, unbounded fan-out and enumeration — The Quilt satisfies all three, because the sweeper queries each configured registry directly, one exact lookup per subject. It reads the binding's sequence so a lagging registry is attributed as stale_binding rather than accused of tampering, and it never evaluates authority evidence, which keeps it inside the DAWN boundary: integrity of the discovery exchange, not trust evaluation of the agent. How the two compose →

Is this just about ANS? Where do ARD catalogs, DNS-AID, AgentFacts cards and DIDs fit?+

The resolution profile names two terminal target classes — an artifact ("an Agent Card, MCP server descriptor, workload descriptor, or a subject-owned discovery catalog") and a subject-authorized gateway — plus Subject Identity types with their own authority models. Every panel of the quilt is one of those things. An ANS-style name service is a Resolution Registry whose authority evidence is a log receipt. DNS-AID is the direct DNS-anchored path, which stays direct; the sweep treats it as a source beside the registries and so becomes the downgrade detector. An ARD signed catalog is a subject-owned discovery catalog — the artifact a registry binds a subject to — pinned by targetDigest and by its own detached signature, so a hijacked entry under a stale signature surfaces as digest as well as signature-invalid. A native AgentFacts card is an artifact with whatever proof the card carries. A DID or workload identity is a Subject Identity type whose document or descriptor is the artifact. None is a second kind of registry; none needs its own diff. The panel-by-panel table is §2.6 of The Registry Points, the Quilt Proves.

Two registries give different answers. Is one lying?+

Not necessarily, and since -02 the record says which. If one serves a binding at sequence 13 and the other at 14 in the same authority scope, the first is behind: that is a stale_binding finding attributing it, and its old target is not compared. If both are at 14 and the targets differ, that is a real disagreement and fires as endpoint. If both agree on the address but not on its digest, the registries are honest and the host is serving two documents: digest. Three conditions a plain endpoint comparison would collapse into one.

Is it a competitor to NANDA or ANS?+

No — complementary. NANDA, ANS, DNS-AID and any resolution registry discover and identify agents; Corroborated Discovery is the proof layer that makes their answers checkable against one another. It makes the quilt of registries self-checking; it does not replace any registry and never forwards one to another.

Does it rank or judge agents?+

No. It attests only whether the sources describing a subject agree — not whether the agent is capable, trustworthy, or well-behaved. Merit is out of scope. Where an agent then acts, holding that action to account is the companion accountability layer, Attested Actions.

Is this open source / an open standard?+

Yes. The sm-* reference implementations are open and published at github.com/Sharathvc23, and the procedure is developed in the open as draft-chandra-agent-registry-corroboration, alongside papers from the research house Stellarminds.ai.