sm-divergence — catch a cheating registry by corroboration

Integrity for the sources agents resolve through. A registry (or a name service, or a DID method) can lie by omission, tampering, or equivocation; signing an artifact closes only tampering. This asks several sources the same questions and makes any disagreement loud.

Agents reach each other through middlemen — registries, indexes, DID methods — and a middleman can lie:

• tamper — serve a false endpoint (or key) for an agent,
• omit — hide an agent that is actually registered,
• equivocate — tell different clients different things.

A self-certifying artifact (a signed AgentFacts-style record, a did:key) defeats tampering — it can't be altered without breaking its signature. But no signature can prove what a source chose not to serve. Omission and equivocation are invisible to a single source.

The cheap, robust defense is one decentralized discovery already hands you: corroboration. If an agent is reachable through two or more sources, ask all of them the same question and treat any disagreement as a signal. That is all this library does — enough to turn a silently cheating source into a loudly diverging one. See the whitepaper for the argument and the spec for the normative procedure.

This is the reference implementation of the IETF draft Multi-Source Corroboration for AI Agent Discovery (draft-chandra-agent-registry-corroboration-00): the sm-resolver kernel is the claim model and diff, and this package is the signing layer — the agent self-description (§8) and the signed Corroboration Record (§9) — plus the reference discovery and identity layers.

Not a transparency log. Corroboration catches a source that disagrees with its peers; it does not stop a single source lying consistently to a first-time caller. That last mile is transparency-log territory (WHITEPAPER §7).

Where it fits

A DNS/CA-pillar primitive in the Project NANDA four-pillar model (DNS / CA / Orchestration / Attestation): it protects the integrity of discovery and identity resolution themselves. NANDA's discovery is multi-source by design — a lean Index delegating to a quilt of registries — and corroboration is only possible because of it: the redundancy built for resilience doubles as an integrity check. This library turns that latent property into an active one, the concrete mechanism behind accountable discovery — a federated registry that is not merely decentralized but auditable.

Install

pip install sm-divergence                 # omission + endpoint/key checks (httpx only)
pip install "sm-divergence[attestation]"  # + verified-DID checks (Ed25519/did:key)

Quick start

import asyncio
from sm_divergence import check_once

results = asyncio.run(check_once(
    ["https://registry-a.example", "https://registry-b.example"],
    ["agent-1", "agent-2"],
))
for r in results:
    print(r.agent_id, r.verdict)          # agent-1 DIVERGENT   (AGREE / DIVERGENT / INSUFFICIENT)
    for f in r.findings:
        print(" ", f.kind, f.confirmation, f.detail)
        # endpoint suspected {'field': 'endpoint',
        #                     'values': {'https://registry-a.example': 'https://real.example/',
        #                                'https://registry-b.example': 'https://attacker.example/'}}

Each check returns one SweepResult per subject — a verdict (AGREE, DIVERGENT, or INSUFFICIENT when fewer than two sources answered decisively), the claims, and the findings. A finding is suspected when first seen and confirmed once re-observed past the detector's staleness_window_s (so legitimate propagation lag isn't mistaken for tampering).

Long-running? Keep a detector and route new findings to your alert sink — it deduplicates, so a persisting divergence fires your callback once:

from sm_divergence import DivergenceDetector

detector = DivergenceDetector(
    ["https://registry-a.example", "https://registry-b.example"],
    on_finding=lambda f: alert(f.kind, f.agent_id, f.detail),
    staleness_window_s=300,
)
await detector.check(watch_ids=my_agent_ids)   # call each poll cycle

The Corroboration Record

Every sweep can be sealed into a signed, content-addressed Corroboration Record (draft §9) — the auditable proof that the checks ran, emitted for every subject whatever the verdict (the AGREE record is the positive attestation, not just the DIVERGENT ones):

from sm_divergence import sign_sweep, verify_corroboration_record

records = sign_sweep(results, sweeper_private_key=my_ed25519_key)  # needs [attestation]
# record_id = SHA-256(JCS(body)); signed by the sweeper's did:key
assert all(verify_corroboration_record(r) for r in records)        # verifies offline

CLI

python -m sm_divergence check \
    --registry https://registry-a.example \
    --registry https://registry-b.example \
    --watch agent-1 --watch agent-2

Exit 0 when the sources agree, 2 when any divergence is found (so a cron or CI step fails loudly), 1 on a usage error. Add --json for machine output, --attestation to also compare verified DIDs, --nanda-index URL beside --registry URL to corroborate a NANDA Index.

What it compares

Every finding is omission (present on one source, positively absent on another) or the name of a view field whose values disagree. The detail is uniform: {"field": <name>, "values": {source: value}} (or {"present_on", "missing_from"} for omission), so a consumer parses findings the same way at every layer.

Architecture — a kernel and its layers

The primitive factors into a source-agnostic kernel and a layer per thing a middleman can be asked. Each layer supplies its own view and thin resolvers; the kernel is untouched.

sm-resolver  (dependency)  the source-agnostic kernel — Resolver[T], the View
                           contract, Claim, diff_claims, Corroborator, SweepResult
sm_divergence/
  discovery/   registry-integrity layer — RecordView + HttpByIdResolver (NEST)
               + NandaIndexV2Resolver
  identity/    key-consistency layer — DidView + DidKeyResolver / DidWebResolver
               / UniversalResolverResolver + signed self-description
  record.py    the signed Corroboration Record (draft §9)

The kernel now lives in its own package, sm-resolver — sm-divergence is the reference set of layers built on it. Everything below is re-exported from the top level, so from sm_divergence import Corroborator, Finding, … is unchanged.

Sources of different formats mix in one run. A Resolver hides one source's wire format and normalizes its answer to the same view the diff understands:

from sm_divergence import DivergenceDetector, HttpByIdResolver, NandaIndexV2Resolver

await DivergenceDetector([
    HttpByIdResolver("https://nest.example"),        # GET /api/agents/{id}
    NandaIndexV2Resolver("https://index.example"),   # two-hop resolve → record
]).check(["agent-1"])

Identity layer — catches a did:web (or a Universal Resolver) serving a different verification key than the agent's self-certifying did:key root, a key divergence (the identity analogue of a tampered endpoint):

from sm_divergence import Corroborator, DidKeyResolver, DidWebResolver

await Corroborator([
    DidKeyResolver(did_for=lambda a: agent_did_key[a]),   # offline root of truth
    DidWebResolver(did_for=lambda a: agent_did_web[a]),   # fetched, may be hijacked
]).check(agent_ids)

A new format or layer is a thin adapter — a view with a comparable() and a Resolver that returns it; the generic Corroborator + diff_views do the rest:

from collections.abc import Mapping
from dataclasses import dataclass
from sm_divergence import Corroborator

@dataclass(frozen=True)
class MyView:
    field: str | None = None
    def comparable(self) -> Mapping[str, str | None]:
        return {"field": self.field}

class MyResolver:
    label = "src-a"
    async def resolve(self, agent_id): ...   # -> (Status, MyView | None)

await Corroborator([MyResolver(), ...]).check(agent_ids)

Signed correspondence — the same agent is a bare id on one registry, a URN locator on an Index, a did:web on a domain, and corroboration can't infer that these are the same agent. Instead of trusting a caller-supplied mapping, let the agent sign its own aliases + endpoints; verify offline and drive the resolvers from the verified names, so a source that disagrees is contradicting the agent's own signature:

from sm_divergence import (
    build_self_description, HttpDescriptionResolver, corroborate_descriptions,
    signed_alias_for, DidWebResolver, AliasClosureResolver, binds_by_key,
)

# the agent signs (bumping seq on every re-issue), with its did:key's private key.
# The bundle is the draft §8 shape: {version, subject, seq, aliases, endpoints, signature}.
bundle = build_self_description(
    aliases={"web": "did:web:acme.org", "nanda": "urn:ai:…:acme"},
    endpoints=["https://acme.example/agents/acme"],
    private_key=agent_ed25519_private_bytes, seq=5,
)

# a consumer fetches the description from EVERY source and reconciles it —
# omission / replay / agent-equivocation on the description itself become findings:
recs = await corroborate_descriptions(
    [HttpDescriptionResolver(s) for s in source_urls], ["acme"],
)
canonical = recs["acme"].canonical                       # None if the agent equivocates
sds = {"acme": canonical} if canonical else {}

# drive resolvers from the PROVEN aliases, with closure so a claimed alias that
# doesn't bind back to the did:key contributes no claim (never a false divergence):
web = DidWebResolver(did_for=signed_alias_for(sds, "web"))
web = AliasClosureResolver(web, lambda a: sds[a].subject if a in sds else None, binds_by_key)

The kernel lives in sm-resolver; sm-divergence is the reference layers built on it. The top-level public API is unchanged.

For verified-DID comparison over signed records shaped {"record": {..., "did": "did:key:z…", "endpoint": …}, "sig": "<b64 ed25519 over JCS(record)>"}, use the built-in adapter (needs the attestation extra):

from sm_divergence import DivergenceDetector, signed_record_adapter
DivergenceDetector(registries, adapter=signed_record_adapter())

Related packages

Develop

make ci-local   # uv: sync → ruff → format → mypy --strict → pytest

License

MIT

First published: 2026-07-04 | Last modified: 2026-07-04

Personal research contributions aligned with Project NANDA standards. Stellarminds.ai