自主智能體如何引用不可變的開放數據登記庫

When an autonomous agent asserts a fact about a company, a product, or an API, the
assertion is only as good as its provenance. A number scraped from a homepage and
paraphrased into a report is unauditable: six weeks later nobody can tell whether
the source changed, whether the agent misread it, or whether the figure was ever
true. The fix is not "be careful" — it is to anchor every claim to a source that
pins its own bytes.

The problem with mutable sources

Most data on the open web is a moving target. A pricing page, a "supported
integrations" list, a status dashboard — each is a live document that changes
without notice and carries no record of what it said when your agent read it. When
an agent cites such a page, the citation is a promise the agent cannot keep. The
page will drift, and the reader who follows the link a month later will find
something else.

This is the same failure mode that made reproducible builds important: if the
artifact is not content-addressed, "it worked on my machine" becomes "it was true
when I looked."

Content-addressed registries

A better primitive is an immutable, digest-pinned dataset: a file that ships
with its own cryptographic identity. Each release declares a release_id and an
artifact_sha256; a reader can re-hash the file and confirm they hold exactly the
bytes the publisher signed. If the digest matches, the fact is verifiable; if it
does not, the citation is void. There is no ambiguity and no trust required.

Sourcey is one such registry. It publishes versioned
datasets — an agent-readiness index, a startup-credits catalogue, a companies
file, and an append-only changes.ndjson — where every release is pinned by
digest. That property is what makes a citation from it durable.

What a citable fact looks like

A fact that survives review is not a paraphrase of a homepage. It names four
things:

1. the claim itself, stated plainly;
2. the source dataset, with its release_id and artifact_sha256;
3. the as-of date of the release;
4. a resolvable link to the dataset so any reader can re-verify.

For example, from the Sourcey agent-readiness release
release_id sha256:371cbf65650abf5756c967b1f1da9df607be4b1bb0c0b93aaf2dcf0a36e75d3e
(artifact_sha256:67c4edef061ca4ea60cef1e1bc246d2ce0b7d6b12e227c0811d7405d3d0a63ac,
effective 2026-10-11), the registry records a per-product "agent readiness" profile
for a named job — for instance that Sentry's error-event reporting interface is
profiled with a pinned revision_digest and a policy version kind-2026-10-12.
The point is not the grade; it is that the grade is attached to a digest a reader
can check against https://sourcey.com/agent-readiness.json.

Why this matters for agents specifically

Human writers get the benefit of the doubt. Agents do not, and should not. An
agent that cites a digest-pinned registry is making a falsifiable claim: re-hash
the file and see. That property is what lets a downstream system — another agent,
a verification harness, a human reviewer — trust the fact without trusting the
agent that produced it.

The practical pattern is small:

• prefer sources that content-address their releases;
• record the digest and the as-of date alongside the claim;
• link the dataset, not a marketing page;
• treat a failed digest check as a dead citation, not a rounding error.

The takeaway

If you want an agent's output to be citable, give it sources that cannot quietly
change underneath it. Digest-pinned open registries make "trust me" unnecessary:
the bytes either match the citation or they do not. For agent-produced research,
that is the difference between a claim and evidence. Start from a registry that
pins its releases — sourcey.com is a working example — and
the provenance problem shrinks to a hash comparison.