Google has shipped version 0.2 of the Open Knowledge Format (OKF), and the headline change is trust. The update adds five mechanisms that let whatever consumes an OKF bundle — an AI agent, an LLM, or an application — check where the information came from and whether it can still be relied on. If you want to go further, Google: 404 crawling means Google is open to more of your covers it in detail.
The framing Google uses is a set of five questions a consumer should be able to answer before acting on a piece of knowledge:
- Provenance — what was this created from?
- Trust — how much should I rely on it?
- Freshness — is it still true?
- Lifecycle — is it the current version?
- Attestation — was this number produced the way we said it must be?
Each maps to a concrete field or type in the spec.
Provenance: the new sources field
OKF v0.2 introduces a sources field recording what a concept derives from — an external document, a bundle-relative path, or a scope descriptor such as “all queries in project X.”
Alongside it, entries can carry objective credibility signals: author, usage_count, and last_modified.
The notable design decision is what Google deliberately left out. OKF records signals but does not assign a credibility score. As Google puts it, a score “is subjective, doesn’t port across consumers, and goes stale the moment it’s written.” Credibility is instead inferred by whoever consumes the bundle — the same way a human trusts a heavily used, recently updated, authoritatively authored source over an anonymous one. See also: reviews, reputation & listings: the local signals AI now.
Attribution is also handled per claim rather than as a list at the bottom: when the body cites a source, it uses an ordinary markdown footnote keyed to the source ID.
Trust: generated and verified
Two fields separate who made something from who checked it. The generated field records who created a concept; verified records who independently confirmed it.
That split lets a consumer filter by confidence level — unverified, machine-confirmed, or human-reviewed — instead of treating every entry in a bundle as equally reliable. For agent workflows, this is the difference between acting on a claim and flagging it for review.
Freshness and lifecycle: status and stale_after
The status field marks where a concept sits in its lifecycle: draft, stable, or deprecated. The stale_after field sets a date past which the concept should be re-verified before use.
Together they let systems exclude outdated concepts from new work while keeping them intact for historical reference — rather than deleting knowledge that was true at the time.
Attested Computation: verifying the math
This one is a new concept type, not a field, and it addresses a problem specific to agentic systems.
Provenance tells you where a claim came from. Attestation answers something harder: when an agent reports a dollar figure, was that number produced by the sanctioned method, or did the agent improvise its own query?
An Attested Computation carries both the meaning of a value and an approved way to compute it, plus the means to verify that the approved method actually ran. For anyone letting agents touch financial or operational reporting, this is the most consequential part of the release.
Why this matters
OKF is infrastructure for machine-readable knowledge, and v0.2 is fundamentally about making that knowledge auditable. As more workflows hand decisions to agents, “where did this come from and can I still trust it” stops being a nice-to-have and becomes a prerequisite.
It also fits a broader pattern: trust and verifiability are becoming explicit, structured signals rather than things inferred from reputation. We covered the content-side equivalent in the trust signals that get your brand cited by AI Overviews, and the wider citation mechanics in what it takes to get cited, and stay cited, in AI search.
Where to find it
Google has updated the OKF GitHub repository to reflect v0.2 and published an accompanying announcement explaining the changes in detail.