AI agents operating on enterprise knowledge bases now have a standardized mechanism for verifying whether the information they consume was generated correctly, by whom, and whether it remains current. Google Cloud published Open Knowledge Format (OKF) version 0.2 on July 25, 2026, via the Google Cloud Blog, adding five trust signal categories to the specification and introducing a new concept type designed to verify how numerical values are computed.
The Five Trust Signals
Google introduced OKF in June 2026 as a format for storing enterprise knowledge, including table schemas, metric definitions, and runbooks, in structured, agent-readable markdown files with YAML frontmatter. As AI agents began generating thousands of knowledge assets automatically, enterprises raised a core accountability question: how can another AI agent know whether generated knowledge is accurate, current, and approved?
OKF v0.2, released July 25, 2026, added a trust layer consisting of fields recording where a concept came from, who produced it and when, who verified it and whether by a human or a machine, and when it should be treated as stale. Google confirmed in its July 25 Google Cloud Blog post that the update organizes these capabilities into five named signals: provenance, trust, freshness, lifecycle, and attestation.
In OKF v0.2, all five of those questions are now answerable from frontmatter, while the format remains as minimally opinionated as v0.1. The update adds vocabulary, not rules: `type` is still the only always-required field, every new field is opt-in, and a bundle that adopts none of the additions is exactly as valid as it was under v0.1.
Provenance: The `sources` Field
The first signal, provenance, is implemented through a new `sources` field in YAML frontmatter. According to Google's July 25 blog post, the field records the materials a concept derives from, an external document, a bundle-relative file path, or a scope descriptor such as "all queries in project X." Each entry in the field can carry objective credibility signals including `author`, `usage_count`, and `last_modified`.
OKF records these signals rather than computing a single trust score. Google's specification states that a credibility score is subjective, does not port across consumers, and goes stale the moment it is written. Instead, the specification leaves scoring to the consuming application. When a concept body cites a specific source, attribution is handled through a standard markdown footnote keyed to the source's `id` field, enabling per-claim attribution rather than a single list at the bottom of a document.
Trust: The `generated` and `verified` Fields
The trust signal is established through two distinct fields. The `generated` field records who or what produced a concept and when the content last changed. The `verified` field records one or more independent confirmations of that content against its sources.
Google's specification derives a trust tier from the `verified` field. A concept with no `verified` key is classified as unverified. Verification by non-human actors only yields a machine-confirmed classification. Verification by a `human:` actor yields human-reviewed status. Google's blog post confirmed that these tiers function as advisory signals, not access control, but enable consumers to filter on trust level before committing resources to reading a concept's full body.
Freshness and Lifecycle: `stale_after` and `status`
OKF v0.2 establishes freshness and lifecycle with the `stale_after` and `status` fields. `status` moves a concept through draft → stable → deprecated, with absent meaning stable. `stale_after` is a single absolute date, a deliberate choice over a relative TTL, so staleness becomes a plain date comparison with no reference to when the concept happened to be read, which is the kind of determinism a non-LLM consumer requires.
Attestation: A New Concept Type
The fifth signal introduces a structural change beyond a new field. Google confirmed in its July 25 blog post that OKF v0.2 introduces `Attested Computation` as an entirely new concept type. Where provenance addresses the origin of a claim, attestation addresses a different question: whether a reported number was produced using the approved calculation method, or whether an agent substituted its own logic.
An `Attested Computation` concept carries a sanctioned computation, typically a SQL query or model invocation, along with an `executor` that defines how to run it and an `attester` that independently verifies the result. The attester is specified as deterministic code with no LLM involvement. According to Google's specification, the comparison is mechanical: a rewritten query, a swapped computation file, or a mutated dependency fails the attestation check. The consumer is then expected to refuse to display the value when the verdict is negative.
Google's blog post noted that the abstraction is intentionally flexible. A Knowledge Catalogue demo shows a bundle round-tripping through Google Cloud's Knowledge Catalog, formerly Dataplex, with trust and provenance signals preserved through the catalogue and back.
Backward Compatibility and Reference Implementation Updates
Version 0.2 is a minor version bump that is additive and backward-compatible, with two deliberate renames: `timestamp` is superseded by `generated.at`, and the body `# Citations` list is superseded by `sources`. In both cases, a v0.2 consumer can fall back to the v0.1 form, and a v0.1 bundle drops in unchanged.
The reference agent in the GitHub repository now emits the provenance and trust families as it generates, so a freshly minted bundle arrives with generated sources and citations already in place. Google also confirmed that updated sample bundles, including GA4 e-commerce, Stack Overflow, Bitcoin, and the Acme Retail example used in the announcement, have been updated to carry v0.2 fields.
What This Means for Marketers and Digital Teams
For digital marketing teams using AI agents to generate or manage structured knowledge, including product catalogues, performance metric definitions, or data governance documentation, OKF v0.2 provides a mechanism to attach verifiable sourcing and expiration metadata directly to those knowledge assets. Teams adopting the format can configure consuming agents to filter on trust tier, exclude stale concepts from active use while preserving them for historical reference, and verify that reported figures were produced using approved calculation methods. Because the format remains plain Markdown with YAML frontmatter and requires no SDK or schema registry, adoption does not introduce infrastructure dependencies.
The OKF v0.2 specification, sample bundles, and reference implementations are available at GitHub.
