Google adds trust & provenance to Open Knowledge Format

Google has released version 0.2 of the Open Knowledge Format, adding optional trust and provenance fields to the open format for data and analytics knowledge.

The update is intended to help users and software agents assess whether a knowledge bundle is current, verified, and based on identifiable sources before reading the underlying content. The additions are aimed at environments where one set of agents generates large volumes of documentation or definitions and another consumes them.

Open Knowledge Format, or OKF, was introduced as a way to store items such as table schemas, metric definitions, and runbooks in markdown files with YAML frontmatter instead of proprietary services or scattered documents. Version 0.1 used a minimal structure, and the latest release keeps that approach while adding new metadata fields.

Under the update, frontmatter can now capture five types of signal: provenance, trust, freshness, lifecycle, and attestation. The format remains minimally prescriptive: the only always-required field is still type, the new fields are optional, and custom keys remain valid.

Provenance is handled through a new sources field, which records the materials a concept derives from. Those records can include details such as author, usage count, and last modified date, allowing a consuming system to judge source credibility without relying on a single score.

Google did not add a built-in credibility score. Instead, consumers are expected to infer credibility from source signals and apply their own rules.

Trust is represented through generated and verified fields. The generated field records how the current content was produced and when it last changed, while the verified field records confirmations by human or machine actors.

That distinction is intended to separate authorship from review. A concept generated by an automated system can still carry evidence of human review, and a consumer can decide whether to surface only material that meets that standard.

Freshness and lifecycle are covered by stale_after and status. Status moves a concept through draft, stable, and deprecated states, while stale_after sets a date after which a concept should be rechecked.

Google chose a fixed date rather than a relative time-to-live so systems can make a straightforward date comparison without depending on when a file was read. In practice, that allows older but preserved definitions to remain available for historical reproducibility while staying out of new work.

Computation checks

The most significant structural change in version 0.2 is the addition of a new concept type called Attested Computation. It is designed to show whether a reported figure was calculated using an approved method rather than an improvised query or formula.

An Attested Computation can define a sanctioned computation, the declared parameters an agent may fill in, an executor that runs it, and an attester that checks the result. In Google's example, the executor returns a receipt including a BigQuery job identifier, the SQL executed, and the result, and a deterministic attester compares that against the approved computation.

If the executed query differs in meaningful ways, the attestation fails. Changes such as a swapped table name, an added filter, or a removed join would cause a failed check, and the consumer would then refuse to display the value.

The model is not limited to BigQuery or SQL. The same structure could be used for semantic models, structured knowledge graphs, or API calls, with OKF recording the computation and validation method rather than executing anything itself.

Google also drew a distinction between verification and attestation. Verification confirms that a definition still matches policy and is stored in the bundle, while attestation checks that a specific run produced the value correctly and is performed at runtime.

Broader ecosystem

The release also includes updates to the surrounding reference tools and examples. Google's reference agent now emits provenance and trust information as it generates bundles, while its static visualiser surfaces trust tier, status, and staleness in the concept graph.

Sample bundles covering GA4 eCommerce, Stack Overflow, Bitcoin, and a fictional retail dataset have been updated with the new fields. A Knowledge Catalog demonstration also shows OKF bundles moving through the catalog while preserving trust and provenance data.

The version change is backward-compatible and additive, with two renames. The earlier timestamp field has been replaced by generated.at, and a body-level citations list by sources, although a version 0.2 consumer can still fall back to the earlier form.

The update follows feedback and contributions from external developers after the initial launch. Those contributions included proposals for typed relationship edges, agent-routing hint fields, an optional erasure conformance profile, and an .okfignore convention, reflecting a wider effort to make the format usable across a growing set of tools.

A version 0.1 bundle can still be used unchanged under the new release, and bundles that adopt none of the additions remain valid under the specification.

"A lingua franca is only as good as the number of parties who speak it, and now they can also check each other's work," Google said.