On July 25, 2026, Google Cloud published version 0.2 of the Open Knowledge Format (OKF), its specification for packaging the metadata, context and curated knowledge that AI agents need. The update adds optional fields that let a consumer judge how far to trust a piece of knowledge before reading it: where it came from, who generated it, who verified it, whether it is stale, and whether a number was produced the way it was supposed to be.
OKF launched as v0.1 on June 13. Google described it then as “a vendor-neutral, agent- and human-friendly standard” and as “a starting point, not a finished standard.” Version 0.2 is the first substantive revision, and it is almost entirely about trust.
OKF formalises what Google calls the LLM-wiki pattern. A knowledge bundle is a directory of Markdown files, each with YAML frontmatter. A file might describe a BigQuery table, a metric or a runbook. The only required frontmatter field is type; others include title, description, resource and tags. Google’s Knowledge Catalog can ingest OKF bundles and serve them to agents, and the spec, reference implementations and samples are published in the knowledge-catalog repository on GitHub.
/ 01What v0.2 adds
The v0.2 post frames trust as five questions a consumer must answer:
- Provenance — what was this created from? A
sourcesfield lists the materials a concept derives from, with signals such as author, usage count and last-modified date. Google deliberately records signals rather than a score, because a credibility score “is subjective, doesn’t port across consumers, and goes stale the moment it’s written.” - Trust — how much should I trust it? A
generatedfield records who or what produced the content and when; averifiedlist records independent confirmations. A consumer derives a tier: unverified, machine-confirmed, or human-reviewed. Google notes these tiers “are advisory signals, not access control.” - Freshness — is it still true? A
stale_afterdate makes staleness “a plain date comparison.” - Lifecycle — is it current? A
statusfield moves concepts from draft to stable to deprecated. - Attestation — was this number produced correctly? A new
Attested Computationtype pairs a sanctioned computation, such as a SQL query, with an executor that returns a receipt and a deterministic attester that checks it. The agent may only fill declared parameters and “must never author or edit the computation.”
The release is backward-compatible: v0.1 bundles still work, timestamp is superseded by generated.at, and the body’s citations list is superseded by sources. The reference agent, visualiser and sample bundles were updated to emit and display the new fields.
/ 02Why it matters
Most agent-context formats so far have focused on getting knowledge into the prompt. OKF v0.2 is notable for focusing on whether that knowledge should be believed. Its design choices are sound. Separating “generated by” from “verified by” makes the difference between machine-written and human-reviewed content explicit. Recording provenance signals rather than a single score leaves judgement with the consumer. And attested computation addresses one of the most common failure modes in enterprise agents: a plausible number produced by a query the model quietly rewrote.
Google’s own distinction is worth repeating: “A stale definition can still attest cleanly; a freshly-verified definition still needs attestation on every run.” Verification of what a definition says and attestation of how a result was computed are different checks, and agents need both.
/ 03What it does not yet cover
OKF is deliberately simple: Markdown, YAML and links between files. That is a strength for adoption and a limit for reasoning. Relationships between concepts are expressed as links in prose, not as typed, machine-checkable statements. There is no shared vocabulary for what kinds of things exist and how they may relate, so two bundles describing “customer” may not mean the same thing. And the new trust fields sit in each concept’s frontmatter, so they describe a whole document rather than each individual statement inside it.
Those are the problems that RDF, OWL and upper ontologies such as the Basic Formal Ontology were designed to solve: typed relationships, shared meaning across independently built datasets, and statement-level provenance. A sensible architecture uses both. A lightweight format like OKF can carry curated knowledge to agents; a governed knowledge graph can supply the entities, relationships and rules underneath, with every fact tied to its source. The two meet naturally at OKF’s new sources and verified fields.
Telling an agent what to know is easy. Telling it how much to believe, and why, is the real standard.
/ 04What to watch
- Governance. OKF lives in a Google repository. Whether it moves to a neutral standards body will shape adoption beyond Google Cloud.
- Consumer behaviour. The trust fields are advisory. The real test is whether agents and catalogues actually withhold unverified, stale or unattested content.
- Mappings to semantic standards. Bridges between OKF bundles and RDF or OWL would let organisations keep one model of their domain and publish it in several forms.
Formats such as OKF that make provenance and verification first-class are welcome, and they work best when the knowledge they carry is grounded in a model of the domain that can be checked.
OKF v0.2 gets the order right: before an agent uses a fact, it should know where the fact came from and who checked it. Its limit is granularity, because trust is recorded per Markdown file, not per statement.
In a knowledge graph, provenance attaches to each assertion. A revenue definition, the table it reads and the analyst who verified it are separate nodes, so one stale join does not taint a whole document, and a verified definition can still be traced to the sanctioned query that produced last quarter’s figure. Typed relations replace prose links, so two teams’ notions of “customer” either map to the same class or visibly do not. The two layers can coexist: a governed graph, built from source systems and refined by subject-matter experts, stays the record of truth, while lightweight formats carry curated slices of it to agents. Deterministic reasoning over that graph complements attestation: the same question gets the same answer with the same chain.
How we’d model it
Assertion—hasProvenance→SourceDefinition—verifiedBy→ReviewerMetric value—computedBy→Sanctioned queryConcept—staleAfter→Date
- Open Knowledge format v0.2 tackles agentic trust — Google Cloud Blog cloud.google.com ↗
- Introducing the Open Knowledge Format — Google Cloud Blog cloud.google.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.