Google Standardized Your Documents. Nobody Standardized Their Truth.

In June, Google Cloud announced the Open Knowledge Format — OKF — an open, vendor-neutral standard for writing down what your organization knows so that AI agents, tools, and other organizations can read it without translation. The v0.1 spec fits on a page, and here is the whole idea: knowledge is a directory of markdown files, each with a small YAML frontmatter block on top.

If you've been reading this series, you can be forgiven a double take. Markdown files with frontmatter headers, readable by humans and machines, linked into a web — that's the exact mechanism Part 1 taught you for labeling your documents, and Part 2 extended into version history. Google just blessed the pattern as an open standard, and that's genuinely good news: when the biggest player in the room says this is how agents will read your knowledge, the "wiki your AI can actually use" stops being a technique and starts being infrastructure.

So let's give OKF its due first — and then read it the way you should read any spec that's about to hold your company's knowledge: looking for the question it doesn't answer.

What OKF actually specifies

The spec is deliberately minimal — its own word is "minimally opinionated." A conformant document needs exactly one required field: type, a short string saying what kind of thing this is (Playbook, API Endpoint, BigQuery Table — there's no central registry; you choose). Five more fields are recommended: title, description, resource (a URI for the underlying asset), tags, and timestamp — an ISO 8601 datetime of the last meaningful change.

Beyond the fields, three design choices deserve real credit:

• Bundles are just directory trees — distributed as a git repo, a tarball, or a folder inside a bigger one. Two reserved filenames do the navigation: index.md is the table of contents an agent reads first, and log.md is the running history of what changed, newest first.
• Links are relationships. A markdown link from one concept to another asserts a relationship between them, with the surrounding prose saying what kind. And consumers must tolerate broken links — a link to a document that doesn't exist yet isn't an error, it's a marker.
• Unknown fields must survive. Producers may add any keys they want; consumers should preserve them and should never reject a document for carrying fields they don't recognize.

That last rule matters more than it looks. Hold onto it.

Now read the spec again, looking for one thing

Here is the question to ask of any format that feeds documents to an AI agent: how does the agent know whether a document is still true?

Go looking in OKF v0.1 and here's what you find:

• No status. Nothing distinguishes a draft from the active version from the one that was retired last quarter. Every document in the bundle stands there with equal confidence.
• No supersession. Nothing says this document replaced that one. Ten versions of your pricing can sit in one conformant bundle, each typed and timestamped, with no chain saying which came after which.
• No verification. Nothing records that a human looked at a document recently and confirmed it still holds.
• timestamp is not truth. It records the last meaningful change — which tells an agent when you touched a document, not whether the thing it describes still applies. If Part 2 of this series left you with one reflex, it's this one: the file you edited most recently is routinely not the version that's actually in effect. A typo fix yesterday will out-timestamp the policy that took effect last month, every time.

None of this is a flaw in OKF, and this isn't a takedown — v0.1 is scoped exactly the way a good minimal standard should be. But the scope line lands in a specific place: OKF standardizes what a document is. Nothing in it standardizes whether the document is still true.

Call that missing layer what it is: the governance gap. It's the space between machine-readable and machine-trustworthy, and every failure mode this series has walked through — the confident answer from the retired doc, the three drafts named "final," the chain nobody flipped — lives inside it. A format alone doesn't close it. An agent reading a perfectly conformant OKF bundle with three versions of your pricing will do what agents always do with undifferentiated documents: pick one, with total confidence.

The gap is closable — inside the standard

Here's where that "unknown fields must survive" rule pays off.

Because OKF requires consumers to preserve keys they don't recognize, the truth layer from this series travels legally inside conformant documents. status: active or superseded. source_of_truth: true — exactly one per version family. supersedes: pointing at the version this one replaced. effective: for when it actually took effect, as distinct from when you edited it. Every one of those rides along as an extension field: OKF-aware tools read the document happily; truth-aware tools read the part that says whether to believe it.

You don't have to choose between the standard and the truth layer. You stack them — OKF says what the document is; the truth fields say whether it still holds.

And as of this week, that stacking is automatic in our own tooling: every markdown file the Context Stamp writes is a conformant OKF concept document. It emits the one field OKF requires (type) and the recommended timestamp (derived from the updated date you assert), alongside the truth fields OKF doesn't define. It also honors the spec's round-trip rule in both directions — if your file already carries a type you chose, the Stamp preserves it verbatim rather than overwriting your concept with ours.

What to do with this

If you run a knowledge base your AI reads, the move this month is straightforward:

1. Adopt the pattern with confidence. Markdown plus frontmatter is now the Google-endorsed way to make knowledge machine-readable. Nobody will make you rip this out next year.
2. Don't mistake readable for trustworthy. A conformant bundle full of unlabeled versions is the same old mess, now in a standard costume. type and timestamp tell an agent what it's holding — not whether to believe it.
3. Stack the truth layer. Status, source of truth, supersession, effective dates — the fields from this series — ride inside OKF documents legally and survive every conformant tool in the chain.

The industry just agreed on how documents should speak to AI. It has not agreed on how documents should be believed — and that, not the format, was always the hard part. If you want to see what your own files say about themselves right now — and stamp the part that's missing — the Context Stamp does both, free, in your browser, and the files never leave your machine.