Executive summary

In June 2026, organizational knowledge acquired a standard shipping container. The Open Knowledge Format — an open, vendor-neutral specification published by Google Cloud — turns institutional knowledge into portable bundles of plain files that any person can read and any AI system can consume. Portability is settled. Governance is not.

This brief is about the question the format deliberately leaves open: once knowledge can move — between teams, tools, organizations, and increasingly between AI agents that read and write it autonomously — who decides what crosses a boundary, what ranks above what, what gets corrected after it has spread, and what deserves trust? These are old questions. What is new is that they now have buildable, architectural answers.

Key findings

• The unit of governance has shifted from the record to the claim in circulation. Classical data governance controls who may open which record. Open-knowledge governance must control reach — what a piece of knowledge is allowed to influence after it leaves its source.
• A format is a contract for reading, not a constitution for behavior. The Open Knowledge Format specifies how knowledge is written down and stays deliberately silent on authority, trust, and correction. Every adopting organization inherits those questions on day one.
• Every durable governance design resolves the same seven tensions. Privacy versus value, immutability versus correction, popularity versus proof, openness versus poisoning, human oversight versus machine scale, one commons versus many boundaries, measurement versus meaning. Naming them turns philosophical arguments into design choices.
• Governance is layered, and the order is load-bearing. Constitution before leases; trust before ranking; disputes before admission; local re-earning before federation. Building out of order recreates the open web's failures — popularity ranking, unaccountable moderation — on new ground.
• Machine-scale writing breaks approval-scale governance. When AI agents maintain the knowledge, thousands of units are written and cross-referenced per week. Per-item human sign-off is arithmetically dead. The workable split: humans govern the rules and the judges; machines govern the instances.
• The cheapest governance decision is made before launch. Every integrity monitor ships in the same release as the mechanism it guards. A ranking layer without its anti-manipulation alarm, shipped “for now,” is how a commons quietly becomes a popularity contest.

What changed: knowledge left the database

The context an organization runs on — what a metric actually means, why the system was built this way and not the three other ways, the runbook for the incident that only happens in December — has historically lived everywhere except in a governable place: wikis, drives, code comments, and the heads of a few senior people. Data governance frameworks governed the databases. The knowledge that decides how the databases get used was ungoverned because it was unstructured, and unmovable because it was scattered.

That is what the Open Knowledge Format changes. Published as an open specification in June 2026, OKF formalizes a pattern the AI era converged on independently: knowledge kept as a directory of plain markdown files with structured front matter, readable by humans without tooling, parseable by AI agents without custom integration, diffable in version control, and portable across tools, organizations, and time. A knowledge bundle is distributed like source code. There is no central authority and no schema registry — deliberately.

“The value of a knowledge format comes from how many parties speak it, not from who owns it.”— Google Cloud, on the Open Knowledge Format

None of this is a defect. A format earns adoption precisely by refusing to legislate behavior. But the silence has a consequence: the day the container arrives, its contents' governance becomes your problem — and the problem is urgent for a reason that did not exist five years ago. Knowledge is now written at machine scale. AI agents produce and maintain it as a byproduct of doing real work, touching fifteen files in one pass, updating cross-references no human would remember to update. The maintenance burden that killed the corporate wiki is finally automated away — which means the volume, velocity, and blast radius of shared knowledge are all about to be multiplied by machines that do not get tired.

An asset class this valuable, moving this fast, governed by nothing but good intentions, has a familiar historical shape. The open web ran the experiment once: open contribution plus popularity ranking plus no provenance produced both the greatest knowledge commons in history and its characteristic pathologies. The chance to do it right the second time — while the ground is still greenfield — is the subject of the rest of this brief.

The seven tensions every design must resolve

Strip away the vendors and the acronyms, and every governance regime for shared knowledge faces the same seven trade-offs. They are genuine tensions, not bugs: each pits two things you actually want against each other. A governance design is nothing more, and nothing less, than a committed resolution to each.

Three of these deserve a closer look, because the intuitive resolution is the wrong one.

Popularity vs. proof (3). The intuitive move — let consumers upvote what they find useful — is the one the open web made, and it is the original sin of every knowledge platform since: it measures agreement, virality, and recency, none of which are correctness. The alternative is now practical: when the consumers of knowledge are agents doing verifiable work, you can observe whether work that used a unit of knowledge actually succeeded — and let only that move rank.

Openness vs. poisoning (4). The intuitive move is a stronger front door: review every contribution for truth before admitting it. This fails twice — it cannot actually detect a well-crafted false claim, and it hands whoever controls the gate a censorship instrument. The design that works accepts that wrong and hostile units will get in, and makes that survivable: every new unit is born on probation with a small, bounded reach, and grows its influence only as verified evidence accumulates. You win at the influence layer, not at the gate.

Measurement vs. meaning (7). Goodhart's law — once a measure becomes a target, it stops measuring — is not a risk in a knowledge commons; it is a schedule. The moment standing affects visibility or money, someone optimizes the proxy. The only cheap defense is temporal: the alarm that detects the gaming ships in the same release as the mechanism worth gaming. Retrofitted monitors arrive after the corruption they were meant to catch.

Where governance starts: promises, then pipelines

Governance does not start with software. It starts with a short list of promises, written where every participant can read them, that the architecture actually enforces. Call it the constitution of the commons. Two properties separate a constitution from a marketing page: every promise is traceable to a shipped mechanism that enforces it, and no promise exceeds what the code does. Ratify what is enforced; never aspire in writing.

Five promises worth ratifying

• “We share the shape, not the substance.” The commons learns from patterns, structure, and aggregate statistics — never from your raw content.
• “Delete severs reach.” When you delete, nothing surfaces your data again. What can survive is only structure already absorbed into aggregates that, by construction, cannot be traced back to you.
• “We never rewrite history; we always govern reach.” The record of what happened is immutable. What any record is allowed to influence going forward is governed, and revocable.
• “Humans govern the evaluators.” The rules, and the judges that score contributions, are set and amended only by accountable humans. No automated process can rewrite the rules or certify its own verdict.
• “Your access is computed, not copied.” What you see is derived at read time from your relationship to each unit. There is no global view and no per-member fork to drift out of date.

The membrane: how “shape, not substance” is enforced

The first promise is the one most organizations get wrong, because they enforce it with labels — humans marking documents “shareable” or “internal.” Labels leak. A document marked general can still name a client, a dollar figure, a case number. The enforceable version routes everything that leaves a private boundary through one pipeline — a membrane — with no bypass path:

• Scrub — detect and strip identifying entities;
• Lift — replace concrete specifics with typed placeholders, keeping the pattern;
• Generalize — induce the structural lesson from the instance;
• Aggregate — emit a pattern only when observed across at least k independent sources;
• Test — scan the candidate output for residual identifying content, from both directions;
• Budget — cap total disclosure under an explicit privacy budget.

The critical property is fail-closed: an error at any stage blocks the emission. A membrane that fails open is a press release waiting to happen. And the membrane clarifies what “sharing posture” means: an open participant publishes everything; a selective participant lets only membrane output cross; a private participant shares nothing — and the honest version of “nothing” is architectural absence of the sharing path, not a configuration flag someone could flip. Where your infrastructure runs and what crosses your boundary are independent decisions; conflating them is how buyers get talked out of the posture they actually need.

The architecture: nine layers, in dependency order

With the tensions named and the constitution ratified, the governance of a knowledge commons decomposes into nine layers. The order is not a roadmap convenience — it is load-bearing. Each layer makes the next one safe to build.

Why the order is load-bearing

• Ranking before trust is a popularity contest with extra steps — the exact failure the open web cannot undo, recreated on your greenfield.
• Admission before disputes turns your front door into a censor: with no channel to contest claims after entry, the gate is pressured to rule on truth, which it cannot do.
• Retraction without a derivation graph is cosmetic: the wrong claim is gone, but every conclusion built on it lives on, unmarked.
• Federation without local re-earning is trust laundering: standing manufactured cheaply in one commons walks into another at face value.

The stakes-gate principle. The moment governance starts paying — ranked retrieval steering decisions, knowledge earning money — is the exact moment gaming arrives. Therefore: no ranked retrieval and no money attached to knowledge until the trust ledger and its integrity monitors are live and green in the same release. Committing to this while it is cheap is the single highest-leverage governance decision an organization can make.

The mechanics of trust: proof, decay, and auditability

The keystone layer deserves its own page, because it is where governance either earns its keep or quietly fails. The question a trust ledger answers: of two units of knowledge that claim the same thing, which should a consumer — human or agent — be steered toward?

What is allowed to move standing

Four properties that make it governance rather than a score

• Standing decays. Without fresh evidence, trust erodes on a half-life of weeks. Yesterday's authority is not permanent tenure; the commons stays current because staleness is priced in.
• The unknown gets sampled, not buried. A new, unproven unit receives deliberate exploration traffic — the multi-armed-bandit insight — so evidence can accumulate. Without this, early winners lock in forever and the commons ossifies.
• The ledger is recomputable, byte for byte. Standing is a pure function of the recorded event history: any auditor can replay the events and must land on identical numbers. The system continuously recomputes and alarms on any divergence — tamper-evidence by construction, not by promise.
• Defense is proportional to evidence. Overturning a unit with fifty verified outcomes takes proportionally more dispute evidence than overturning one with two. Well-earned standing is real protection — and a rejected challenge is trust-neutral, so no one farms wins by surviving manufactured disputes.

Note what is absent: no committee assigning quality scores, no moderator deciding truth, no black-box relevance model. Humans appear at exactly two points — writing the rules, and signing off on changes to the verdict thresholds — which is the only division of labor that survives machine-scale volume while remaining accountable.

How to evaluate any knowledge commons

Whether you are joining a shared knowledge network, buying a platform that maintains one, or building your own: ask these seven questions. Answers should be specific and mechanical, not aspirational.

1. Can we export the knowledge — and does governance travel with it? The bundle should leave in an open format with its provenance and standing attached. Portability that strips the governance metadata exports the words and abandons the trust.
2. What exactly crosses our boundary, and what enforces that? The answer must name a pipeline, not a policy. If sharing is governed by labels and good intentions, one mislabeled document is a breach. Fail-closed or it is not a membrane.
3. When something is retracted, what happens to everything built on it? Ask to see the derivation graph and the measured propagation. If corrections do not ripple to derivatives, every error that spread is permanent.
4. What moves rank? If any part of the answer is engagement, votes, retrieval counts, or recency — walk away. The only acceptable inputs are verified outcomes and adjudicated verdicts, with the manipulation alarm live in the same release.
5. Who can change the judges? The thresholds that define a passing verdict must be changeable only by named, accountable human sign-off — never by an automated process, and never silently.
6. How does a wrong-but-confident claim get contested? There must be a structured dispute channel whose verdicts are deterministic given the evidence — not moderator discretion — and whose rejected challenges change nothing.
7. What does imported knowledge count for on arrival? Anything arriving from outside — another team, another organization, another commons — should arrive probationary and re-earn standing locally. “Trusted because the source is big” is how poisoning goes wholesale.

A commons with confident, mechanical answers to all seven is rare today. A commons with fewer than four is an ungoverned feed with a governance page.

Next steps

If you are the decision-maker reading this brief

Three actions this quarter:

1. Inventory what is already moving. Which of your organization's knowledge already leaves its boundary — in vendor tools, in AI assistants, in shared documents — and under what governance? The honest answer is usually “more than we thought, and none.”
2. Write the constitution you can already enforce. Ratify the promises your current architecture actually keeps — even if the list is short. A true page beats an aspirational framework. Every promise you cannot yet enforce becomes a roadmap item, in dependency order.
3. Adopt the stakes-gate rule now, while it costs nothing. Commit in writing: no ranked retrieval and no monetized knowledge until outcome-based trust and its integrity monitors ship together. This one sentence, adopted early, is the difference between governing a commons and moderating a feed.

If you are the champion who handed this brief upward

Two actions this week:

1. Schedule the conversation. The seven-question checklist is the agenda. Any knowledge network, platform, or internal initiative your organization is evaluating can be scored against it in one meeting.
2. Frame governance as the precondition of the asset, not overhead on it. If institutional knowledge is becoming a balance-sheet asset — and it is — then governance is its accounting. Nobody calls accounting overhead on the money.

References

• Google Cloud, How the Open Knowledge Format can improve data sharing (June 12, 2026), and the Open Knowledge Format specification, v0.1 draft (GoogleCloudPlatform knowledge-catalog repository)
• Andrej Karpathy, public commentary on LLM-maintained knowledge bases (2026), as cited in the Open Knowledge Format announcement
• Elinor Ostrom, Governing the Commons: The Evolution of Institutions for Collective Action (1990)
• Marilyn Strathern, “Improving ratings” (1997) — the canonical formulation of Goodhart's law
• Cynthia Dwork, Differential Privacy (2006)
• Latanya Sweeney, k-Anonymity: A Model for Protecting Privacy (2002)
• Peter Auer, Nicolò Cesa-Bianchi, and Paul Fischer, Finite-time Analysis of the Multiarmed Bandit Problem (2002)

Version 1.0 · July 2026 · This brief is provided as vendor-neutral category education. Distribute freely.