Andrej Karpathy had called it: the real value of a model-maintained wiki is that models don’t get bored. This week Google formalized it with the Open Knowledge Format, and the reaction was immediate. Charly Wargnier, ex-Streamlit, wrote on LinkedIn that it could end up replacing Notion or traditional wikis for developer teams.

He’s probably right operationally. And still, the format isn’t the part that matters.

The thesis up front: a knowledge standard standardizes structure, not truth. The moat was never the format, which is open and anyone can copy. It’s who curates the knowledge and who verifies it before an agent replicates it across the whole base. That’s the part no specification solves, and it’s exactly where Socio Tecnológico works.

What OKF is, without the noise

The Open Knowledge Format is deliberately simple. A folder of Markdown files with YAML frontmatter, kept in version control next to the code. A small set of agreed conventions so wikis written by different teams can be consumed by each other without translation.

No compression scheme. No central registry. No SDK. If you can cat a file, you can read it. If you can clone a repo, you can deploy it. Anyone who has used Obsidian, Notion, or any of the model-wiki patterns that emerged over the past year recognizes the shape instantly.

The argument in its favor is real and worth stating well. Human wikis don’t die for lack of a tool, they die of boredom. Someone stops updating a cross-reference, then another, and in six months the documentation lies. An agent doesn’t get tired of touching fifteen files in one pass to keep them coherent. On pure consistency, it beats any human team. That part is true.

The hole almost nobody named

Under the enthusiasm, the sharpest comments that same week pointed at one thing. An agent doesn’t update the documentation to match reality. It updates it to match what the agent believes at that moment.

And if that belief is wrong, the format doesn’t notice. It propagates it. Faster and more uniformly than any distracted human.

There’s the uncomfortable inversion. Humans abandoned wiki upkeep partly out of tedium, yes. But also partly because judging what is still true is hard, and that judgment doesn’t disappear because the writer is tireless. A standard for writing context doesn’t solve the harder problem, which is verifying it. It moves it downstream and, worse, makes it invisible: when everything is neatly structured and linked, it looks trustworthy even when nobody confirmed it’s true.

There’s a second, more technical hole. OKF promises entity relationships through cross-links, no graph database required. It works while you read one hop at a time. The moment you ask a multi-hop question, what depends on X that also touches Y, you’re back to one of two options: load the whole corpus into context or rebuild the index you said you didn’t need. The structure lives in the files. The ability to query it does not.

Curated knowledge is the asset, not the format

This connects to something I already argued about CRMs. In the post on the CRM as infrastructure the point was that value moved from the system of record to the reasoning layer that lives on top. The same thing happens here with knowledge. The format is the cheap, commoditized part. The expensive part is curation: deciding what enters, verifying it’s true, holding the judgment of what’s authoritative when two sources contradict.

An agent reading a flawless OKF full of unverified claims is no more trustworthy than a confident intern. It’s faster, which is not the same as better. The difference between an agent that helps and one that repeats errors with confidence isn’t in how it stores the text. It’s in who decided that text was true before letting it in.

That’s why the real problem isn’t format, it’s ownership. Who owns the knowledge base that feeds your agents? Who verifies it? Who designs the human checkpoint where a claim goes from “the agent believes it” to “the company confirmed it”? If the answer is nobody, OKF didn’t hand you an advantage. It handed you a more efficient way to propagate whatever your agents happened to assume. The cross-document speed that thrills Karpathy is also the speed of cross-document error.

What IQ Source does about it

Socio Tecnológico treats curated context as what it is: the asset that defines the quality of any agent you build. We don’t sell a format, formats are free. We own the role that decides what knowledge enters the base, verifies it before the agent propagates it, and designs where a human confirms before a claim replicates across two hundred related files.

The prior discovery, in AI Maestro, includes mapping where the knowledge your processes assume as true actually comes from. Often that knowledge lives in the heads of three people and a Slack channel, not in any document. Adopting OKF before that mapping is neatly organizing a base nobody confirmed. Order of operations matters: first you define who owns the truth, then you pick the format to store it.

Run a test with your team this week. Take the most-repeated claim in your internal documentation, the one everyone cites without thinking, and ask who last verified it and when. If nobody knows, you don’t have a format problem. You have a shared belief your next agent is about to treat as a fact. OKF will copy it faster. Verifying it is still your job.

Define who owns the truth before you pick the format