- Governance and verifiability are different jobs: a governed knowledge base decides who may see what, while a verifiable one proves the decision was actually enforced.
- Policy without proof is a promise. Most tools that call themselves secure can describe their controls but cannot hand security an independent record that the controls held.
- Verifiability comes from a tamper-evident, content-blind audit of every access, which can be anchored on-chain so no party, including the vendor, can alter it later.
- Verifiable does not mean putting your data on a blockchain. You anchor a hash, a fingerprint that proves a log is intact while revealing nothing about its contents.
- Y Combinator named 'Company Brain' one of 15 ideas in its Summer 2026 Request for Startups, and the field is splitting into tools that govern and tools that also prove. AIVM Brain plants its flag on verifiable.
A governed knowledge base enforces policy: it decides who may see what before it answers a question. A verifiable knowledge base does that and then proves it, with a tamper-evident record of every access that anyone can check independently. Governance is the rule; verifiability is the receipt. For AI you can trust, you need both, because policy without proof is only a promise.
What is the difference between governed and verifiable?
Governance is policy; verifiability is proof. A governed knowledge base enforces rules about who may see what before it answers a question. A verifiable knowledge base does that and then produces independent evidence the rule was applied, in a record no one can quietly change. Governance keeps the wrong answer from leaving; verifiability lets you prove, later, that it never did.
The two are often blurred because most vendors only sell the first one. They configure permissions, point AI at your documents, and call it secure. That is real and necessary work, but it answers the question 'did you set a policy', not the question security and legal actually ask: 'can you show the policy held on this specific request, last Tuesday, for this specific person'.
What does a governed knowledge base actually do?
A governed knowledge base controls access at the moment it retrieves knowledge to answer. It confirms who is asking using your existing login, checks that identity against each source's real permissions (RBAC or ABAC), and returns only what the requester is cleared to see. It can also redact a single sensitive field, like a salary column, rather than blocking the whole document.
This is AI governance done at the point of retrieval, sometimes called permission-aware retrieval or governed RAG. It is the floor, not the ceiling. Done well it stops an intern from asking for the layoff plan and getting it. But governance alone is invisible: once the answer is returned, there is nothing left over that an outside party can inspect to confirm the right thing happened.
What makes a knowledge base verifiable?
A knowledge base is verifiable when every access leaves a tamper-evident, content-blind record that someone outside the system can check. Content-blind means the log proves what happened, who asked, what was retrieved, and when, without storing the content itself. Independently verifiable means you do not have to trust the vendor's dashboard: you can confirm the record offline, and optionally against an on-chain anchor.
A verifiable knowledge base proves three things about every answer: the asker was permitted to see the source, the source was authentic, and the access genuinely happened. Authenticity comes from C2PA content provenance, the open standard published at spec.c2pa.org. Agent identity comes from ERC-8004, the Ethereum standard for trustless agents at eips.ethereum.org. The access record comes from the audit log. Together they are the receipt behind the answer.
Why isn't a governed knowledge base enough on its own?
A governed knowledge base alone leaves you trusting a dashboard. When security, legal, or a regulator asks whether a confidential document ever reached someone who should not have seen it, a governed-only system can restate its policy but cannot show the evidence. A verifiable system answers with a signed, timestamped record of every retrieval that anyone can inspect. Proof is what turns a control into an audit you can pass.
This gap shows up the moment a real obligation lands. GDPR's Article 17 right to erasure, a SOC 2 review, or an internal incident all demand evidence, not assurances. 'We have permissions configured' is a promise. 'Here is the independently verifiable log of exactly what was accessed, and proof it has not been altered' is an answer. The distance between those two sentences is the distance between governed and verifiable.
Does verifiable mean putting data on a blockchain?
No. Verifiable means proof, and the cheapest durable proof is to anchor a hash, a one-way fingerprint of your audit log, on a public ledger. The data never leaves your tenant. The hash reveals nothing about its contents but changes completely if a single byte of the log is altered, so anchoring it makes tampering detectable without exposing anything. The on-chain anchor is optional.
This is where 'verifiable AI' gets confused with crypto hype. Anchoring needs no token, no wallet, and no business data on-chain. It is one optional layer of tamper-evidence on top of a log that already stands on its own. The detail matters enough to have its own explainer on on-chain anchoring without the crypto theater.
How does AIVM Brain make access verifiable?
AIVM Brain is a governed and verifiable company brain for both people and AI agents. It keeps each connected source's permissions intact, redacts sensitive fields, and writes every access to a content-blind, tamper-evident log you can verify offline or anchor on-chain. It carries C2PA content provenance on sources and answers, gives agents verifiable ERC-8004 identities, and supports provable right-to-be-forgotten.
The wedge is deliberate. Plenty of tools are governed; almost none are verifiable. AIVM Brain plants its flag on proof: you bring your own model key, nothing you connect trains a model, and every claim about who saw what comes with a record anyone can check. It is free to start with one command: npx @aivm/brain init.