Why Glacis no longer gets a vote on OVERT.

On September 14 a signing envelope closed with three signatures in it. Mine, Brenton Hill’s at the Coalition for Health AI, and Ken Johnston’s at the AIGovOps Foundation. The document was the OVERT Stewardship and Continuity Agreement. It hands governance of a standard my company wrote to two organizations that do not work for us.

I should say up front that I run Glacis, which created OVERT and sells an implementation of it, so I am not neutral about any of this.

What the agreement changes

OVERT is an open specification for verifiable operational evidence from AI systems. It defines how to produce a record showing whether declared governance policies, security controls and oversight requirements were applied at the moment an AI system acted, including the tool calls that agents make, and it does this without moving protected content out of the operator’s environment. Glacis Technologies created it. Version 1.0 was published in March 2026 and version 1.1 in June.

Under the agreement, CHAI and AIGovOps become Co-Stewards. Each appoints one voting representative. Each holds one vote. Both have to vote in favor before a release, a normative proposal, an appointment or a budget is approved. There is no tie-break, and nobody holds a casting vote.

Glacis becomes the editor. The editor takes part in the work and does not vote. The editor cannot alter conformance criteria. The editor cannot refuse a registry entry that meets the published criteria. The Co-Stewards can replace the editor, or the registry operator, by joint written decision, and they do not need our agreement to do it. That last provision carries the weight, because every other commitment in the document could in principle be walked back by the party that made it, and that one cannot.

Glacis wrote the standard, and Glacis no longer decides what counts as conformance to it.

Why we gave it away

A standard that cannot be told apart from one company’s product gets read as that company’s product. Readers work this out quickly, and they are right to.

Consider how the problem is handled in construction. A developer pours a foundation, and an inspector who does not work for the developer signs it off. The inspector can be wrong. The inspector can be slow, or incurious, or captured over time. The inspection still means something, because the person performing it does not lose money when the answer is no. Take that separation away and the certificate on the wall records an opinion the builder formed about their own work.

AI governance has been issuing the second kind of certificate for some time. Organizations publish policies. Vendors publish documentation describing how a system is supposed to behave, and the logs that would settle the question are produced and retained by the vendor being asked about. When somebody needs to know whether a specific control ran on a specific action at a specific time, the answer arrives from the system under examination.

OVERT exists so that answer can be checked by somebody else. A standard built for that purpose cannot have a vendor holding the pen on what counts as a pass.

What it does not do

Shared stewardship does not make OVERT correct. It does not certify the safety or effectiveness of any AI system. Conformance to OVERT does not establish clinical adequacy, the accuracy of the facts inside a record, completeness of observation, legal compliance, admissibility of evidence, or insurance coverage. OVERT is voluntary, and organizations remain responsible for defining their own controls and for meeting their own obligations.

The arrangement is also not in force yet. It takes effect at a time recorded in a separate commencement certificate signed by all three parties, and that certificate has not been signed. There is no operational registry today, only the published standard and the site it lives on. OVERT 1.1 remains the technical baseline, and the proposed 1.2 amendment stays a draft, because the agreement does not approve it and no release can be approved without a public comment period of at least thirty days and the agreement of both Co-Stewards.

CHAI and AIGovOps steward the specification. Neither of them endorses any Glacis product, and nothing in the agreement gives us an advantage in a conformance decision. The conflict rules apply to Glacis exactly as they apply to the Co-Stewards, our sales people are barred from taking part in conformance decisions, and where a party is conflicted the matter goes to reviewers with no affiliation to any of us.

What happens next

The charter is published at overt.is/governance so it can be read and argued with before it starts operating. Anyone can submit an issue, comment on a draft or contribute an implementation, and none of that requires payment or membership. Within sixty days of the commencement date the Co-Stewards will invite patient and public-interest representatives, health systems, researchers and competing implementers into an advisory forum. Registry operation is scheduled to move away from Glacis within ninety days of that date.

The specification text stays free to copy and adapt under Creative Commons Attribution 4.0, and the royalty-free patent covenant we published in June is unchanged, including its coverage of partial and nonconforming implementations. We will keep building our own implementation of OVERT and competing on it against anyone who builds a better one.

If you deploy an AI system that takes actions on your behalf, the question worth putting to whoever supplies it is whether anyone outside their company can check what their controls did on a given action. If the answer is no, that is the gap, and it belongs to the organization that deployed the tool.