What is OVERT AI evidence?
OVERT (Observable Verification Evidence for Runtime Trust) is an open, voluntary technical specification for producing verifiable records that an AI system's declared safeguards actually ran at runtime. It defines signed, hash-linked evidence for actions such as AI agent tool calls, human approvals, and overrides, so auditors can check controls without seeing patient data.
Who wrote it and who stewards it
OVERT was developed by Glacis Technologies. Version 1.0 was released on March 25, 2026, and version 1.1, the current stable version, on June 11, 2026.
On September 21, 2026, the Coalition for Health AI (CHAI) announced that OVERT will move to shared stewardship by CHAI and the AIGovOps Foundation, with Glacis serving as a non-voting editor and initial registry operator. It is accurate to call OVERT a standard headed for CHAI co-stewardship. It is not accurate to call it a CHAI-authored framework, and the arrangement is not yet active. The OVERT governance page states that the stewardship agreement was executed on September 14, 2026 but "takes effect at the date and time stated in a separate commencement certificate signed by all three parties," and that the certificate "has not yet been signed." Until then, Glacis remains steward.
CHAI's announcement is explicit about scope: "OVERT is a voluntary technical specification. It does not establish regulatory or legal requirements, certify the safety or effectiveness of an AI system, or determine whether an organization has complied with applicable law."
What it attests, and what it does not
OVERT attests that controls executed and how strong the evidence is. The standard states that it "does not attest the truthfulness of model outputs, the absence of hallucination, the absence of compromise, or the adequacy of the operator's policies."
Examples of its requirements for AI agents include TOOL-1, "Every tool call by an AI agent SHALL be evaluated against policy and attested before execution," tamper-evident tool-call logging (TOOL-5), human approval gates for sensitive operations (TOOL-4), and attestation of human corrections and overrides (HITL-3). A design principle is that only hashes and signed receipts cross organizational boundaries, while protected content stays with the operator.
Assurance and conformance levels
OVERT defines four Attestation Assurance Levels, from AAL-1 (policy documents) to AAL-4 (cryptographic attestation by a structurally independent attestation provider). Conformance claims use a declared Protocol Profile. Under section 22.6.3, a self-declared profile, marked with an `SD-` prefix, is valid only for Level 1 and Level 2 claims. Levels 3 and 4 require a registry-registered profile. The standard also describes qualified assessors who examine conformance claims.
For buyers, the practical takeaway is to ask vendors which level they claim, which profile they use, and whether anyone independent has examined the claim.
How Skovos handles this
Skovos exports signed, hash-chained evidence records in the OVERT v1.1 style under a self-declared profile. This is not a conformance claim and not a certification. It gives auditors a consistent, verifiable format for Skovos permission decisions and audit entries.
Frequently asked questions
Is OVERT a CHAI standard?
Not yet in the full sense. Glacis wrote it, and CHAI and the AIGovOps Foundation agreed to co-steward it, but the arrangement starts only when a commencement certificate is signed. As of late September 2026, the OVERT site said it had not been.
Is OVERT required for hospitals or AI vendors?
No. It is voluntary and does not create legal or regulatory requirements.
Does OVERT certify that an AI tool is safe?
No. It provides evidence that declared controls ran. It does not judge output accuracy or whether the policies themselves are adequate.
What does "OVERT-style evidence" mean?
It usually means records shaped like OVERT receipts, such as signed, hash-linked entries, without a formal conformance claim. Ask for the level and profile before relying on it.
Related reading
Sources
- CHAI, announcement on OVERT stewardship, September 21, 2026: https://www.chai.org/news/glacis-to-place-overt-its-open-standard-for-verifying-ai-safeguards-under
- OVERT governance charter and stewardship status: https://overt.is/governance
- OVERT v1.1 standard text: https://overt.is/latest.md
- OVERT version history: https://overt.is/versions.json