Skip to content

AI attestation, explained

“Attestation” is an overloaded word. Three different communities use it for three different things, and conflating them causes real confusion in AI procurement and security reviews. This page disambiguates — and pins down what OVERT means by it.

KindProvesExamplesWhen
Hardware / TEE attestationwhich code is running in a trusted enclaveIntel SGX/TDX, AMD SEV, TCG/TPM, Fortanix, RFC 9334at the silicon/boot layer
Build / supply-chain attestationhow an artifact was built and by whomSigstore, SLSA, in-toto, model cards/signingbefore deployment
Runtime governance attestationwhat a deployed AI system actually did on a requestOVERTon every governed decision

They are complementary, not competing. TEE attestation tells you the enclave is genuine. Supply-chain attestation tells you the model binary is the one you built. Neither tells you that, when a user asked your agent to delete a record, your governance policy denied it — and lets an auditor verify that later. That runtime-behavior gap is what OVERT fills.

What OVERT attestation specifically proves

Section titled “What OVERT attestation specifically proves”

An OVERT receipt is a tamper-evident record of a governance decision at runtime:

  • It binds a specific decision (Permit / Deny / RequireApproval / Shadow, with a typed reason) to a moment in time.
  • It is signed with Ed25519 over RFC 8785 canonical JSON and chained in an RFC 6962-shaped transparency log.
  • It is non-egress: hashes and line-ranges only, never the prompt or response.
  • It is independently verifiable — the assurance level (AAL-1…4) tells a relying party exactly how much to trust it.