The Offline Attestation Record: A Tamper-Evident Provenance Mechanism for Sovereign, Offline Enterprise AI

Enterprise adoption of capable artificial intelligence in regulated sectors is constrained less by model quality than by data sovereignty: to obtain an answer, an organisation must typically transmit prompts, documents and records to infrastructure it does not control. This paper describes a sovereign architecture in which inference runs entirely offline, on hardware the organisation owns, and presents its central accountability mechanism, the Offline Attestation Record (OAR). The OAR binds the identity of the inference engine, the provenance of the loaded model, and a description of each action into a post-quantum-signed, append-only record that can be verified independently and without network access against a signed checkpoint held off the executing machine. We characterise the record deliberately as tamper-evident rather than tamper-proof: its guarantee is that silent, unnoticed alteration becomes detectable, not that alteration is impossible. We give the system model, the record construction, and a threat model covering silent log alteration, undisclosed model substitution and covert egress, and we discuss how offline, provable inference reframes obligations such as data residency and auditability from a matter of trust into a matter of evidence. The approach is model-agnostic, and its capability scales with the hardware the operator chooses to own.