Abstract As AI systems increasingly participate in consequential decisions, organizations face a widening gap between operational visibility and evidentiary accountability. Existing observability infrastructure records how software executed through logs, traces, metrics, and runtime telemetry, enabling reconstruction of system behavior after the fact. It does not necessarily preserve the decision itself as an independently verifiable object. This paper introduces decision proof as a distinct infrastructure primitive. Rather than reconstructing consequential decisions from distributed runtime residue, decision proof packages the decision at the moment it is made into a portable, canonical decision object containing the evidence boundary, governing policy, authority chain, runtime reference, decision output, and cryptographic commitment required for later verification. The result is a shift in evidentiary architecture: from probabilistic reconstruction toward deterministic verification under a declared regime. Decision proof complements observability rather than replacing it. Observability explains system behavior; decision proof preserves the decision object that organizations ultimately need to defend. A generic financial-services example illustrates the distinction, followed by an architectural treatment of rollback, recursive decision governance, and federated verification. The paper concludes by positioning decision proof as a general infrastructure category for governance, regulation, audit, appeals, and long-term evidence preservation in autonomous systems. VERIFY is presented only as a reference implementation of the broader architectural primitive.
Devin Bostick (Fri,) studied this question.