Artificial intelligence systems are increasingly relied upon in regulated workflows, including finance, healthcare, employment, and consumer-facing decision support. In these settings, post-incident scrutiny rarely turns on whether an AI output was factually correct. Instead, liability and regulatory exposure arise when organizations cannot reconstruct what was produced, under which controls, and how the output was subsequently relied upon. This paper describes the minimum characteristics of an evidentiary control mechanism for AI systems operating in regulated environments. It defines when AI outputs become record-relevant, specifies the evidence objects required to make those outputs reconstructable, and outlines an operating model that distributes accountability across legal, compliance, risk, product, and operational functions. The focus is deliberately narrow and implementation oriented. The mechanism does not assess model accuracy, guarantee correctness, or prescribe optimization strategies. Instead, it addresses evidentiary survivability under audit, investigation, or litigation by describing how AI-influenced outputs can be made auditable, attributable, and correctable in practice. This work is intended as a governance reference artifact for organizations, auditors, and regulators evaluating the evidentiary implications of AI reliance.