AI Governance Is Not Enough to Prove Responsibility: A Conceptual and Testable Architecture for Demonstrable Responsibility in AI Systems

Artificial intelligence governance frameworks increasingly define organisational duties relating to risk management, documentation, transparency, and human oversight. Those frameworks are necessary, but they do not necessarily produce a reconstructable account of responsibility in a specific operational event: who held authority, who accepted responsibility, what action followed, what evidence supports the outcome, and how the resulting claim was assessed. This paper defines that limitation as the Operational Responsibility Gap and proposes Responsibility Infrastructure as a conceptual architecture for representing and reconstructing responsibility claims across organisational boundaries. Its central claim is limited: where responsibility must support external reliance, allocation, agency, and accountability may need to be jointly evidenced rather than inferred from fragmented governance records. The proposal is conceptual and testable; it has not yet been empirically validated across organisations, sectors, or implementations. This paper does not claim to have solved the problem or established the effectiveness of Responsibility Infrastructure; it defines a bounded and testable architectural hypothesis and identifies diagnostic and pilot conditions through which its necessity, proportionality, and practical value may be independently assessed.