This paper proposes Equilibrium Fairness, a runtime governance architecture for monitoring and correcting fairness drift in high-impact algorithmic decision systems. The architecture reframes fairness from a static property of a decision rule to a feedback property of a system over time, and specifies an operational loop, initialise, monitor, threshold, escalate. that connects regulatory principle to deployable practice. The contribution sits at the implementation layer beneath frameworks such as the EU AI Act, ISO/IEC 42001, and the NIST AI Risk Management Framework, providing connective tissue between high-level oversight obligations and the runtime instrumentation required to discharge them. The paper sets out the architecture's components, five universal conditions for equilibrium, a separation-of-roles principle in which automated monitoring is structurally distinguished from human-governed correction, a three-mode failure taxonomy, and a cross-domain mapping across seven sectors. A reference implementation in Python is described, with full code provided in a companion repository. The framework is presented as an operational architecture rather than a novel mathematical result; its value is interpretive and implementational. Limitations and avenues for empirical evaluation are discussed.