Reference Integrity: A Framework for Detecting and Preventing Reference Standard Corruption in Adaptive Systems

This paper presents Reference Integrity (RI), a framework for detecting and preventing reference standard corruption in adaptive systems. RI addresses a specific failure mode in which a system’s internal evaluative standard drifts silently through recursive self-reference, producing compounding degradation invisible to standard monitoring. The paper formalizes the failure condition, proposes cross-channel divergence as an early detection signal, and situates RI within existing literature on model collapse, reward model overoptimization, and evaluation drift. An agent-based simulation provides mechanistic demonstration of the detection hypothesis. Cross-domain applications include ML systems, embodied robotics, and organizational systems.