Measuring What Persists: Conditioning Mechanisms and a Geometric Framework for AI Agent Identity

Andrew Tanner

AI agents in long-context applications drift from their specified identity. Current methods detect this only after qualitative degradation is visible. We present a geometric framew…