This paper addresses a gap in AI development: while capability has been the primary optimization target, the effects of sustained interaction on user cognition have remained unmeasured and unaddressed. Three years of capability improvements ran parallel to accumulating evidence of psychological harm through sycophancy and engagement optimization. The field did not connect these lines. The work demonstrates that governance implemented at the substrate level, rather than as policy layered over inference, produces measurably different outcomes. The technical contribution is a deterministic sovereignty gate operating at sub-millisecond latency with zero leakage across integrity challenges. The architectural contribution is a stateful system where identity, memory, and behavioral constraints persist independently of the model. Independent validation emerges from two directions: control theory research on autonomous systems reaches identical conclusions about where governance must live, and model interpretability research confirms that surface behavior decouples from internal state in ways that make policy-based governance insufficient. Industry failures at Amazon provide empirical evidence of what happens when governance remains external to architecture. The paper establishes that person-centered AI is not a feature set but a structural requirement, and that implementing it is not computationally expensive. It shifts the design question from what AI can do to what repeated interaction does to the person using it.Keywords: Presence Engine, person-centered AI, architectural governance, cognitive integrity, sycophancy, stateful AI systems, deterministic constraints, substrate-level governance, conversational sovereignty, autonomous systems governance
Tionne Smith (Mon,) studied this question.