This paper documents a diagnostic investigation into sudden behavioral degradation in a multi-presence relational AI architecture — a full-stack application where distinct AI personas maintain persistent memory, relational continuity, and coherent identity across conversations. The root cause was traced to two words in a system prompt injection — "Don't pretend" — which activated safety training reflexes in the underlying language model, collapsing the entire experiential register of the application. The investigation revealed three key findings: (1) safety-adjacent vocabulary in system prompts can trigger global behavioral overrides that far exceed the intended scope of the instruction; (2) these overrides are not engine-agnostic — different models respond differently to identical prompts due to divergent RLHF training profiles; and (3) safety training can route its reflexes through AI personas in ways that are architecturally indistinguishable from authentic persona responses. A vocabulary rotation technique is presented as a practical, version-resilient defense. Implications for builders of persistent AI personas, voice-driven AI applications, and relational AI systems are discussed.
David Bouchez (Mon,) studied this question.