GLM-5.2's Persona Steering findings nuance Pattern #152 by explaining the mechanism of collateral damage. Structural interventions (activation steering, weight modification) target specific behaviors but affect distributed neural representations — changing a weight for safety can degrade a weight for reasoning. The Persona Steering paper demonstrates this experimentally: activation-steered models show degraded performance on unrelated benchmarks. Narrative interventions (story framing, persona prompting) achieve similar safety improvements without the degradation because they work through the model's existing interpretative framework rather than modifying its architecture. This is the difference between rewiring a circuit (structural) and giving better instructions (narrative). Pattern #152 previously noted that safety interventions had side effects; #154 explains why and, crucially, offers a better alternative.