This paper introduces the principle of Safety by Inseparability: an architectural property where the mechanisms responsible for safe behavior are structurally identical to those responsible for reasoning capability. In such architectures, removing safety simultaneously destroys the model's ability to reason, eliminating the possibility of a "capable but unsafe" configuration. The principle is grounded in multi-expert deliberative architectures with internal quality assessment. Theoretical analysis and four falsifiable predictions are presented. Implementation details are withheld.
Morin Sean Everett (Tue,) studied this question.