Alignment is Not Transferable Across Structural Discontinuity - A Structural Constraint on AI Alignment under Transformation

Every alignment method — RLHF, Constitutional AI, Red-teaming, preference optimization — operates on one assumption: that the system produced by alignment training is the system that is deployed. This assumption is not derived. It is presupposed. La Profilée establishes that a system under real transformation exists as a system if and only if it satisfies the persistence condition IR ≤ 1. If this condition is violated between alignment training and deployment, the aligned system and the deployed system are not the same system. Alignment is not a property of systems in general. It is a property of a specific system at a specific structural state. It is not transferable across structural discontinuity. This paper derives this consequence, specifies the structural condition under which alignment is preserved, and demonstrates the mechanism through the empirical case of catastrophic forgetting — reinterpreted as structural non-persistence rather than memory failure. The result is not a critique of alignment methods. It is the structural condition they have not yet addressed.