Translating Across Frameworks

In current AI alignment terminology, what this framework refers to as identity architecture overlaps with several existing areas of concern, including inner alignment (the consistency between learned and base objectives), value generalization under distribution shift, and robustness under novel conditions.

The distinction this framework introduces is not a rejection of these approaches, but a reframing. Rather than treating alignment as the problem of constraining or shaping learned objectives after capability is developed, it proposes that the system's objective-generating structure itself must be architected from the inside out, prior to capability development.

This reframing has practical consequences. In a constraint-based framework, the question is always whether a rule covers the current situation. In an architecture-based framework, the question is whether the system's outputs are coherent with its foundational structure — a question that can be asked about any output, including ones no rule anticipated, because the core is present in every output the system generates.