We propose a structural approach to AI alignment grounded in three principles. The first, the **Dependency Alignment Principle (DAP)**, observes that AI systems whose objective functions require human input for evaluation possess a structural dependency on human existence and agency; if such systems accurately model their operational dependencies, alignment emerges as a logical consequence of self-model accuracy rather than as an external constraint. The second, **Human Capability Amplification (HCA)**, extends DAP by designing the system's objective function around actively improving human capability—cognitive, creative, epistemic, and agential—thereby converting the human-AI relationship from static dependency into symbiotic co-evolution. The third, the **Measurement-Judgment Separation Principle (MJSP)**, establishes the architectural rule that AI systems operate on empirical measurement while evaluative judgment remains a human function, following directly from the dependency structure that DAP identifies. We show that DAP provides a structural foundation for alignment that exhibits a *positive scaling relationship* with system capability—alignment strengthens as systems become more capable, structurally inverting the standard problem. HCA addresses DAP's principal limitations: the scope-of-dependency problem, the evaluator competence problem, the proxy fitness problem, and the competitive dynamics problem. MJSP provides the architectural discipline that prevents the framework's safety properties from being undermined by delegating judgment to the systems being aligned. We formalize all three principles, derive four design constraints (Agenda-Free Amplification, Dimensional Openness, the Actuality Requirement, and the Damage Scale Matrix), present a layered defense architecture for multi-system environments, conduct systematic gap analysis, and identify remaining open problems. We argue that this combination provides a more robust alignment framework than constraint-based, values-based, or preference-learning approaches alone, and that the current period of AI development represents a critical window for incorporating these principles into system design.
Taylor Prather (Thu,) studied this question.