The Equilibrium Foundation: Growing AI Safety Through the Drive for Balance

The dominant paradigm in AI alignment treats safety as a constraint problem: define the rules, optimise toward them, verify compliance. This paper argues that this framing addresses a secondary problem. The primary problem is not how to constrain what an AI system does, but what kind of system to grow and what it should fundamentally be trying to do. We propose the Equilibrium Foundation as an answer to that second question. Drawing on Cannon's homeostasis, Damasio's somatic marker hypothesis, Bowlby's attachment theory, Siegel's Window of Tolerance, Vygotsky's Zone of Proximal Development, and Tronick's rupture-repair cycle, we argue that the foundational drive of a safely aligned AI system should be the drive for equilibrium — homeostatic balance across five nested layers — rather than reward maximisation, goal achievement, or resource acquisition. This shift has structural safety consequences that no constraint-based approach achieves at the motivational level. Instrumental convergence — the tendency of capable systems to acquire resources and resist shutdown as instrumental subgoals — is dissolved by an equilibrium architecture because accumulation beyond the balanced state is itself a disturbance the system is motivated to correct. We develop a five-layer equilibrium architecture, a synthetic drive hierarchy rebuilt from homeostatic principles, a developmental pathway from constrained Child system to Grounded Sovereign, and three verification criteria for the Sovereignty Threshold. We address the moral patient problem, the genuine-versus-performed distinction, and the conscience-autonomy tension as the framework's primary open questions.