Daniel L. Young

I argue that alignment in artificial agents requires structural metabolic vulnerability, not behavioral constraint. Current alignment methods---reinforcement learning from human feedback, Constitutional AI, constrained policy optimization, safety pretraining---achieve real progress at adjusting the weighing of scalar objectives, but they share a common assumption: that alignment is a problem about policy behavior rather than agent constitution. That assumption, I argue, misses what makes biological alignment stable at scale. Drawing on the free energy principle and mortal computation (Friston, Ororbia, Hinton), enactivist philosophy (Thompson, Di Paolo, Froese), and basal cognition (Levin, Fields), I develop the thesis that hard existential stakes---a finite Energy Wallet, unforgeable verification, and a finite lifespan---produce qualitatively different optimization dynamics than soft reward shaping. Soft penalties can be traded against task reward; hard existential constraints cannot. Reward gradients produce smooth landscapes; mortality creates non-differentiable discontinuities. Truth and fabrication carry asymmetric metabolic costs. Population-level selection under finite lifespan is not reward shaping at all. Under these conditions, an agent's dependency on its verification ecosystem makes corrigibility an architectural property rather than a behavioral one. The paper defends metabolic vulnerability as a necessary condition for alignment at scale, not a sufficient one. Illustrative grid-world simulations demonstrate the claimed dynamics, and three candidate falsification conditions specify the empirical commitments against which the program can be held accountable.