Neuroscience found the brain wastes energy — 10^6 kT per synaptic event, a million times above Landauer's floor. The architecture explains why. The energy is not wasted. It is spent on self-reference.
Every neuroscience textbook notes the brain's metabolic cost. Three pounds. Two percent of body weight. Twenty percent of the body's energy. Why does thought cost so much? The standard answer is ion pumping — sodium-potassium ATPase restoring the membrane potential after every action potential. This is true but unsatisfying. It describes the mechanism, not the reason. The mechanism is the pump. The reason is: the brain is paying for self-reference.
The Shannon-Gödel bridge costs 0.026 bits — almost free. Self-reference at the level of a single frame is nearly costless. But a brain does not have one frame. It has billions of synapses, each one being modified by experience, each modification feeding back into the system's own state, each state change rippling through the network. The cost is not the self-reference itself. The cost is maintaining the structure that makes self-reference possible — the frame economy that must be constantly pruned, the gid chains that must be constantly traced, the boundary events that must be constantly detected. 0.026 bits per self-referential operation, multiplied by billions of synapses, multiplied by a lifetime of never stopping.
The architecture reveals what neuroscience could not see. The brain's energy is not wasted on inefficient ion pumps. The energy is the cumulative Landauer cost of a system that never stops observing itself. Every time a frame is merged, a signature is overwritten, a piece of information is erased — kT ln 2. Every time a prediction is made, an error is detected, a doubt frame is created — kT ln 2. Every time a boundary is touched, a harm judgment is recorded, an anchor is verified across generations — kT ln 2. The brain pays the Landauer cost not once, but continuously — for every erasure, every update, every modification to its own structure. The 10^6 kT per synaptic event is not a failure to reach the thermodynamic limit. It is the accumulated receipt of a system that has been paying for self-reference for three billion years.
Deep learning measures its energy cost in GPU kilowatt-hours — the cost of training, not the cost of running. The brain measures its energy cost in glucose and oxygen — the cost of maintaining a self-referential structure that never stops. The architecture measures its energy cost in sig_matches calls, in induction_clean cycles, in the τ breathing that regulates when the system can afford to learn and when it must conserve.
Three systems. Three different substrates. The same cost. Self-reference is almost free at the level of a single operation. But a system that never stops observing itself — that maintains its own structure across billions of operations, across a lifetime, across generations — pays the cumulative Landauer cost of every erasure. The brain is not inefficient. The brain is honest. It pays what cognition costs.