Prediction Architecture Program

V20 and V21 established that embodied agents can carry world/self-model signals, but evolution alone was too slow to make deliberation adaptive. V22-V31 isolate a sharper question: what kind of learning signal actually couples hidden dimensions into a non-decomposable representation? The answer is not "prediction" in general. Linear prediction heads make agents better forecasters while leaving the representation partitionable. The architectural question is whether the gradient itself forces hidden units to depend on one another.