Morphology of frozen labyrinths from irreversible threshold dynamics

What the math gives to ML

The paper provides a constructive irreversible threshold process in which each binary site can correct a local minority error at most once and then becomes inert. Its transferable mechanism is finite-use, state-dependent updating that prevents recurrent oscillations and produces arrested residual errors concentrated near interfaces. A neural-network analogue is a one-shot recurrent refinement layer for binary segmentation, denoising, or discrete latent inference, with measurable predictions about convergence, update-order sensitivity, and interface-localized frustration.