FLOWRUNNER — the morphogenetic network plays a maze

An environment hides a law: a winding high-conductivity corridor through a walled maze, excited by chaotic sources along the path. The network never sees the maze layout — only noisy potential-field snapshots. It identifies the local law with its streaming weak-form RLS, and its angle channel self-organizes into a corridor that steers a token from entry to goal.

The clip — watch the shape grow and the token follow the discovered law

Panels: (a) observed activity with the token path and goal star on the final frame; (b) the network's shape — 3D angle arrows over the chemical field, with the hidden law as the white dashed line; (c) the geometric coupling a_ij (the angles as a map) forming along it.

Final frame — the learned shape vs the hidden law

Measured (4 episodes, noisy observations, no labels, no backprop)

The channels are complementary: the shape alone fails where the weights alone succeed and vice versa; the composed material routes 100% at exactly oracle-level steps — the network plays the maze as well as a model given the true law, from observations alone.