Structure ahead of flexibility.

The two-body problem. Each network learns the motion from a family of orbits of equal energy and is then rolled forward 290 steps, two full periods, on an orbit it has not seen. White is the true orbit. Blue is ARCHON's dynamics module. Amber is the network named on the rail. The strip is each rollout's distance from the true orbit, on a log scale.

ARCHON's pane shows the module's state on its four manifolds and its learned network firing on the orbit being drawn, over its 40,000-step training loss. With 19,397 parameters it is 110 times closer in position than a Hamiltonian network, 187 times in energy and 572 times in angular momentum, and 34, 83 and 121 times ahead of a standard network. Given ARCHON's own coordinates a standard network still trails by 1.9, 6.1 and 3.5 times, which is the share that belongs to the structure alone. For ARCHON the readouts are its own medians, and for each network its error as a multiple of ARCHON's, all over 50 seeds.