models.NeuralClock

models.NeuralClock(clock, weights, scale=None, notes=list(), reference=None)

A feed-forward network over a fixed probe set.

Forward pass only. Nothing here trains, and the architecture is deliberately the plainest one that fits published aging networks: dense layers, one activation, no dropout at inference, no batch norm state to get wrong.

THE ONE ARCHITECTURE HERE WHERE A GPU EARNS ITS TRANSFER. A linear clock is a single dot product over a few thousand features, and docs/dev/gpu.qmd measures the card losing to the CPU on those by up to 4.6x because moving the matrix costs more than multiplying it. A network is a chain of dense layers with real depth: AltumAge is 20,318 inputs, so the arithmetic per byte transferred is an order of magnitude higher and the transfer amortises. The whole pass therefore runs through the xp handle, activations included, and honours device= like the linear path does.