Action and consequence · explore recorded action-credit training and frozen transfer →
Rows: adaptation strength 0, 0.3, 0.9. Tile score: median held-out R². One means perfect prediction; zero is the held-out mean baseline. Negative values remain visible.
Pale lines: individual controllers. Gold: their median. Dashed: direct-current-input control. Sampling interval: model time.
These are recorded experiments with the native Julia controller kernels. The readout learns to decode state; the cells are not trained to solve these tasks. Physiology and neighbors are clamped. A readable memory is a capability to investigate in the coupled material—not proof that a cell uses it.
Independent training and test signals; training-only feature normalization; fixed ridge penalty; no test-set fitting. Features are hidden state and slow adaptation. Identical-input twins begin with a small hidden-state difference. Sustained twin separation may reflect different attractors or sensitivity; it is not by itself a Lyapunov exponent or evidence of criticality.
The matrix changes recurrent-to-hidden weights and adaptation strength. Sensory weights, founder identities, response time, adaptation time, and the existing hidden plasticity law remain matched. Every condition keeps its full score vector. This view selects recordings and does not run Julia in the browser.
Download measured results · Half-step comparison · Methods and limitations