Homeostatic surprise-modulated online learning — concept provenance

online.py (WM_ONLINE_LR_ENABLED, default off) is a CLEAN-ROOM implementation. Two concepts are inspired by the temporal-neuron research project (wizzense/temporal-neuron, TNRL v1.0 — non-commercial, share-alike):

1. Surprise-gated plasticity — a learner should change its weights most when its prediction was most violated, and consolidate (learn slowly) when the world is behaving as expected. temporal-neuron expresses this as surprise-modulated STDP; here it is an LR multiplier driven by the ratio of the current latent-MSE to its own slow EMA.
2. Homeostatic set-point regulation — a slow negative-feedback loop that holds an activity statistic inside a band rather than optimizing it. temporal-neuron applies this to firing rates; here the statistic is mean per-dim latent variance and the effector is LeWM’s SIGReg weight (sig_lambda): variance under the band means the representation is contracting toward collapse → raise the isotropy pressure; over the band → relax it.

License boundary: no code from temporal-neuron was read into, copied into, linked into, or derived into this file — an automated boundary scan asserts that mechanically on every change upstream. This document exists so the concept credit is explicit while the code stays clean-room.

Other inputs: LeWM’s SIGReg (Epps–Pulley isotropy test) is the collapse sensor/effector pair the homeostat drives; the surprise signal is the same latent MSE the LeWM gate scores.