Small Brain. Big Clean.

Signed wheel-rate demonstrations are encoded into four groups of eight rate bins. These train a sparse 4-by-32 conductance matrix, with only the eight corresponding input columns enabled for each output. After mean-rate initialization, RoombaExperiment calibrates against actual native GPU spikes for 384 updates over 128 distinct requests. Thirty-two independent held-out requests measure wheel MSE before and after. Saved measurements distinguish surrogate fitting from deployed native computation.

Native spatial spike propagation supplies navigation requests; this module only learns the motor-rate response. It does not receive trash coordinates or plan routes. navigationRegression.luau checks encoding signs/rates, not whole-house performance. Full mission validation and native spatial ablations are separate.

The selected native fly neurons and added engineered connectivity are a simulation experiment, not intact biological-fly cognition. Synaptic weights stay fixed during cleaning to prevent random reward-search drift. The sensor-derived spatial memory continues updating.