Small Brain. Big Clean.

Live cleaning uses a deliberately engineered intervention in the imported male fly CNS simulation, not an intact biological fly learning housekeeping.

Local depth, floor pixels, underbody dirt sensing and wheel odometry populate an initially empty spatial memory. CPU code stores observations and projects dirt onto nearby body-clear poses. It does not search for a route or receive hidden trash coordinates. 6,144 identified native interneurons propagate spike waves through observed free cells. A fixed latency/heading decoder selects a neighbouring pose from actual native spikes. Fixed differential-drive encoding stimulates 1,024 identified input neurons; trained synapses feed 128 native motor neurons. Their measured forward/reverse spike rates are the sole wheel actuators. The engineered wave circuit gates/reset its selected neurons between waves; native network dynamics are therefore intentionally modified.

The motor matrix has 128 slots, 32 trainable within-channel rate-bin weights and 96 zero cross-channel slots. These define 16,384 added motor connections. Training fits signed rate demonstrations and then updates weights against actual GPU neuron measurements. The checkpoint records independent held-out motor error. Motor weights remain stable during missions; observed spatial memory changes. The old random online weight search is disabled.

Physical body collision, individual debris contact, suction, track spreading and right-side-brush fluid transport determine cleaning. The brush transports mass inward; suction alone removes it. Completion still requires 99.9% liquid removal, and residual film remains physically simulated. No completion counters are synthesized.

New mess resets debris, memory and route. Learning exposes retraining, zero motor weights and motor-neuron silencing. public:status(), :trainingReport(), :spatialMemory(), :auditNeurons() and :calibrate() expose evidence. Overview/follow orbit controls, low right-shoulder GoPro, actual trajectory and Hide UI are preserved.