Reward-trained circuit in the native fly connectome

Loads 164,587 native neurons and 25,563,197 biological synapses. The 32-channel observation stimulates 1,024 selected native sensory neurons. Ninety-six parameters are shared across 49,152 engineered added connections into 96 native decision neurons. Biological weights stay fixed; reward search changes the added connections.

An adaptation of the installed flyOnlineLearning parameter-search approach tests positive and negative random weight perturbations on matched 2015 episodes. All 768 training actions come from actual GPU native spike counters. The objective is net portfolio log return after costs. Cash-only episodes earn zero, profitable episodes positive reward and losing episodes negative reward. Equal returns cause no parameter update. Both initial action drive and perturbation distributions are symmetric across actions; no BUY/HOLD/SELL target labels are fitted.

Sixteen paired updates are a limited demonstration, not a strong optimization budget. Profitable training trials, active trading or a winning historical replay do not establish forecasting skill. Evaluation dates have been inspected during development and are no longer a pristine research holdout.

Evaluation freezes added weights as well. A decision uses a locked native spike sample while the brain continues evolving for live telemetry. The cloud uses broker.neuralActivity: ongoing trace is compressed without clipping at 1, and actual current-step spikes flash white. Quiet neurons stay dim. Display samples at frame rate and cannot expose every millisecond spike.