Native historical trading experiment

Trains the engineered added connections inside the native fly simulation with paired reward trials on 2015 data. Sixteen pairs of 24-day rollouts (768 actual native decisions) compare opposite perturbations of the same weights on identical historical dates and starting cash. Reward is 10000 × log(final equity / initial equity), including commissions, spread, slippage and dividends. No action or inactivity bonus, supervised action labels, forced order, or test-date feedback enters training.

The two returns move the 96 shared synapse parameters toward the better perturbation, with bounded updates. This is a small derivative-free reward-search experiment, not proof of predictive skill or a biologically validated learning rule. After an actual decision-neuron silencing check, weights freeze for the entire 2016–17 replay. Native motor spike counts alone select BUY/HOLD/SELL; the exchange may reject an infeasible order but cannot substitute a different action.

Starting cash is $1,000. Automatic two-second transitions carry cash, shares, costs, dividends, drawdown and comparison accounts across all five contiguous periods. The final dataset date stops playback and preserves the account; Restart replay explicitly starts over and is marked as repeated evidence.

A trading day takes 10/2.5/0.833 seconds at 1x/4x/12x. The decision window locks the action and its sampled rates; native activity continues between decisions without changing that locked action. Pause stops both. Random comparison accounts and a one-time 25% passive allocation use the same exchange assumptions.