Verification

Run accounting-and-sensors.luau for historical data integrity, chronological separation, known real prices, sensor noninterference, fees, risk constraints, dividend entitlement/payment and ledger conservation.

native-regression.luau is an isolated native runner; it refuses to allocate a competing brain beside a live scene. Advance until periods finish, then finish to persist the audit. Ablation tests measured zero spikes and HOLD, without requiring a profitable or active policy.

historical-evaluation.json preserves the first historical experiment, including its then price-only classification score. historical-evaluation-final.json is the final dividend-correct source evaluation. Each contains all five windows and actual native evidence. No hyperparameter was selected against these windows. These are five reset accounts, not one compounded return. Repeated test exposure is disclosed and the sample is too small to establish general predictive skill.

Current reward version: reward-invariants.luau verifies positive/negative/zero rewards, additive log-return accounting, paired 2015-only tapes, and actual dollar value of a one-share $1,000-account fill. reward-training-1000.json records 768 native training decisions, 16 reward-driven weight updates and actual silencing. reward-evaluation-1000.json records the full continuous smaller-account replay. reward-training.json and reward-evaluation.json preserve the earlier $10,000 run with 5% lots; they are historical artifacts, not current account defaults. continuous-replay.json is the older supervised all-cash policy's automatic-transition test. neuron-texture-diagnosis.json records actual GPU trace saturation; live-neural-activity.jpg is a two-frame motion comparison. Older classification-case regression remains as data/audit coverage, not the current training objective.