Before the run
Specify the hypothesis, data snapshot, strategy family, parameter budget, split usage, costs and failure criteria. Register the experiment before opening the result. Changing the design creates a new experiment, not a revised historical success.
During the run
Keep signals separate from fills. Apply previous-close decisions only at the next eligible execution time. Record costs, rejected orders and terminal handling. A model should not invent information about intrabar order sequence or queue position.
After the run
Inspect the ledger and an independent accounting sample, not only the summary. A positive screen remains private. A negative result may be published only when rights and disclosure scope permit it. The report must distinguish absence of evidence from evidence of absence.