For each usable expanding-window forecast origin C, calculate log loss separately for the baseline model and the relevant PrF or construct-discrimination model for every eligible held-out polity at that origin.

Within each forecast origin, calculate the unweighted arithmetic mean of log loss across eligible held-out polities separately for the baseline and relevant model. This produces one baseline log-loss value and one model log-loss value for each usable forecast origin.

Pooled temporal baseline and model log loss are then calculated as the unweighted arithmetic means of their respective origin-level log-loss values across all usable forecast origins. Each usable forecast origin therefore receives equal final weight regardless of the number of eligible held-out polities at that origin.

Define pooled temporal relative log-loss improvement as:

R_T = (LogLoss_Baseline,T - LogLoss_Model,T) / LogLoss_Baseline,T

This temporal log-loss pooling rule applies wherever pooled temporal relative log-loss improvement is required for the full IDP model and the registered O* and E* construct-discrimination models.

This clarification changes no other aggregation rule, model specification, predictor, transformation, outcome, horizon, exclusion rule, feasibility threshold, grouped-cross-validation procedure, bootstrap procedure, robustness analysis or evidence-category rule. No alternative temporal log-loss weighting scheme may be selected after outcome access.