Paired variance of the factorial preprocessing comparisons

Quantify the paired variance of the eight-cell factorial comparison of training-time normalization, inference-time normalization, and the feature-fusion branch by repeating the factorial across additional training runs.

Background

The paper reports the factorial comparisons using identical corpora, hyperparameters, and seeds within each paired comparison, but each configuration is otherwise based on a single training run. Although multi-seed experiments were conducted for one configuration, the authors state that they did not repeat the full factorial and therefore could not measure the variance of the paired cell-to-cell effects. Establishing this variance would clarify the statistical reliability of smaller differences between factorial conditions.