How does Empirical Bayesian Kriging of ArcGIS Pro handle sampling bias?

In ArcGIS Pro, simple kriging approaches allow you to de-cluster your data before applying a normal-score transformation. According to the Esri guideline, normal-score transformations are only feasible if the input data does not exhibit a strong sampling bias.

Empirical Bayesian Kriging does allow you to apply a normal-score transformation to the data before interpolation, but does not allow you to (at least manually) apply de-clustering algorithms on the data. Following that logic, EBK must either apply some sort of de-clustering technique automatically or must otherwise be suited to properly produce a prediction surface from geographically biased input data. Neither the Esri guidelines/publications nor other sources provide any information on this.

How then does EBK handle sampling bias?