The Geostatistical Analyst (GA) radial basis functions are exact interpolators which means the input point values equals the associated layer values. And since zonal stats tools work on rasters (not GA layers), the user has to convert the GA layer to a raster. However... the export process doesn't reproduce the input point values at the associated raster cell.
The workaround is to use the GA Layer to Points tool on the raster centroids' points, perform a spatial join between the point centroid file and the polygons (in my case research plot boundaries) and calculate the average point centroid value within each plot polygon.
Converting the point centroid values to a raster created the same distortion as exporting the GA layer.
My suggestion is to let zonal stats tools work on GA layers based on user defined raster properties (i.e. cell size and extent). The zonal stats tool's code would use the matrix of GA layer values associated with the raster centroids in its analysis.
Also... I'm guessing that the GA Layer to Points tool works on a function stored within the GA Layer. Would it be difficult to expose that function for coders?