Did you use any ESDA tools to check the kriging assumptions? If the predictions are vastly different than the input points, it may be that your data is violating the assumptions for kriging, or it might be that your fitted semivariogram model isn't well specified.
By default in ArcGIS 10, the filled contours map is a surface for the predictions, and the map will display predictions at the input points, not the value of the input points themselves. If you want Ordinary Kriging to give exact predictions at the input point locations, you can set Measurement Error to 0. However, this will cause a jump in the surface at the input point locations.
