Peptides are key immune targets. They are generated by fragmentation of antigenic proteins, selected by major histocompatibility complex (MHC) molecules and subsequently presented to T cells. One of the most selective requirements is that of peptide binding to MHC. Accurate descriptions and predictions of peptide-MHC interactions are therefore important. Quantitative matrices representing MHC class I specificity can be used to search any query protein for the presence of MHC binding peptides. Assuming that each peptide residue contributes to binding in an additive and sequence independent manner, such "crude" matrix-driven predictions can be expressed as a quantitative estimates of binding strength. Crude matrix-driven predictions are reasonably uniform (i.e. precise), however, there is a general tendency towards overestimating binding (i.e. being inaccurate). To evaluate and possibly improve predictions, we have measured the MHC class I binding of a large number of peptides. In an attempt to further improve predictions and to include sequence dependency, we subdivided the panel of peptides according to whether the peptides had zero, one or two primary anchor residues. This allowed us to define unique anchor-stratified calibrations, which led to predictions of improved precision and accuracy.