Triplet phases recorded from insulin crystals were used to measure the improvement of phases during model refinement and to quantify the contribution made by each step in the refinement. Conventional amplitude data were recorded to 1.5 A resolution from rhombohedral pig insulin crystals using 1.54 A Cu Kalpha radiation. An initial atomic model and starting phases were obtained from a published structure and the atomic model was refined against the amplitude data using CNS. The refined phases were compared with 800 triplet phases that were measured from similar crystals using a three-beam interference technique and 1.1 A wavelength synchrotron radiation. The solvent region was improved further using a novel density-modification procedure. Calculated triplet phases were obtained from the model after each step in the refinement and were compared with the recorded triplet phases. The average difference between the recorded triplet phases and the calculated triplet phases was used as an unbiased measure of the correctness of the model at each stage in the refinement. The average individual phase error was estimated from discrepancies from triplet phases after each refinement step. Conventional atomic refinement of an approximate starting model reduced the average individual phase error from 21.6 to 14.7 degrees. Improvement of the solvent region, including the difference-map flattening procedure, reduced the individual phase error by a further 2.6 degrees. Modeling the discrete disorder of four amino acids accounted for an additional 0.5 degrees improvement and the final individual phase error was 11.6 degrees.