A strategy is presented for refining anomalous scattering models and calculating protein phases directly from the Bijvoet and dispersive differences of a macromolecular multiwavelength anomalous diffraction (MAD) experiment. This procedure, incorporated in the program MADPHSREF, is especially amenable for exploiting the weak perturbations to normal scattering produced by inner-shell electronic transitions of asymmetric metal and protein ligand assemblies. The protocol accounts for more than one type of anomalous scatterer, incorporates stereochemical restraints, treats the data in local scaling groups, and partly compensates for correlated errors. Approximating maximum likelihood by averaging observation variances and covariances over all values of phase considerably improved error estimation. Probabilistic rejection of aberrant observations, re-evaluated before each refinement cycle, improved refinement convergence and accuracy compared with other less flexible rejection criteria. MADPHSREF allows the facile combination of MAD phase information with phase information from other sources. For the suifite reductase hemoprotein (SiRHP), relative weights for MAD and multiple isomorphous replacement (MIR) phases were determined by matching histograms of electron density. Accurate metal-cluster geometries and the associated errors in atomic positions can be determined from refinement against anomalous differences using normal scattering phases from a refined structure. When applied to MAD data collected on SiRHP, these methods confirmed the Fe(4)S(4) cluster asymmetry initially observed in the refined 1.6 A resolution structure and resulted in a MAD-phased, experimental, electron-density map that is of better quality than the combined MAD/MIR map originally used to determine the structure.