In this article we present a statistical fitting method for evaluation of atomic reconstructions which does not require a coarse-graining step. The fitting compares different models of chemical structure in their capability to explain the measured data set by a least square type merit function. Only preliminary qualitative assumptions about the possible chemical structure are required, while accurate quantitative parameters of the chosen model are delivered by fitting. The technique is particularly useful for singular defect structures with very high composition gradients, for which iso-concentration surfaces determined by coarse-graining become questionable or impossible. We demonstrate that particularly detailed information can be gained from triple junctions and grain boundaries.