By characterizing a very large number of might-have-been evolutionary trajectories starting from a resurrected ancestral protein, the authors show that the evolution of an essential modern protein was contingent on extremely unlikely historical mutations. Can evolutionary biology become a predictive science? The answer to that question depends largely on whether it proves possible to develop a quantitative measure of the role of chance historical events in shaping evolutionary paths. With that objective in mind, Michael Harms and Joseph Thornton start from a database of thousands of variants of an ancestral form of the glucocorticoid receptor to look for mutations paving the way for the appearance of a larger-effect mutation creating a new ligand specificity, and they find none besides the historical permissive mutations. Their result shows that the evolution of this class of hormone receptors is critically dependent on rare non-deterministic events, constrained by protein biophysics. Evolutionary contingency is often seen in terms of chance external forces such as extinction by asteroid impact or climate change, but this work points to the internal organization of biological systems as a further powerful source of contingency. Understanding how chance historical events shape evolutionary processes is a central goal of evolutionary biology1,2,3,4,5,6,7. Direct insights into the extent and causes of evolutionary contingency have been limited to experimental systems7,8,9, because it is difficult to know what happened in the deep past and to characterize other paths that evolution could have followed. Here we combine ancestral protein reconstruction, directed evolution and biophysical analysis to explore alternative ‘might-have-been’ trajectories during the ancient evolution of a novel protein function. We previously found that the evolution of cortisol specificity in the ancestral glucocorticoid receptor (GR) was contingent on permissive substitutions, which had no apparent effect on receptor function but were necessary for GR to tolerate the large-effect mutations that caused the shift in specificity6. Here we show that alternative mutations that could have permitted the historical function-switching substitutions are extremely rare in the ensemble of genotypes accessible to the ancestral GR. In a library of thousands of variants of the ancestral protein, we recovered historical permissive substitutions but no alternative permissive genotypes. Using biophysical analysis, we found that permissive mutations must satisfy at least three physical requirements—they must stabilize specific local elements of the protein structure, maintain the correct energetic balance between functional conformations, and be compatible with the ancestral and derived structures—thus revealing why permissive mutations are rare. These findings demonstrate that GR evolution depended strongly on improbable, non-deterministic events, and this contingency arose from intrinsic biophysical properties of the protein.