Errors are a fact of life, including in the Protein Data Bank (PDB) (which we rely on for a lot of applications including training and evaluating deep learning models). These authors had previously built a tool to spot inconsistencies between a structural model and a computationally-predicted one, and decided to apply it to detecting register errors (for example, when the main chain may be broadly correct, but residues are systematically assigned the identity of a residue a number of amino acids up or down in the sequence). Their tool identified thousands of likely register errors, but they could also suggest corrections for these errors that mostly improved refinement statistics. This could be a useful approach for detecting and fixing database deposit errors to make sure we're working with the best data possible.