On what to do when there is not sufficient agreement. It would indeed be great if we could have all unique representations (ideally already ranked most to least frequent, but this I can implement locally). My thinking was that in case there is no trustworthy, automatic resolution, at least we could have a good suggestion for manual cleaning. Pruning (incorrect) suggestions is easier than coming up with entries yourself.
From Robbert: