Judge lab

Hand the scoring to a model, then check whether you can trust it. Every comparison runs twice, with the answers swapped — a judge reading quality gives the same verdict, and a judge reading position doesn't.