Random referees: two controls every ranking model should beat

Shuffled rankings and shuffled labels tell you whether a model learned anything, and whether your pipeline leaks.