Does your model get better at task T when you rank by estimated probability p(T) ?
System design of an Early Ranker
Early (Stage) Ranking in recommender systems