gold reward values: [0.0, 1.333, 2.667, 4.0, 5.333, 6.667, 8.0]
Bayes-optimal accuracy on the test pairs = 0.763;  log-loss floor (true probabilities) = 0.461

 n pairs  accuracy  log-loss  Spearman
      50     0.618     2.039     0.462
     200     0.677     1.680     0.689
    1000     0.696     1.679     0.751
    5000     0.754     0.505     0.942

calibration of the last model (n = 5000, seed 2): bins of predicted probability that the first response wins
  bin  mean predicted  empirical   count
    1           0.054      0.100    8000
    2           0.233      0.290    8000
    3           0.502      0.505    8000
    4           0.769      0.709    8000
    5           0.947      0.896    8000
