16.1 tests needed for a wrong program with error rate eps to survive with probability at most delta
  eps=0.01   delta=0.05  exact n=299    sufficient bound ln(1/delta)/eps -> 300
  eps=0.01   delta=0.01  exact n=459    sufficient bound ln(1/delta)/eps -> 461
  eps=0.1    delta=0.01  exact n=44     sufficient bound ln(1/delta)/eps -> 47
  eps=0.001  delta=0.05  exact n=2995   sufficient bound ln(1/delta)/eps -> 2996

16.2 mean and standard deviation of the number of distinct tasks among N draws from M equally likely tasks
  formula: mean 136.48, sd 4.62;  simulation: mean 136.44, sd 4.61

16.3 mean-field time to reach the target (p_9 >= 0.9) under fixed schedules, eta = 2, theta_0 = -2
  greedy (best level each step): 41
  uniform over levels          : 192
  hardest level only           : 3000

16.4 exact error rate of the upper-median program by enumerating count vectors (lengths 1..9, values 0..9)
  exact error rate = 0.3494   (Listing 16.4 simulated 0.3462)

16.5 leakage by seed range: probability that a test task occurs among 1000 training tasks
  M =    243: predicted 0.9838   (Listing 16.2 observed 0.970 on 200 test tasks, standard error 0.012)
  M =  13122: predicted 0.0734   (Listing 16.2 observed 0.085 on 200 test tasks, standard error 0.020)

16.6 a program that is correct on every input without duplicates can still be wrong with duplicates (set-based median)
  agrees with the truth on 1800 distinct-valued lists: True ; on [0, 0, 0, 5, 9]: 5 vs 0
