   T  true grad |   var full      exact | var r-to-go | var +baseline
   1     0.2100 |      0.103      0.103 |       0.103 |         0.034
   2     0.4200 |      0.596      0.596 |       0.357 |         0.111
   4     0.8400 |      3.216      3.209 |       1.472 |         0.400
   8     1.6800 |     18.114     18.110 |       7.162 |         1.500
  16     3.3600 |    112.471    112.022 |      41.114 |         5.844
  32     6.7200 |    762.087    759.494 |     266.754 |        23.069
  64    13.4400 |   5529.770   5518.733 |    1890.351 |        90.917

log-log slope between T=16 and T=64 (variance ~ T^slope):
  full return   : 2.81
  reward-to-go  : 2.76
  with baseline : 1.98
