A = A
B = B

phase                           A /entry    B /entry    change       n
──────────────────────────────────────────────────────────────────────
iteration/compile                      —      50.0ms only in B       1
iteration/train_step              40.0ms      60.0ms   +50.0%?     10
iteration                         28.0ms      38.3ms   +36.9%~     50
…checkpoint_to_object_store       47.5ms      47.5ms    +0.0%?      2
iteration/env_step                 5.4ms       5.4ms    +0.0%      40
iteration/mcts                    16.9ms      16.9ms    +0.0%      40

  ? = fewer than 30 entries in one run; too thin to read
      as a regression. These are means, not a significance test.

  ~ = the mean moved but the median did not (or vice versa), so the
      change is in the tail rather than a shift of the whole phase:
      iteration/train_step    mean 1.50x   median 1.05x
      iteration               mean 1.37x   median 1.02x