verify rows: 3910   drafted/row: 7.00   accepted/row: 4.727   committed/step-row: 5.727

accepted-length distribution (accepted drafts per row)
   0      306    7.8%  ####
   1      389    9.9%  #####
   2      322    8.2%  ####
   3      304    7.8%  ####
   4      292    7.5%  ####
   5      253    6.5%  ###
   6      204    5.2%  ###
   7     1840   47.1%  ############################

first rejection position (2070 rows rejected something, 52.9%)
  slot  0      306   14.8% of rejections
  slot  1      389   18.8% of rejections
  slot  2      322   15.6% of rejections
  slot  3      304   14.7% of rejections
  slot  4      292   14.1% of rejections
  slot  5      253   12.2% of rejections
  slot  6      204    9.9% of rejections

the four cases, over 20552 TESTED slots (slots past a rejection are excluded)
  drafter right, draw kept it      -> no loss     16277   79.2%
  drafter right, draw REJECTED it  -> SAMPLING loss      287    1.4%
  drafter wrong, draw rejected it  -> DRAFTER loss     1783    8.7%
  drafter wrong, draw kept it      -> no loss      2205   10.7%
  -> of all rejections, 13.9% were tokens the drafter got RIGHT (sampling) and 86.1% were drafter misses

per-slot decomposition  (hit = draft token IS target argmax)
  hit%(all) is over every DRAFTED slot: unbiased drafter quality.
  hit%(tested) is over slots the prefix reached: conditioned on surviving,
  so it runs HIGH at depth and its n shrinks. Read the trend off hit%(all).

slot   n_all hit%(all)  tested hit%(tst)   E[p'top1|hit]    E[p'draft|miss]  alpha_obs
   0    3910     85.7     3910     85.7     0.9391+-0.0048     0.1451+-0.0120     0.922
   1    3910     75.4     3604     76.4     0.9351+-0.0055     0.0935+-0.0088     0.892
   2    3910     79.0     3215     83.0     0.9421+-0.0051     0.1149+-0.0114     0.900
   3    3910     73.2     2893     78.4     0.9426+-0.0057     0.1007+-0.0102     0.895
   4    3910     74.7     2589     83.6     0.9569+-0.0049     0.0963+-0.0124     0.887
   5    3910     64.6     2297     72.5     0.9605+-0.0058     0.0616+-0.0076     0.890
   6    3910     68.1     2044     82.9     0.9614+-0.0052     0.0849+-0.0129     0.900

reading: hit%(all) falling with slot = the DRAFTER degrades with depth.
         hit%(all) flat and high but alpha_obs low = SAMPLING rejects tokens
         the drafter got right; the bucket table above says so directly.

value of one more accepted token per row: +17.5% tok/s at the measured accept length (5.73 committed), 40 us/token commit cost assumed
