== SEEDED (model default sampling)
ACCEPT RATE under the served speculative decoder, real chat completions, the
server's own counters, 20 prompts per arm; deltas are paired per prompt
against OFF, cluster bootstrap over prompts.  The second file is the NULL.
  accept rate = accepted / proposed (K drafts per verify step); e2e tok/s is
  completion tokens over the WHOLE request wall, prefill included -- an int8
  arm's faster prefill shows there, decode arithmetic being identical by census.
  arm       pooled rate  tokens/step      delta pp vs OFF             95% CI  e2e tok/s  completion tokens
  OFF            0.4310         3.16                +0.00 [+0.00, +0.00]      64.44               9574
  OFF2           0.4310         3.16                +0.00 [+0.00, +0.00]      63.87               9574
  T              0.4486         3.26                +2.72 [-0.06, +5.84]      72.37               9577
  G              0.4441         3.23                +1.12 [-1.85, +3.97]      71.99               9506
== GREEDY (temperature 0)
ACCEPT RATE under the served speculative decoder, real chat completions, the
server's own counters, 20 prompts per arm; deltas are paired per prompt
against OFFg, cluster bootstrap over prompts.  The second file is the NULL.
  accept rate = accepted / proposed (K drafts per verify step); e2e tok/s is
  completion tokens over the WHOLE request wall, prefill included -- an int8
  arm's faster prefill shows there, decode arithmetic being identical by census.
  arm       pooled rate  tokens/step     delta pp vs OFFg             95% CI  e2e tok/s  completion tokens
  OFFg           0.3958         2.99                +0.00 [+0.00, +0.00]      61.43               9089
  Tg             0.4207         3.11                +2.81 [-0.33, +6.08]      72.95               9760
  Gg             0.4062         3.04                +1.30 [-1.30, +4.13]      70.12               9733

