== DFLASH2 K=7 chain
  arm's faster prefill shows there, decode arithmetic being identical by census.
  arm       pooled rate  tokens/step   delta pp vs DFLASH             95% CI  e2e tok/s  completion tokens
  DFLASH         0.3608         3.54                +0.00 [+0.00, +0.00]      72.06               9669
  DF_OFF2        0.3608         3.54                +0.00 [+0.00, +0.00]      72.18               9669
  DF_T           0.3717         3.61                +1.07 [-1.26, +3.79]      82.93               9554
  DF_G           0.3628         3.54                +0.38 [-2.97, +3.88]      80.36               9541
== NATIVE MTP-5 chain
  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

