========================================================================
POOLED COST (RTK ON vs OFF, reward-agnostic)   results/sweep-20260807-072640
========================================================================

task                           OFF n  ON n  OFF mean  ON mean   Δmean   Δmed
  dialogue-parser                  5     6      0.36     0.32    -13%   -15%
  fix-build-google-auto            5     5      2.36     2.50     +6%   -24%
  simpo-code-reproduction          6     5      0.92     0.91     -1%   -11%

------------------------------------------------------------------------
clean cost trials: OFF=16  ON=16   (3 task-pairs)
median of per-task MEAN   deltas = -1.1%
median of per-task MEDIAN deltas = -15.1%   <- skew-prone (compare_sweep view)

POOLED normalized (1.0 = each task's OFF mean, all trials):
  OFF mean = 1.000   ON mean = 0.968   ->  ON -3.2%   permutation p = 0.801
------------------------------------------------------------------------
Read-out: the POOLED line is the honest cost delta. p>=0.05 => cost-neutral.
If MEDIAN and MEAN deltas disagree in sign, the sample is skewed -- trust pooled.
