"Change-averaged runs at about twice nominal alpha" is not what it does: measured¶

interval-calibration-change-averaged.json, interval_calibration.py --estimator change_averaged, the same null, window, arm, trials and resamples as the pooled interval-calibration.json. The claim came from the audit of this log (item 3), which found no artifact behind it.

Against a nominal 0.05 (Monte Carlo error 0.003 to 0.004, each row's standard_error). Change-averaged runs somewhat above pooled at every size and near double nominal only at ten changes; across the quoted 48 to 91 (ladder points 45 and 91) it is 1.1 to 1.3 times nominal, not twice. The estimand decision does not rest on this: pooled is registered and re-derives on the current H1 (audit). The sentence is corrected wherever it is quoted.