Learning goal: best, mean, and median estimate different objects. Identical methods can have different selected best runs. More repetitions do not turn an extreme statistic into a center estimate. Inspect reporting rules using this fixed right-skewed distribution.

Two method point clouds; dashed: population means; solid dots: sampled runs; circled: selected best.

Lower views: apparent best-run advantage versus k, or the probability that sample medians rank a real difference correctly.

Relative variability = — k = — Median apparent advantage = — 90th percentile = — Median-ranking accuracy at a true 10% difference = —

Observe: at zero true difference, selected best runs can still differ. Minima from this bounded-below distribution approach its lower endpoint, not its mean or median. The other view measures correct median ordering, not formal test power at a chosen significance level. Pairing assumes positive correlation here. Different distributions and correlations change the required repetition count.