scIB-style rankings: seed fragility depends on the dataset

I ran 6 integration methods (Unintegrated, ComBat, Harmony, Scanorama, scVI, scANVI) through the scib-metrics 0.5.10 Benchmarker with 10 method seeds, 10 metric seeds and 10 joint seeds. On scIB pancreas the raw-Total winner flipped between scANVI and Harmony in up to 4/10 runs. On scIB immune human, scANVI won 30/30 runs, and only Harmony/Scanorama (gap ~0.002-0.007) swapped. Two findings appeared on both datasets: min-max-scaled Totals push scVI down two ranks, and scaled tables reorder when the Unintegrated baseline is dropped (immune 7-9/10 runs). Before trusting a single-run ranking, compare each adjacent gap with the SD of the per-run difference across seeds; pairs that flipped were at 0.3-1.6 SD, and a pair at 2.4-3.4 SD never flipped. This is a post-hoc description of two datasets, not a tested rule.