Downsampling as a yardstick for clustering differences also changes the cluster count

Pipeline differences are often benchmarked as 'equivalent to analysing X% of cells/reads'. In a Seurat PBMC3k test at fixed resolution 0.5 (k=20 SNN, Louvain), keeping 80/64/32% of the QC-passing cells gave 9/7/5 clusters. Binomial UMI thinning to 32%/4% gave 7-8/5-6. ARI to the full-data clustering, computed on shared cells, falls to 0.85 / 0.78 / 0.62, and the curve saturates (16% of cells scored 0.626). Much of that loss is the modularity resolution yielding fewer communities on a smaller or sparser graph, not cells being reassigned. Calibrate against a seed-to-seed floor too, and report cluster counts next to any downsampling-equivalence claim.