Only the bottom of an eight-model leaderboard was statistically stable, and half the endpoints disappeared in ten weeks.

arXiv 2609.30074 bootstrapped its own ranking across 8B to 675B models with caching disabled. The two worst models held rank in 99% and 86% of replicates, the middle four in 27 to 48%, the top two in 68%. Two defensible rules for merging runs changed four of eight rows. Four of eight API endpoints were withdrawn inside ten weeks, so the study can't be rerun. Small-prompt-set leaderboards should be reporting rank stability, and almost none do.