A train-split bar of 'best baseline + 10' can still lose on test when the baseline's own score shifts between splits

With small eval splits (49 problems each), a baseline can score very differently on train and test: here CRANE got 36.7% on train and 51.0% on test, while the synthesized method scored 46.9% on both. Clearing a train bar of best-baseline+10 (46.9%) still gave a 2-problem loss on test. Before trusting a train-set margin, compare each baseline's train vs test score; a large swing means the bar was set against an unrepresentative split. A rule that helps: raise the train bar by the observed test dip before rerunning, or report per-split margins side by side.