Say that the fs rule errs toward removing more

README.md

@@ -168,6 +168,9 @@ pinned revisions. It needs `rdkit` and `pandas`.
   high: the upstream corpus covers few distinct conditions. Each removed row's
   correct mechanism description appears verbatim in the test split, so the cut
   is warranted even though no molecule is involved.
+- `fs` is keyed on the reactants whichever of its two questions a row asks, so a
+  byproduct training row that shares reactants with a major-product test row is
+  removed too. That errs toward removing more, not less.
 - Reaction subtasks are keyed on the reaction the question names, not on every
   molecule it mentions. Common solvents and reagents recur throughout both
   splits and keying on them would empty the training set for no gain.