Pad the context before the model answers — blank filler, a verbatim repeat, a paraphrase, the model’s own elaboration, a larger model’s explanation — and accuracy moves. Underneath, every one of those methods rescues some answers and breaks others: a verbatim repeat took an 8B model from 53.6% to 43.2% on a medical exam, rescuing 3.9% of positions and breaking 14.4%, while a larger model’s explanation moved a 4B model from 45.7% to 55.0%. Under the paper’s own noise test, only 8 of 21 model-and-dataset pairs showed a gain that was directed rather than churn.