Papers use ablations to justify an architecture, a data mix, or a training trick: drop the component, rerun the eval, report the delta. A missing ablation often means the claimed ingredient was never isolated. Results still depend on the rest of the recipe, so one ablation is evidence, not proof in every setting.