For Finding 8 we re-checked test labels: the messages where at least 2 of the 4 models agreed on a different category than the Banking77 answer key. Each was judged blind, with the two labels shown in random order. This page lists every verdict, so you can check our calls.
Changed (): the alternative is right. Both fit (): either is fair. Original kept () and unclear (): the official label stays.
Who decided: Claude drafted every verdict without seeing which label was official or which model said what; the author approved them in bulk, without overrides. The rule was written down before any re-scoring.
In the model columns, this style means the model picked the alternative, plain means the original label, and this style means something else.