provision
A tax benchmark that scores the reasoning and the answer as two separate results.
A tax answer has two ways to be wrong. It can miss, or it can land while the reasoning underneath it is nonsense. In tax the second is the dangerous one: a position is worth exactly as much as the argument filed alongside it, and a confident wrong route reads like a right one.
More than 1,500 questions authored and validated with practitioners, spanning compliance and application, comparative analysis, numerical reasoning, problem solving, semantic analysis, and the current-year updates that go stale fastest. Every response is scored twice — once against ground truth, once against the way an expert would have worked toward it.
The two scores come apart, and the gap is the finding. Models argue through tax questions considerably better than they compute through them, which tells a firm where a reviewer still has to sit and where one no longer does. Buying on a single blended number would have hidden it completely.
