Swapping citations between identical numbers leaves the arithmetic alone, so the test asks whether the cited fact actually supports the calculation. The researchers report strong results on natural records, but also missed financial-role errors and sensitivity to table labels. These exploratory tests do not establish a dependable financial checker.
How you can use it
A report may contain the same number for different years or categories. This study tests whether an AI checking a calculation can distinguish the right citation from one that merely has the right value.
Its method is to swap citations between identical numbers and check whether the AI notices the changed financial role. That lets researchers examine what financial evidence checking can and cannot do beyond matching numbers.