When "taxpayer" beats the client's name, and when it does not

1 · Learn the move · Cognitive verifier

Replacing the name is the first move and the one most people stop at. 'Taxpayer A' is enough when the facts are common: a wage earner with a rental, a partnership with a loss, a sale of a residence. It is not enough when the facts are rare: the only distillery in the county, a settlement reported in the newspaper, a founder whose exit was a press release. For those, the name is already gone and the client is still in the paste. The verifier is a question you make the model ask of each fact: how many taxpayers in the region share this fact? Then of the combination: how many share all of them? If the honest answer is 'a few', the fact gets generalised further or the question gets asked differently, through a hypothetical you build, not a case you describe. The model can run the count as a reasoning exercise. It cannot know the real number, and it cannot declare the result de-identified; that question is unresolved, and the line saying so goes under every result [M].

For each fact in this de-identified pattern, estimate how many taxpayers in the stated region plausibly share it, as an order of magnitude (thousands / hundreds / a few / one), and say why. Then estimate the same for the combination. For any fact or combination that comes out at 'a few' or 'one', propose a further generalisation that keeps the tax question intact, or a hypothetical I could ask instead. Do not tell me the result is de-identified; end with the line that this is not a safe harbor.