Skip to content
LedgerLedger
← Your pathLevel 1 · Beginner

The three ways a model fabricates authority (case, section, revenue ruling)

Lesson 3 of 6 · Tax research that survives review

1 · Learn the move · Examples + few-shot

Fabrication has three shapes, and once you have seen each you spot it on sight. The case: a plausible name against the Commissioner, a Tax Court memo number in the right range, a holding that is exactly what you wanted, and no pin cite; the court has never heard of it, and the Deloitte report that invented a judge's quotation was this shape [S]. The section: a Code section number that exists, attached to a rule it does not contain, because the model learned that s. 162 is about expenses and filled in the rest. The revenue ruling: a number in the right format for a year that had rulings, saying the thing a ruling would say, which has either a different holding or no existence. The tell across all three is the same: confidence without a trail. A real citation can be opened; a real holding can be quoted; a real figure has a source and a year. Teach the model the three shapes with examples, and ask it to grade its own draft against them before you do. It will catch some. You catch the rest, by opening every one.

Here are three examples of fabricated authority and why each is fabricated: [paste the three shapes above, one invented example each]. Now grade this draft against them: [paste the draft]. For every citation, say which shape it most resembles if any (case without pin cite / section with a rule it does not contain / ruling number with no trail) and why, or 'no resemblance I can see'. Mark all of them VERIFY regardless. Do not remove or rewrite any citation; I decide after I open them.

2 · Your turn. You write the prompt

A model-drafted memo on Taxpayer C's worker classification cites 'Hendricks v. Commissioner, T.C. Memo. 2019-114' for the proposition that a signed contractor agreement is dispositive, 'IRC s. 3508(c)' for a safe harbor, and 'Rev. Rul. 87-41' for a twenty-factor test. One of the three is real and says something narrower. One is a real section about something else. One does not exist.

Write the prompt for this scenario as given, or swap in a real one from your own work. The score works the same either way.

Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.

Optional. These shape the output when you run your prompt below, not your score.