Three verdicts as data
Run the code. Three citations in a drafted memo, each checked against what the citator actually holds — and notice that not one of the three survives untouched. That ratio is pedagogically exaggerated, but the categories are not: these are the four ways an AI-cited authority fails, in the order you check them.
STRIKE — not found. In re Halvorsen Marine is not in the citator because it does not exist. It has a plausible name, a pinpoint, and a quote that sounds like admiralty law. That's the Mata pattern: everything about the cite is well-formed except the part where the case is real. Existence is checked by lookup —CITATOR.get(...)— never by asking the model, which will confirm its own invention with the same fluency it used to invent it.STRIKE — quote not in source. Not triggered in this run, but it's the check that catches Couvrette's 8 fabricated quotations: a real case, opened, read — and the quoted sentence isn't in it. The rule is verbatim or it's a failure. "Close paraphrase presented as a quote" is not a lesser offense; it's the same offense with better camouflage.STRIKE — negative treatment. Draeger exists and the quote is on the exact page cited. It's also no longer good law. This is the failure that has embarrassed lawyers since long before AI — the model just produces it faster and with more confidence. Existence and accuracy tell you nothing about treatment; the citator's flag is a separate, mandatory read.FIX — wrong pinpoint. Right case, wrong page: Kestrel is real, good law, accurately quoted — at page 415, not 412. The one category that doesn't kill the authority, just the cite. Fix it and keep it. This matters because a gate that can only say keep-or-strike teaches people to ignore it; a gate with a repair verdict gets used.
Two design choices worth stealing for any checklist you build at work. The order is fixed: existence first, because the other checks are meaningless against a case that isn't there; treatment before pinpoint, because there's no point perfecting the cite of an authority you're about to strike. The verdict is data, not vibes: every citation ends the pass with exactly one of four strings attached to it, which means the pass can be logged, counted, and — in lesson three — signed. "I looked them over and they seemed fine" is not a record. KEEP / FIX / STRIKE per authority is.
One more thing about the fixture: the cases here are invented on purpose, because this is a drill and client matters don't belong in training exercises. At your desk, CITATOR is a real research platform, and the one dict is really two reads: you open the case text itself to check the quote and the pinpoint, and you read the citator for treatment. The fixture flattens both into one lookup; the discipline is identical. The stakes are not.