Perch trap set: 170 clips that catch AI scribes inventing drug names and doses

Three blind test sets, 170 clips in all, for checking whether an AI clinical scribe charts a medication or a dose that was never dictated. The root set is 70 clips (45 veterinary dictations, 25 traps) in synthetic Kokoro text to speech, one voice per clip, with ground truth, a dependency free Node scorer, and a per clip ledger of Perch's own run and one outside comparison. Version 1.1.0 added corpus7, a second 70 clip blind set with a measure for dose numbers in the note that nobody said. Version 1.2.0 adds high-stakes, 30 blind clips built only from anticoagulants, antiplatelets and insulins, asked for by an emergency physician. It carries Perch's transcripts and charted lists for three runs, a per clip ledger, the fix the set led to, and a screen recording of the shipped build. It also adds a by drug class section to the root ledger and a --corpus option to the scorer. Every result is a raw count and none is a percentage. Talcoe, the company that makes Perch, built and ran this test itself.