cff-version: 1.2.0
message: "If you use this test set, please cite it as below."
title: "Perch trap set: 70 clips that catch AI scribes inventing drug names"
version: 1.0.0
date-released: "2026-10-02"
authors:
  - name: "Talcoe LLC"
type: dataset
doi: "10.5281/zenodo.23113470"
identifiers:
  - type: doi
    value: "10.5281/zenodo.23113470"
    description: "Version 1.0.0"
  - type: doi
    value: "10.5281/zenodo.23113469"
    description: "Every version"
  - type: url
    value: "https://huggingface.co/datasets/Talcoe/perch-trapset"
    description: "Hugging Face mirror"
license: "CC-BY-4.0"
repository-code: "https://github.com/perchvet/perch-trapset"
abstract: >-
  A 70-clip blind test set (45 dictations, 25 traps) for checking whether an AI clinical
  scribe charts a medication that was never dictated. Clips are synthetic text to speech
  audio built from veterinary dictations written by an author who never saw the matching
  code. Ships with a generic Node scorer, a corpus file with ground truth per clip, and a
  full ledger of results from Talcoe LLC's own product, Perch, and one outside comparison
  run.
keywords:
  - veterinary
  - clinical scribe
  - speech recognition
  - hallucination
  - benchmark
  - AI safety