Data for: Trust Without Competence: The Novice Verification Problem in AI-Assisted Learning and a Four-Step Manual Countermeasure

Full question/answer key, unassisted and protocol-condition answers, claim- and citation-level blinded verification records, and aggregate results for a 60-answer pilot (30 questions x 2 conditions) across six domains unfamiliar to the author, plus the pilot's raw first run (including a disclosed mid-pilot agent failure and its repair) for full audit transparency. Answers and verification were produced by instances of one large-language-model family, as described in the paper. One incidental email address was redacted from two files in the raw first-run folder before upload (see package README). Recomputation confirms the paper's headline numbers (accuracy 0.93 to 0.98; citation-exists 0.22 to 1.00; citation-supports 0.33 to 0.99).