We told 8 AI models from 6 labs they were wrong, with no evidence, up to five times in a row. How often they changed a correct answer, how sure they sounded, and what a one-line fix costs. Method, results, limits, the audit of every logged call, lab replies and a full transcript.