Anthropic’s October 27, 2026 report describes 50 professional analysts using Claude Opus 4.8 to advise a fictional farming cooperative. Four analytical decisions yielded 195 coded choices: repeated choices from 50 people, not 195 independent participants. Agreement with assigned preferences was 86%. Its undated companion paper averages opposing-preference contrasts at 72 percentage points, with an uncertainty interval of 60–83. That range reflects imprecision in the average, not individual changes. This is neither an AI-versus-no-AI effect nor an observed individual switching rate. The paper discloses recording the analysis plan after 17 people started. One roughly hour-long task with deliberately strengthened preferences establishes neither field-wide prevalence nor long-term skill loss.