Replication dataset for a manuscript submitted to AI & Society. This deposit contains coded datasets and tone classification results supporting an autoethnographic study of AI-assisted amateur research. An emergency medicine physician used ChatGPT-4o as an epistemic collaborator for three research projects (Collatz conjecture, cosmology, STS). All manuscripts were rejected after peer review. Core dataset: 2,566 unique assistant messages across 47 conversations, classified into four tone categories (E/C/M/N) via full census using Claude Opus 4. v3 update: Iterative calibration with refined criteria, v2 re-coding of all messages, and IRR verification (κ=0.557, 83.6% agreement). Key findings: E:C=3.2:1 overall; Collatz 1.1:1, ITC 4.6:1, CST 2.6:1. 172 files in full_census_v2.zip including raw data, v1/v2 classifications, input batches, and IRR validation files.