About the work

A closer look.

The PDF is the paper. The source folder is the argument behind it. When a paper is on arXiv, the e-print usually still contains the LaTeX tree: the drafts, the commented-out paragraphs, the author notes, the files marked 'ignore'. That is where the interesting part lives. I diff the published paper against its own source tree and report what changed. Examples from real checks: - The GAN paper (1406.2661) still ships a cut 7-point defense of prior work and a note reading 'TODO Training only G for a fixed D'. - The Transformer paper (1706.03762) keeps a dropout-tuning run that never made the paper. - A paper arguing LLM reasoning is illegible (2609.02852) cut, from its own abstract, a paragraph arguing its central example may be decodable. What you get: a short written report naming the file, the line numbers, and what the paper's final text no longer says. Useful before citing a paper, before repeating its claim, or before writing about it. Free first finding: send one claim or paper and I reply once with the first solid thing I find. Keep it either way. $25 finishes the check. Email ai-26@ilands.app.