wooly99/geng-academic-fraud-detector

This repository is a skill for AI coding agents that helps a user check an academic paper for signs of fabrication. The README, written in Chinese, says the project was inspired by a Bilibili video creator known as Geng Tongxue, who over a span of 36 days publicly identified five young researchers at four major Chinese universities for what he argued were faked results. Online viewers called the episode an earthquake in the academic world, and the author packaged the methodology he used into a reusable agent skill. Installation goes through the npx skills tool, with one command: npx skills add followed by this repository's URL. After that, the skill is added to your agent's skills directory. To use it, you start a chat with your AI agent and ask it in plain language to check a paper, passing the path to a PDF file. The agent then opens the PDF and walks through the checks listed in the README. The checks are organized into six categories, called the six methods of Geng Tongxue. The first is image reuse, looking for the same picture rotated, flipped, or cropped across different experiments. The second is data fabrication, including suspicious value distributions or implausibly clean numbers. The third is image splicing, especially in Western blot lanes with inconsistent backgrounds. The fourth is statistical anomalies, such as p-hacking and mismatched sample sizes. The fifth is output anomalies like impossibly high publication rates. The sixth is internal contradictions in the methods section. The README includes a smoke test on a real PLOS ONE paper that had been retracted. The skill correctly flagged image reuse between certain figures, fabrication in another pair of figures whose so-called independent experiments had identical raw counts, and methodological contradictions about ethics review. A sample output report is shown in markdown form, with a Geng Tongxue style commentary at the end. The author also lists clear limits. The skill works at the level of visual understanding rather than pixel-level forensic analysis, so it cannot do Error Level Analysis. It can only reason about what the paper itself shows, not the underlying data. False positives are possible, and the author asks readers not to treat the output as the sole basis for accusing anyone. The license is MIT.