exergyleizhou/paperguard-demo

PaperGuard Demo

Public demo for PaperGuard — a statistical anomaly screener for tabular research data. 40 detectors (academic 36 + industrial 4) covering terminal-digit, Benford, GRIM/GRIMMER/SPRITE, statcheck, Carlisle baseline balance, image duplication (pHash), EXIF temporal forensics, docx RSID forensics, file metadata forensics, AI-text heuristics, and more.

Paste a DOI or manuscript text → get back:

• —A Markdown summary of all findings, grouped by severity
• —A WCAG 2.1 AA HTML report (same one the CLI exports)
• —The full machine-readable JSON

What this is not

A fraud detector. Read the disclaimer in the app and the v10 recall study and T7/T8 endpoint scope for the honest performance characteristics.

Source code

• —Library + CLI: https://github.com/exergyleizhou-ux/PaperGuard
• —Install locally: pip install paperguard
• —This demo's app.py is a thin Gradio wrapper around the PaperGuard detector pipeline.
• —New in 2.15.0: OCR table extraction for scanned PDFs, Chinese scholar search via Semantic Scholar, batch author audit, author name disambiguation, EXIF temporal forensics, docx RSID forensics, decimal consistency detector.

License

MIT.

Why these pins?

• —python_version: "3.11" — HF Space default container moved to Python 3.13, which removed the stdlib audioop module that pydub (a gradio dependency) needs. Pin 3.11 to keep stdlib audioop.
• —sdk_version: 5.34.0 (gradio 5.x) — gradio 4.44 imports HfFolder from huggingface_hub, which was removed in huggingface_hub 1.x. Gradio 5.x uses the new API. (2026-05-22 ship learning.)