Paper Data Fabrication Detector

Problem

When reviewing a paper, numeric tables are hard to inspect by eye for copy-paste, constant offset, scaling, formula-generated values, or cross-table reuse. This skill focuses on research PDFs that contain numeric tables. It does not infer author intent or produce a verdict; it reports quantifiable statistical red flags such as digit bias, constant differences, perfect arithmetic sequences, and repeated values across tables.

How It Works

• Input: an absolute PDF path provided by the user; Excel and CSV are not supported yet.
• Extraction: numeric tables are extracted from the PDF, and the report name may use PMID from metadata when available, falling back to a timestamp.
• Detection: nine built-in statistical checks run on the tables, including last-digit chi-square, fixed differences, fixed ratios, arithmetic/geometric patterns, decimal repetition, rounding grid effects, Benford’s law, duplicate columns, and cross-table duplicates.
• Output: a standalone HTML report is generated, for example outputs/check_<PMID>.html, with risk levels from no anomaly to high risk. A risk_score >= 80 indicates that the raw data should be manually reviewed.

Boundaries

This is a data-forensics tool, not a judge. Text- or image-heavy papers may return “insufficient evidence,” which should not be read as passing. GRIM is deferred to v2.0 because it requires extracting sample size n. High scores mean statistical anomalies, not proof of misconduct; conclusions still require lab records and raw data verification.