🌐 **Live Demo:**https://fraudscreen-enwajvuqeykfe4prwaw8z7.streamlit.app/ Research Source Data Numerical Anomaly Screening Demo
fraud_screen is a Streamlit-based demo for screening numerical anomaly patterns in scientific paper Source Data. It helps reviewers, editors, research integrity teams, and meta-researchers inspect uploaded Source Data files, identify selected numerical patterns, and organize manual review priorities.
Important: Statistical anomalies are not evidence of research misconduct. fraud_screen is a screening and prioritization tool, not a misconduct determination system.
fraud_screen is a web demo for first-pass numerical screening of paper Source Data, Supplementary Tables, CSV files, and related research data tables. It reads uploaded files, extracts numerical records, applies anomaly-pattern detectors, and generates structured reports for human review.
The tool is designed to answer practical review questions:
- Which files and sheets were successfully read?
- Which numerical records were extracted?
- Are there repeated vectors, rounded duplicates, fixed mathematical relationships, unusual summary statistics, or weak digit-pattern signals?
- Which figures, sheets, or files should be reviewed first?
- Were target figures mentioned in a case note or retraction notice found in the uploaded Source Data?
- Can this run be traced back to the same input files, settings, and configuration?
Scientific papers increasingly publish figure-level Source Data and supplementary spreadsheets. These files are valuable for research integrity review, but manual inspection can be slow, inconsistent, and difficult to scale.
fraud_screen provides a structured screening layer. It does not replace expert judgment. Instead, it helps reviewers move from a large collection of files to a documented list of review priorities, source-data coverage notes, and reproducible run metadata.
- Excel / CSV Source Data ingestion with sheet-level inventory
- Processing Activity Log so long uploads do not look frozen
- Numerical anomaly screening for selected repeated, relational, digit, smoothness, and summary-statistic patterns
- Risk summary and manual review priority queue
- Target Figure Review Lite for matching user-specified figure/panel labels
- Case Review Report for organizing known concerns, target findings, anomaly summaries, limitations, and review recommendations
- Reproducibility metadata including input fingerprint, version, configuration hash, and run parameters
- Downloadable outputs in HTML, JSON, CSV, Markdown, and ZIP formats
- Built-in synthetic demo for testing without real paper data
| Category | Example signals |
|---|---|
| Duplicate vectors | Exact duplicate vectors, rounded duplicate vectors, highly similar standardized patterns |
| Constant relationships | Constant offsets, constant ratios, near-perfect linear relationships |
| Summary-statistic checks | SEM / SD / n inconsistency, repeated error values |
| Digit-pattern checks | Terminal digit patterns, Benford-style first-digit screening |
| Smoothness checks | Overly smooth or monotonic trends requiring manual review |
| Source-data coverage | Empty sheets, skipped sheets, text-only sheets, extracted record counts |
fraud_screen does not and cannot:
- determine whether research misconduct occurred;
- prove that data are fabricated or authentic;
- evaluate raw microscopy images, flow cytometry gates, lab notebooks, instrument exports, or author intent;
- replace expert human review, journal procedures, institutional review, or domain-specific interpretation;
- provide legal, editorial, or institutional evidence on its own.
A flagged result is a review signal, not a conclusion. A clean result does not prove data authenticity.
- Upload Source Data Excel / CSV files, supplementary tables, or a ZIP package.
- Optionally enter target figures or a case-review note.
- Click Start screening.
- Watch the Processing Activity Log while files and sheets are processed.
- Review the risk summary, manual review queue, source-data inventory, target review, and case review outputs.
- Download HTML, JSON, CSV, Markdown, and ZIP reports for manual follow-up.
pip install -r requirements.txtstreamlit run app.pyThen open the local Streamlit URL in your browser, usually:
http://localhost:8501
- Push this repository to GitHub.
- Log in to Streamlit Community Cloud.
- Create a new app from this repository.
- Use:
Main file path: app.py
- Deploy the app.
Recommended Python version: 3.11 or 3.12.
Use public or synthetic data for demo deployments:
- paper Source Data Excel files (
.xlsx/.xls); - supplementary tables;
- CSV files containing numerical data;
- ZIP packages containing multiple public Source Data files.
Do not upload confidential, unpublished, sensitive, or private research data to public deployments.
A run can generate:
- HTML screening report;
- JSON report for programmatic access;
- anomaly-detail CSV;
- manual-review priority CSV;
- source-data inventory CSV;
- target-figure review CSV;
- case-review summary CSV;
- case-review Markdown report;
- current-run manifest;
- ZIP package containing all run outputs.
Target Figure Review Lite lets users enter labels such as:
Fig. 5c, Fig. 5d, Extended Data Fig. 5c
The tool searches uploaded source-data metadata, extracted records, and anomaly details for matching figure or panel labels. It reports whether each target was matched, whether current numerical flags were found nearby, and what kind of manual review may be appropriate.
This is a lightweight text-matching review helper. It is not a deep forensic analysis.
The Case Review Report organizes case notes, target figure review results, anomaly summaries, data-reading coverage, limitations, and manual review recommendations into a structured Markdown report.
It is intended for research integrity case organization and manual follow-up. It does not make misconduct determinations.
Each run records metadata to help explain and reproduce results:
- fraud_screen version;
- input fingerprint;
- session ID;
- uploaded file names and sizes;
- configuration hash;
- detector configuration hash;
- target figures;
- deterministic-run flag;
- report paths and manifest entries.
This helps distinguish changes caused by different input files, versions, configuration, or run parameters.
- Local runs process files on the local machine.
- Public Streamlit deployments should only be used with public, synthetic, or non-sensitive files.
- Uploaded files and generated reports may be stored temporarily during a run.
- Do not upload unpublished, confidential, personal, or restricted data to a public deployment.
Statistical anomalies are not evidence of research misconduct. fraud_screen is intended for initial screening, exploratory review, and manual review prioritization only. All findings require expert human review, source verification, and contextual interpretation.
- Improve PDF table extraction and source-data linking
- Expand benchmark datasets for known anomaly patterns
- Improve cross-paper batch workflows
- Add richer documentation for detector limitations
- Explore additional report export formats
MIT License. See LICENSE.