paperconan — paper data sanity check

Tool repository: https://github.com/zixixr/paperconan

When to use

Use this skill when the user:

• Hands over a directory of supplementary source data (.xlsx / .csv / .tsv) and asks for a data integrity / sanity check
• Wants to prep a PubPeer post and asks you to surface suspicious numeric patterns first
• Asks about a paper they suspect has fabricated data and wants statistical signal before deciding next step
• Says things like "帮我看看这篇论文数据有没有问题", "扫一下这个 source data", "查查这些表格的可疑模式", "audit this paper's data"

When NOT to use

• The concern is image fraud (Western blot, microscopy, gel splicing) — paperconan only inspects numeric tables
• The data is not tabular — paperconan reads .xlsx / .csv / .tsv only (not .xls, PDFs, or images)
• The user wants statistical methodology review or peer-review-style scrutiny — paperconan is forensic, not statistical

Prerequisites

pip install paperconan          # base install (xlsx / csv / tsv)
pip install "paperconan[all]"   # + supplementary PDF / Word table extraction
# Dev install from a clone:  pip install -e /path/to/paperconan

Verify with paperconan --help (or paperconan --version).

A complete worked example — synthetic data + the report it produces + a guided

walkthrough of every finding — lives in the repo's

examples/ directory. Read it to see the output

shape before running on real data.

How to invoke

Single command, takes a directory of data tables (.xlsx / .csv / .tsv):

paperconan <input-dir>
# Default output: <input-dir>/audit/scan.json + <input-dir>/audit/report.html

Common variants:

paperconan <input-dir> --out /tmp/audit-X     # custom output dir
paperconan <input-dir> --md                   # also write REPORT.md
paperconan <input-dir> --no-html              # only scan.json (CI / scripted use)

Exit code is 0 even when findings are present — findings are not errors.

Fetching a paper's data automatically

If the user gives a paper (DOI or title) instead of a local directory:

paperconan fetch "<DOI or title>"                 # list candidate datasets + match signals
paperconan fetch "<DOI>" --download <cand_id> --out data/   # download chosen candidate's tabular files
paperconan data/                                  # then audit as usual

Workflow:

1. Run paperconan fetch "". Each candidate has match_signals

(doi_in_related, title_overlap, author_overlap).

1. You decide the match — prefer doi_in_related: true; otherwise weigh title/author

overlap. If unsure, show the user the candidates and ask. Repository full-text search

(especially figshare/zenodo) often returns unrelated deposits, so fetch --auto

refuses to download a candidate with no DOI match / weak title overlap (it falls back

to journal guidance), and fetch --download of such a candidate requires --force.

A candidate flagged ⚠ no DOI/title match in the listing is probably not this paper's data.

1. Download the chosen candidate, then run paperconan on the output.

Honesty rules (REQUIRED)

• Searched repositories are Zenodo / Figshare / Dryad / Europe PMC.
• If a candidate has no .xlsx/.csv/.tsv, say so and name the other file types.
• If nothing matches, fetch now prints a journal-guidance block derived from the

DOI (publisher + doi.org article link + where that publisher puts source data,

e.g. Nature's ...MOESM_ESM.xlsx). Relay it — never imply "checked = clean".

• Do not bypass paywalls or scrape publisher sites.
• Download works keyless for Zenodo and Figshare. Europe PMC is also keyless:

for open-access papers it serves supplementary material as one zip, and fetch

downloads it and extracts the tabular members automatically. Dryad is

discovery/listing only — its download API needs authentication, so report Dryad hits

to the user and point them to the Dryad dataset page to download the files manually.

How to read the output

Three artifacts may exist in the output dir:

scan.json top-level schema

{
  "tool": "paperconan",
  "tool_version": "0.4.0",        // for provenance when the report is archived / shared
  "scanned_at": "2026-05-29T02:08:53+00:00",
  "input_dir": "...",
  "paper": {"doi": "10.1038/...", "title": "..."},  // provenance, or null (see below)
  "n_files": 3,
  "n_blocks_with_findings": 8,
  "relations_blocks": [
    {
      "file": "ED_Fig8b.xlsx",
      "sheet": "Sheet1",
      "block": {"rows": "6-15", "cols": "1-30", "header": [...]},
      "relations": [...],              // cross-column relations
      "progressions": [...],           // arithmetic progressions
      "equal_pairs": [...],            // pairs of columns with many equal rows
      "within_col": [...],             // within-column anomalies
      "identical_after_rounding": [...] // cells matching after rounding
    }
  ],
  // per-sheet last-digit χ². Each: {label, n, chi2, p, p_adj, fdr_significant, counts, top}
  // Filter on fdr_significant (BH-FDR q ≤ 0.05), NOT raw p — dozens of sheets are tested.
  "digit_distribution": [...],
  // per-sheet two-decimal ending counts. Each: {label, n, n_unique, top}
  "decimal_endings": [...],
  // bit-identical / value-overlap across sheets (same file OR cross-file). See fields below.
  "cross_sheet_findings": [...]
}

paper provenance is populated from a paperconan_source.json sidecar that

paperconan fetch --download/--auto writes alongside the data, or from

paperconan