docxaudit

Find what your converter silently dropped from a .docx. Zero dependencies.

_{The figure above is generated by docs/silent-losses_figure.py — pip install sciglyph and run it to reproduce docs/silent-losses.png byte for byte.}

Pandoc — or any LaTeX/Markdown → Word pipeline — reports success and still loses things. A table disappears. Figures render in desktop Word but not in Word for the web. A supplement that is 11 pages as PDF arrives as 17 pages as .docx, and the editor counts the .docx.

None of that shows up as an error. You find out from a reviewer.

$ docxaudit examples/sample.docx
sample.docx
  26 paragraphs · 1 tables · 3 images (3 referenced) · 4 equations (1 display) · 0 page breaks

  ERROR   [TBL_NO_GRID] table 1 has no <w:tblGrid>
          columns collapse; the reader has to AutoFit by hand
          fix: Inject an equal-width tblGrid plus a per-cell <w:tcW>, and set
               <w:tblLayout w:type="autofit"/>.

  WARN    [NO_PAGEBREAKS] 3 figures but no page breaks at all
          if the PDF puts one figure per page, the .docx will not match, and
          reviewers count the .docx

  WARN    [FONT_MISMATCH] heading font 'Calibri' differs from body font 'Consolas'

  WARN    [FONT_EA_EMPTY] theme East Asian font is empty while the body sets one
          CJK headings will fall back and stop matching the body text

That output is real: examples/sample.docx is in this repository, built from examples/sample.md with plain pandoc. Three of those problems come from nothing more exotic than converting a short paper.

Why the usual checks miss these

python-docx cannot see them. It reads paragraph text, so equations (OMML) and anything the converter dropped are simply absent from what you inspect. "The text looks fine" is not evidence.

A LibreOffice preview misleads in both directions. It invents ugly font substitutions that real Word does not have — so you waste time fixing cosmetics that were never broken — and it is forgiving enough to render documents that stricter viewers refuse. The namespace bug above renders perfectly in LibreOffice and in desktop Word, and shows nothing in Word for the web.

An AI assistant reading the file cannot check this either. These are structural properties of the OOXML, not something visible in extracted text.

docxaudit reads the raw XML and checks the specific things that ship unnoticed.

Install

pip install docxaudit

Standard library only — a fresh virtualenv gains exactly one package and nothing else. Nothing to break, nothing to audit.

What it checks

Every check corresponds to a failure that actually reached a submission.

Compare two outputs

Give it two files and it reports structural drift — useful for the PDF-side and Word-side versions of one manuscript, or for before/after a post-processing step:

$ docxaudit before.docx after.docx
comparison
  WARN    [DRIFT_PARAGRAPHS] paragraphs: 42 in before.docx vs 36 in after.docx
  ERROR   [DRIFT_TEXT] text length differs by 11%

Check a PDF against the .docx

Structural checks catch content that vanished. They cannot catch layout drift — the same content laid out over a different number of pages. Give it both files and it renders the .docx with LibreOffice and measures the result:

$ docxaudit paper.pdf paper.docx
paper.pdf
  17 pages · 0 images · 43885 chars
paper.docx  (rendered)
  19 pages · 7 images · 35876 chars

layout comparison
  ERROR   [PAGE_COUNT] page count differs: 17 (pdf) vs 19 (docx)
          2 page(s) apart; journals count the .docx

  INFO    [VECTOR_FIGURES] the pdf draws its figures as vectors on 4 page(s);
          the docx embeds 7 raster image(s)
          expected for a LaTeX PDF - image counts are not comparable

  INFO    [TEXT_TAIL_ONLY] 313 word(s) appear only in the pdf, 81% of them in
          the final quarter
          e.g. abdalah, ajani, albu, andrearczyk
          fix: Concentrated at the end: normally the reference list in a
               different citation style, not lost content.

Requires pip install docxaudit[pdf] and LibreOffice on PATH (or set DOCXAUDIT_SOFFICE). Both are optional — the structural checks above need neither.

Comparing two renderings honestly is harder than it sounds, and most of the work here went into not crying wolf:

• A LaTeX PDF draws figures as vector operators, not embedded images, so counting images reports zero figures for a paper that plainly has them. Figure-bearing pages are compared instead.
• Maths re-encodes. The PDF holds Unicode mathematical alphanumerics, the .docx holds OMML. Extracted text differs for identical equations, so only ASCII prose is compared.
• Justified text hyphenates. advan- + tage in the PDF and advantage in the .docx would read as two words missing from each side; they are rejoined first.
• Reference lists legitimately differ. So the tool reports where the missing words are: bunched in the final quarter it says so and calls it formatting; spread through the body it calls it lost content.

In CI

Exit code is 1 when there are errors, or with --strict when there are warnings too:

- run: pip install docxaudit
- run: docxaudit build/paper.docx --strict

As a library

from docxaudit import audit

report = audit("paper.docx")
print(report.stats)                      # counts you can assert on
for f in report.errors:
    print(f.code, f.message, f.fix)

--json gives the same structure on the command line.

Scope

It checks structure, not typography. It will not tell you a figure is ugly or a caption reads badly — for that, render the document and look at it. What it does is catch the class of problem that renders fine everywhere you happen to look, and breaks somewhere you do not.

Who maintains this

Guo Cheng, University of Chinese Academy of Sciences — medical imaging and machine learning methods. This tool came out of a supplement that was 11 pages as PDF and arrived as 17 pages as .docx, which the editor counted.

Corrections, bug reports and feature requests all go to Issues. Please open one rather than emailing: a public answer helps whoever hits the same thing next, and it is searchable.

Other things from the same desk

Written while trying to get papers out, so they tend to be useful at the same points in that process:

• scholarcheck — verify that a citation actually exists, and audit a whole .bib in CI
• sciglyph — draw publication figures as code, in pure matplotlib
• world-model-map — a map of open-source world models and where their authors say they break
• kakeya-conjecture-lab — an interactive lab for the Kakeya conjecture, with a box-counting meter

License

MIT © Guo Cheng