Every number traceable to its source.

The problem

A number typed is a number that can drift

Say the abstract reads 4,000 reports. The data are cleaned, the analysis is run again, and it now says 3,986. The abstract still says 4,000, because someone typed it.

In a manuscript-guard project the abstract holds {{results.cohort.n_reports}} instead, and the build writes in the number the analysis wrote. Run the analysis again and rebuild, and every bound number follows. A number in a paper comes from one of three places:

1. Your results

   A binding into the results file your Python or R code writes.

2. The literature

   A binding into a ledger, beside the sentence of the stored source it was quoted from.

3. A convention of writing

   The common ones are built in, such as the 95 of a 95% confidence interval. Any other is justified once, in paper.yaml.

A number that is none of these is a finding of the check, not something left for a reviewer to notice.

The demo

See it catch a mistake

In under a minute: a number typed over its binding in the worked example that comes with the tool, and what check says of it; then some of the mistakes the gates catch, and what is left to a co-author. The video has no sound; its words are on screen, and below it.

The words of the video

1. A paper is full of numbers.
2. Each one should come from somewhere.
3. Type one in by hand, and it drifts from the analysis.
4. manuscript-guard finds it, and says where.
5. Bind it to the result, and it follows the analysis.
6. A number typed into the text, bound to nothing — G2
7. A results file edited by hand — G1
8. A retracted work, still cited — G7
9. A DOI that names another work — G7
10. The plan changed after the first results, and nothing said — G12 warns
11. The analysis changed since the Methods were read — G9
12. Some mistakes no check can see.
13. A co-author reads those. manuscript-guard asks them, one page each.
14. Check a paper in your browser, or start one.

Who decides

The gates decide, an agent helps, a person judges

• The gates

  Seventeen gates, each code that decides from the project's records, offline and in CI. No gate asks a model anything; two of them read a recorded review, which a person or a model may have written.

• An agent, if you use one

  Claude Code, Codex or another agent tool, with a skill for each step, drafts the plan, writes and runs the analysis, binds the numbers, drafts the prose and fixes what the check finds. Everything runs without one.

• A person

  Approves the manuscript, signs off what is theirs to close, and judges whether a passage supports its sentence. You are asked before the manuscript is sent to a model provider.

What it catches

Vord and the mistakes the gates catch

Vord, the guard, stands for the gates. Each villain is a mistake that a gate catches, named with that gate.

• 

  The Stray Number

  A number typed into the text, bound to nothing.

  G2
• 

  The Forger

  A results file edited by hand.

  G1
• 

  The Ghost Citation

  A retracted work, still cited.

  G7
• 

  The Impostor DOI

  A DOI that names another work than the one cited.

  G7
• 

  The Plan Rewriter

  The plan changed after the first results, and the Deviations say nothing.

  G12
• 

  Stale Methods

  The analysis changed since the Methods were last read against it.

  G9

What each gate checks

What no gate catches

A mistake no gate sees is a shadow. Vord's sword passes through it, so Vord calls a co-author, who reads what no gate can. Each co-author gets a page of things to confirm: whether a sentence says what its source says.

• 

  Confirmation bias

  It lies in what the authors chose to look at.

• 

  Data leakage

  No gate reads the analysis code for test data used in training.

• 

  The co-author

  The person who reads what no gate can.

Status

What it rests on so far

It is alpha.

• A real paper. One manuscript, a narrative review by the toolkit's author, has been through check, build and verify, which ran its analysis again and reproduced each of its 1,483 values.
• A measured set. The reference checks, on the 740 references of 16 published open-access pharmacology and epidemiology papers: as published, 1.49 false alarms per 100 references.
• Not measured. How much of a study an agent can carry with it, and whether the research is better for it. No study has yet gone from its plan to publication with the toolkit.

The full status

Install

Two ways in

uvx manuscript-guard audit paper.docx

pip install manuscript-guard
manuscript-guard init my-paper
cd my-paper
manuscript-guard next