authorecon

Four checks that run before research output leaves your hands. No dependencies, each one exits non-zero when it finds something, so they gate a build or a send.

pip install authorecon

submission-scrub

Refuses to let a file reach a human with agent artifacts still in it.

submission-scrub paper.tex
submission-scrub --dir submission/

It looks where prose review does not: LaTeX % comments, HTML and markdown comments, code comments, and instruction-shaped sentences addressed to the author rather than the reader. Reading a file over is not a check.

This exists because a .tex formatted with AI assistance was submitted to a journal with the assistant's own instructions still in the source. The file read fine. Nobody scrolled the comments. The editor issued a one-year ban.

slop-scan

Finds the structural habits that make prose read as machine-written.

slop-scan paper.md
slop-scan --strict draft.md      # fail on any hit, not only blocking ones

The usual advice is to grep for a word list — delve, tapestry, leverage. That catches the vocabulary and misses the writing. What actually reads as generated is structural: the contrastive pair that asserts a thing by denying its opposite, the announced candour that tells the reader a sentence is honest instead of being honest, the triad padded from two real items to three, the paragraph that ends by restating itself, em dashes three to a paragraph.

None of those contain a flagged word. All of them survive a word-list scan. So this looks for shapes. Hits are a reading list rather than an edit — a genuine contrast is a genuine contrast — and the blocking classes are the ones that have never once been right.

software-inventory

Reports which of your repositories are in a state anyone could actually cite.

software-inventory ~/code/project-a ~/code/project-b --json out.json
software-inventory . --markdown

Per tree: remote, branch, licence, citation metadata, entry points, and every module with its first docstring line. Then the gaps as a list. A tree with no licence cannot be reused by someone who read the paper it belongs to; a module with no docstring cannot be understood without reading it.

It reads the working copy rather than the remote, which is how it caught a repository whose licence existed on GitHub while the code did not exist there at all.

zenodo-precheck

Refuses to deposit something already deposited.

zenodo-precheck "A Title You Are About To Deposit"
zenodo-precheck --all-drafts papers/
zenodo-precheck --software ./my-repo --related 10.5281/zenodo.1234567

Set ZENODO_TOKEN to include your own unpublished drafts, which no public search can see.

The title mode catches a record you forgot you made. The --software mode answers a different question: it lists the files attached to every related record, looks inside any archive among them, and compares that against what your repository actually tracks. A title check answers has this name been used. It cannot answer has this content been deposited, and for software the second question is the one that matters — source is routinely published as a zip attached to a paper, under a name that resembles nothing.

Both failures these guard against happened, a few hours apart, to the author.

Licence

MIT.