tmlr-verify

Pre-submission verification for TMLR (Transactions on Machine Learning Research) manuscripts — packaged both as a standalone shell validator and as a Claude Code skill.

Born from a real desk rejection: a \setcitestyle{numbers,square,comma} override plus a hand-typed thebibliography violated TMLR's "submissions that do not use the unmodified TMLR stylefile" ground, and the paper was rejected roughly twelve hours after submission — no content review, boilerplate letter, empty forum. The entire class of defect is mechanically detectable in seconds. This repo detects it before TMLR does.

What it checks

Ground 2 — unmodified stylefile / citation compliance

• No \setcitestyle, \bibpunct, or direct natbib load in the document (tmlr.sty mandates author-year; any override is a desk-reject)
• No manual thebibliography / hand-inserted References heading — BibTeX via tmlr.bst only
• tmlr.sty present and carrying the official authoryear line (detects modified copies)
• Compiled PDF contains zero [n]-style numeric citations and exactly one References heading

Ground 1 — anonymity

• Plain \usepackage{tmlr} — fails on [accepted]/[preprint], which reveal the author block
• Blank PDF metadata (Author/Title/Subject/Keywords)
• No identity strings (you supply the regex), email addresses, code-hosting URLs, or acknowledgments in the PDF text
• Warns on traceable job identifiers (SLURM numbers, compute-service job IDs)

Compile hygiene

• No undefined references/citations, no natbib numeric-fallback error, no dangling ??, overfull-box count

Grounds 3–5 (overlap, author quota, perceived fit) are not file-checkable; the skill prints them as a human checklist instead of pretending otherwise.

Standalone use

./check_tmlr_submission.sh main.tex 'surname1|surname2|your-university|your-cluster'

Run from the project directory after a full pdflatex → bibtex → pdflatex → pdflatex cycle. Exit 0 = all hard checks pass. Requires poppler (brew install poppler / apt install poppler-utils).

As a Claude Code skill

git clone https://github.com/mmurshed/tmlr-verify ~/.claude/skills/tmlr-verify

Then /tmlr-verify inside any TMLR project. The skill adds what grep cannot judge: a fresh clean-cycle compile (a stale .aux from an old numeric build makes natbib silently fall back to numeric citations — reproducing the desk rejection), identity-regex derivation from your own \author{} block, alphabetical-order and citation-rendering spot checks on the PDF, a two-minute screener-speed read of the abstract and section titles for skim-surface risks, and the short-vs-long submission-type page count.

One hard rule the tools can only remind you of

Upload the PDF and supplementary zip only — never the .tex. Under anonymous tmlr.sty your real author block compiles to "Anonymous authors", so the source file de-anonymizes a submission whose PDF is perfectly clean.

License

MIT