Anonymize paper

Anonymize Paper

Make a LaTeX submission genuinely double-blind — then undo it cleanly after acceptance. Anonymization leaks are documented desk-reject grounds (CHI and NeurIPS state this explicitly, including leaks in supplementary material and linked repos), and de-anonymization done by hand routinely leaves placeholder authors or dead anonymous.4open.science links in the published PDF. This skill runs the deep sweep in both directions and records every change so the reversal is mechanical, not archaeological.

When to use

• "Anonymize my paper for NeurIPS / CHI / ICML / KDD ..." / "blind this"
• "Did I leak my identity anywhere?" / "double-blind check" before submission
• "Rewrite my self-citations in third person"
• "My code/data links identify me — what do I do?"
• After acceptance: "de-anonymize", "restore the authors", "prepare the camera-ready author block"
• Called from tailor-to-venue (anonymization sweep step) or before preflight-check (final gate).

Inputs

1. The main .tex file (with \documentclass); \input/\include files are followed automatically.
2. The venue profile venues/conferences/<venue>-<year>.yml (schema in venues/schema.yml) — supplies the blind level (single/double/triple) and the cfp_url. No profile? Ask the user for the blind level or create a profile with parse-cfp.
3. Optional but recommended: the compiled PDF (metadata check), the supplementary directory, and the author/institution names to grep for.

Process — anonymize (submission)

1. Resolve the blind level, then re-verify it live — mandatory. Read the venue profile; fetch the cfp_url and confirm the blind level and the venue's anonymization policy wording (what counts as a violation, whether acknowledgments must be removed, whether anonymized artifact links are allowed). Single-blind venues (e.g. SIGSPATIAL) need no anonymization — tell the user and stop instead of mangling a fine paper.

2. Run the deep scan:

   python3 scripts/scan_anonymization.py paper.tex \
       --venue venues/conferences/<venue>-<year>.yml \
       --supplementary <supp-dir> --names "Jane Doe,Example University"
   

   Flags: --blind double (no profile), --pdf paper.pdf (explicit PDF), --no-pdf, --json, --strict (warnings also fail), --force (scan at single-blind venues anyway), --no-inputs. Exit codes: 0 clean, 1 leaks found, 2 bad arguments. The scanner covers: author/affiliation/email/ ORCID/\thanks blocks, acknowledgments, funding/grant ids, identifying links, bare emails, first-person self-citations, institutional self-references, pdfauthor, LaTeX comments, .bib files, home-directory paths, compiled-PDF metadata bytes, and the supplementary tree.

3. Fix findings with reversible edits. Work through ERRORs first, then WARNs (each is a judgment call — discuss, don't bulk-delete). For the fix recipe per leak class — author block, acknowledgments, self-citations, repo/dataset links, metadata, supplementary — follow references/leak-catalog.md. Two rules:

   • Prefer the \ifanon toggle so the camera-ready flip is one line; fall back to a submission-anon git branch when source will be uploaded (arXiv, supplementary zips) — toggles leak the real names in source.
   • Rewrite self-citations in third person ("Doe et al. [12] showed"), never as "Anonymous [12]" — unless the cited work is itself unpublished.

4. Record the reversal manifest. For every change write one line into anonymization-manifest.md (kept OUT of the submission zip): what was removed/rewritten, file:line, and the exact original text (grant numbers, acknowledgment paragraph, real repo URL). Format in references/camera-ready-reversal.md.

5. Sweep the supplementary material. Re-run step 2 with --supplementary after fixes. Ship code as a clean export (no .git), clear notebook outputs/metadata, remove LICENSE/AUTHORS copyright names, and host artifacts on an anonymized mirror (anonymous.4open.science) — never Drive/ Dropbox/GitHub links that expose the account.

6. Verify the compiled PDF. Recompile, re-run the scan so the PDF metadata check runs on the fresh PDF, and do the manual pass the scanner cannot: figures with lab logos or terminal screenshots showing usernames, dataset descriptions that name the institution, watermarks.

7. Re-run until clean, then gate. Iterate scan → fix → scan to exit 0. Then run the full preflight-check (it validates the documentclass invocation and the rest of the desk-reject surface).

Process — de-anonymize (camera-ready)

8. After acceptance, reverse using the manifest plus:

   python3 scripts/scan_anonymization.py paper.tex --mode camera-ready
   

   This flags the leftovers: placeholder/empty author blocks, lingering anonymous.4open.science links, [review,anonymous] class options, neurips_<year> without [final], \anontrue toggles, "omitted for review" wording, and missing acknowledgments. Restore the author block so it matches the copyright form (ACM eRights / IEEE eCF) EXACTLY — walk references/camera-ready-reversal.md — then hand off to prepare-camera-ready for the venue rail.

Output

• A findings report (text or --json) with severity, check id, file:line.
• The edited .tex/supplementary files (with the user's approval, one leak class at a time) plus anonymization-manifest.md for the reversal.
• At camera-ready: a leftover report and the restored sources.

Relationship to preflight-check

preflight-check runs the same source-level anonymization checks (shared check ids, anonymization/*) as one gate among many; this skill is the deep variant — comments, .bib, PDF bytes, supplementary trees, name grep — plus the fix workflow and the camera-ready reversal. Quick gate → preflight; full sweep or de-anonymization → this skill.

Adapt to your discipline

The leak patterns are field-agnostic; the policies are not. Fork and swap the venue profiles for your field's journals (many use single-blind — the scan then auto-skips), and extend --names conventions for institutional review boards or clinical-trial registry ids that identify groups in your field.

Guardrails

• Never claim the paper "is anonymous" — say "no machine-detectable identity leaks remain"; writing style, self-datasets, and niche topics can still identify authors, and reviewers actively search.
• Never delete scholarly content to anonymize (e.g. dropping a self-citation entirely is misconduct-adjacent); rewrite in third person instead. Citation integrity questions route through verify-citations.
• Re-verify the blind level and anonymization policy against the live cfp_url before editing — a wrong blind level mangles a correct paper.
• Never submit to any system on the user's behalf; stop at the report/edits.
• Quote at most the flagged line in reports; never paste large paper portions into outputs.