RecSys Submission

Run this audit before anything is uploaded to the RecSys submission system. RecSys is a single-community but multi-track venue: the same manuscript can be legal in one track and desk-rejectable in another, so the audit begins with "which track is this PDF for?" and only then checks format. Reopen the current edition's Call for Contributions — every number below is a 2026-cycle anchor (verified 2026-07-09), not a permanent rule.

Pick the track before the format

For 2026 RecSys reorganized its lineup: it dropped the Late-Breaking Results (LBR) poster track and added Research & Practice (R&P) Notes. Route the contribution first.

References are excluded from the page count; appendices count inside it. Use the ACM two-column template unmodified — no margin, font, or spacing surgery.

Anonymization for recommender papers specifically

Generic double-blind advice misses the leaks recommender papers actually carry:

• Proprietary logs and platform. "Interaction logs from our video platform in region X" can name the company. Keep the statistics, genericize the source.
• Deployed system names. If the ranker already runs in production under a public name, the name de-anonymizes; rename for review or argue from the current policy on preprints.
• The anonymous repository. RecSys expects a code/data link inside the paper rather than a supplement upload — mirror it to an anonymized host and strip the owner, commit authors, and cluster paths from history.
• Self-citations in third person; acknowledgements and funding removed; PDF metadata stripped.

Blocking risks

• Wrong track for the contribution (a deployed-system paper squeezed into a Main long paper, or a reproduction study submitted as a novel-method paper).
• Over budget once appendices are counted inside the limit.
• Identity leak via logs, platform, deployed system name, or repository owner.
• Untuned baselines or a random split on sequential data — not a format desk reject, but the fastest route to a low score (see recsys-experiments).
• Parallel submission of the same work to another venue.

Desk-reject and triage table

Final-week sequence for an offline-evaluation paper

1. Freeze the split: regenerate every ranking table from the versioned dataset and a temporal split so the PDF and the anonymous repository agree.
2. Re-run baseline tuning last under an equal budget; a stale, under-tuned baseline is the classic reviewer catch when the numbers move.
3. Compress into the content-page budget by demoting per-dataset analyses to the appendix or repository, never by shrinking floats.
4. Anonymity sweep: PDF metadata, acknowledgements, log provenance, deployed-system name, repository owner.
5. Confirm the submission-form abstract matches the PDF abstract and the track is correct.

# Pre-upload checks that catch the silent killers
pdfinfo paper.pdf | grep -iE "author|creator"                 # metadata identity leak
grep -inE "acknowledg|our (platform|service|company)" *.tex   # textual / platform leaks
pdftotext paper.pdf - | grep -icE "tuned|grid search|equal budget"  # baseline tuning stated?
pdftotext paper.pdf - | grep -icE "temporal split|leave-one-last"   # leakage-aware split?

Format anchors

• RecSys uses the ACM two-column proceedings template; dense ranking tables and multi-panel offline/online figures overflow columns fast, so compress early rather than on deadline night.
• The page counts above describe the 2026 cycle; treat every number as provisional and recheck the current Call for Contributions before relying on it.

Output format

[RecSys readiness] Ready / Needs fixes / Wrong track / Not ready
[Track] main-long / main-short / past-present-future / reproducibility / industry / resource / R&P-note
[Budget check] <pages used>/<budget>, appendix-inside-budget confirmed yes-no
[Anonymity risks] <logs / platform / deployed name / repository owner / metadata>
[Evidence risk] <baseline tuning / split leakage / offline-only claim>
[Fix order] <ordered blocking fixes>