NSDI Submission

The audit below is ordered by irreversibility: calendar errors first, then format, then blindness, then policy. Numbers are the NSDI '27 cycle as rendered on 2026-07-08; the fall HotCRP /deadlines page outranks everything here.

Gate 0: the right site, the right clock

NSDI runs a separate HotCRP instance per deadline (nsdi27spring.usenix.hotcrp.com, nsdi27fall...). Uploading to a stale spring site bookmarked in March is a real failure mode. For the '27 fall gate:

• Abstracts (title, author list, track, conflicts): September 10, 2026.
• Full papers: September 17, 2026.
• Both 11:59 pm US EDT — an Eastern-US midnight, not AoE and not the Pacific cutoff NSDI '26 used. Same-named deadlines moved time zones between editions; compute each coauthor's local equivalent from the HotCRP page, not from memory.

The abstract gate is a prerequisite: no registered abstract, no paper upload a week later. Register even if 60% sure — withdrawing is free, materializing an unregistered paper is impossible.

Gate 1: format

Page-cap pressure at NSDI is figure pressure — figures count. The fix is content triage, never \vspace surgery or font tricks; production chairs and reviewers both notice geometry games.

Gate 2: blindness, per track

Baseline sweep for all tracks — no author names, affiliations, acknowledgments, or funding numbers; self-citations in third person; references never omitted for anonymity; concurrent NSDI submissions cited with the CFP's fixed sentence ("Under submission. Details omitted for double-blind reviewing.").

Then the fork:

• Research/frontiers: deployment context must not identify the institution by implication. A trace "from our global CDN" plus a author-count and a topology detail can triangulate to one company; neutralize to scale-descriptive terms.
• Operational systems: the CFP relaxes object anonymity — real system names, company names, and links may stay because judging real-world use requires them — while author names stay hidden. Do not import this exception into a research-track paper.

# Mechanical pass on the exact upload candidate
pdftotext final.pdf - | grep -nEi '(acknowledg|funded by|grant|@[a-z]+\.(com|edu))' | head
pdfinfo final.pdf | grep -Ei 'author|creator'          # PDF metadata identity
grep -rnEi '(github\.com/[a-z0-9-]+|/home/|/Users/)' *.tex figures/ | head
# research track only: search for your institution's giveaway nouns
pdftotext final.pdf - | grep -niE 'ourcorp|internal-codename' | head

Mechanical passes catch strings; only a human pass catches implication. Have a non-author read the PDF asking one question: who wrote this?

Gate 3: policy

• Dual submission / prior publication / plagiarism: prohibited by USENIX; concurrent review anywhere else is disqualifying, and the ban survives notification-date arithmetic ("they'd decide before NSDI" does not help).
• Resubmission ban: if this work was rejected (without a revision option) from the immediately preceding NSDI deadline, it cannot enter this one. Check the paper's own decision history — spring '27 reject means no fall '27 entry (nsdi-review-process).
• One-shot revisions enter with their auxiliary packet complete (nsdi-author-response); an ordinary-looking resubmission that ignores its issue list is terminally rejected.
• Author cap: at most eight submissions per author across all '27 deadlines — reconcile the group's ledger, counting spring entries.
• AI use: grammar-and-clarity assistance on human-written text is endorsed; anything more, ask the co-chairs before the deadline, not after.
• Ethics: measurement of real users or third-party infrastructure needs its handling addressed in the paper; the '27 CFP's exact ethics wording is 待核实 — read that section live.

HotCRP form fields deserve the same audit as the PDF

Reviewers meet the form before the paper: title and abstract must match the PDF verbatim (drift reads as sloppiness and can confuse bidding); topic selections route the paper toward the sub-community you want reviewing it — choose for the reviewer profile, not maximal coverage; the full author list is entered at abstract time for conflict computation even though the PDF hides it, and late author additions after review assignment are a co-chairs-level incident. Declare conflicts by the site's definitions, completely and no wider.

Final sequence

1. T-8 days: fall HotCRP account, abstract fields drafted, conflicts entered.
2. T-7 (Sept 10): abstract registered; track locked deliberately.
3. T-2: content freeze; Gates 1-2 executed on the near-final PDF.
4. T-1: coauthor sign-off; HotCRP title/abstract verbatim-match the PDF.
5. T-0 (Sept 17): upload, status complete, then download the receipt copy and read it — the uploaded bytes, not the local file, are the submission.

Output format

[Clock] site URL + local-time cutoffs confirmed from /deadlines
[Gate 1] pages / template / track — pass or blocker
[Gate 2] mechanical findings + human-pass verdict (track-appropriate)
[Gate 3] dual-sub / ban / cap / AI / ethics — each cleared or flagged
[Receipt] uploaded PDF re-read? fields match?
[Blockers] ordered, with owner and time cost