osdi-supplementary

OSDI Supplementary Material

OSDI inverts the supplement habit most authors bring from ML and ACM venues. Rules

below are the OSDI '26 cycle (verified 2026-07-08): submissions get **12 pages plus

unlimited references and nothing else** — the CFP permits extra pages for references

"but not for supplementary material in appendices." There is no supplement upload, no

appendix, no ZIP. Accepted papers later expand to **14 pages plus references and

appendices, plus an optional two-page Artifact Appendix** for papers passing

artifact evaluation.

The three destinations

Every piece of material has exactly one correct home:

| Destination | What belongs there | When it exists |

|---|---|---|

| 12-page reviewed body | Everything a reviewer needs to judge correctness and significance | Submission (December) |

| Final-paper appendix (post-acceptance) | Depth for readers: full protocol details, extended results, proofs written out, the Artifact Appendix | Final papers (June) |

| Public artifact | Code, traces, configurations, raw data, runbooks | Archived by the May artifact deadline |

The triage question for any candidate paragraph, proof, table, or figure: **"Does a

skeptical reviewer need this to accept the paper?"** Yes → it earns body space, and

something weaker must leave to make room. No → it waits for the final-paper appendix

or the artifact, and the body carries a one-line summary of what exists.

Living without an appendix at submission

Techniques for material that would be an appendix elsewhere:

• Proofs and invariants — state the invariant and the proof's one load-bearing

step in the body ("safety follows because a lease transfer requires quorum

acknowledgment of the epoch bump"); the full proof is final-paper appendix

material. If the paper is unacceptable without the full formal argument, OSDI's

format is telling you something about venue fit (osdi-topic-selection).

• Extended result grids — the body shows the RQ-answering figure; the sweep

across every configuration goes to the artifact as data plus plotting scripts,

referenced in one sentence without a link that breaks anonymity.

• Protocol/config detail — compress into a table; systems reviewers extract more

from a dense parameter table than from two pages of prose.

• Never point to an external URL as a de facto appendix. An anonymous link

carrying material the paper depends on violates the spirit of the rule reviewers

enforce, and any non-anonymous link is a double-blind breach (osdi-submission).

Planning the June expansion in December

Write the submission with a marked expansion plan so the accepted version is an

unfolding, not a rewrite:

% submission-time annotations, kept in comments in the source
% [FINAL-14pp] restore lease-protocol failure cases cut from §4.2
% [FINAL-APPENDIX-A] full safety proof (draft in proofs/safety.tex)
% [FINAL-APPENDIX-B] full workload sweep grid (artifact: plots/sweep/)
% [ARTIFACT-APPENDIX] runbook skeleton -> becomes the 2-page AE appendix
% rule: nothing in FINAL-* may be load-bearing for acceptance

The rule in the last comment is the audit: if any FINAL-* item is something a

reviewer would have needed, it must be promoted into the body now — after

acceptance is too late to have convinced anyone, and OSDI '26 offered no response

period in which to supply missing material on request.

Triage decision flow

Run each candidate item through this flow when the draft first crosses 12 pages —

not on deadline night:

item: proof / table / protocol detail / extra experiment / config dump
  |
  needed by a skeptical reviewer to ACCEPT? --yes--> body (cut something weaker;
  |                                                  osdi-writing-style compression
  no                                                 pass runs first)
  |
  needed by a READER or REPRODUCER later? --yes--> is it prose/math?
  |                                                  yes -> final-paper appendix
  no                                                 no (code/data/runs)
  |                                                       -> artifact
  cut entirely (deadline-week sunk cost is not a reason to keep it)

Two calibration notes: reviewers differ, so when genuinely unsure whether an item is

acceptance-critical, it goes in the body — the no-response process gives no chance

to supply it on request; and the body must summarize every deferred item in one

sentence ("the full sweep, 240 configurations, accompanies the artifact"), because

an unmentioned deferral is indistinguishable from a gap.

The Artifact Appendix (the one supplement OSDI formalizes)

For papers that pass artifact evaluation, the final paper may add up to two pages

documenting the artifact in a standard format: what the artifact contains, hardware

and software requirements, how to reproduce which claims. Treat it as the public,

citable interface to the artifact — written from the runbook the AE committee already

exercised (osdi-artifact-evaluation), not composed fresh in June.

Triage failure modes

• Submitting a paper that silently assumes appendix space exists — discovered at

page 13, two days before the deadline.

• Cutting load-bearing evidence to fit, instead of cutting redundancy

(osdi-writing-style's compression pass comes first).

• An external "anonymous supplement" link standing in for the forbidden appendix.
• Deferring so much to the final version that reviewers correctly judge the

submission incomplete — conditional acceptance mandates fixes, but only for papers

strong enough to conditionally accept.

Output format

[No-appendix compliance] submission = body + refs only? violations: <list>
[Triage table] each deferred item -> destination (final-appendix / artifact) + body summary line present?
[Load-bearing audit] any FINAL-* item a reviewer actually needs? <list or none>
[External links] none that carry evidence / none that break anonymity? yes/no
[June plan] expansion annotations in source? artifact-appendix runbook started?