brycewang-stanford CIKM Supplementary
CIKM 2026 gives supporting material no free rider: every track's budget counts the appendix inside the page limit — 10 pages for full papers, 4/7/4/4 for short/applied/resource/demo — and only the references (full track: up to 2 pages) plus the GenAI Usage Disclosure section sit outside it (source map, 2026-07-08). There is no separate supplementary-upload channel in the verified 2026 calls. The design question is therefore three-way: costed appendix pages, the anonymously cited artifact, or cut.
The three containers
| Container | Costs | Right contents |
|---|---|---|
| Main body | Budget pages | Everything needed to evaluate the claim unaided |
| In-PDF appendix | Budget pages too | Only material reviewers must see during review that the body cannot hold: a proof, a prompt set, a schema diagram |
| Cited anonymous artifact | Zero pages, but discretionary attention | Configs, full hyperparameter grids, extra per-dataset tables, code, data cards |
Decision rule per candidate item: must a reviewer see this to score the paper? If yes and it is short, appendix — knowing it displaces body content one-for-one. If yes and it is long, restructure the body so a summary suffices, and put the whole in the artifact. If no, artifact or delete.
Building the anonymous artifact
- Host with an anonymity-preserving service (anonymized repository links of the usual kind); the URL, commit history, issue templates, CI badges, and README must not resolve to authors or institutions.
- Strip notebook metadata, license headers naming people, internal hostnames, and
data paths like
/home/username/— the classic leak channels. - Structure for a ten-minute reviewer visit: a top-level README mapping paper section → directory, and one script per headline table. Reviewers sample; make the sample they take be the strongest material.
- Reference the artifact from the PDF at the point of relevance ("full grids: artifact §3"), not only in a final availability statement, so the pointer is found when the question arises.
What the uncounted sections may and may not carry
The GenAI Usage Disclosure section and the references are outside the budget — and they are not overflow space. Substantive results, extra experiments, or method details placed there violate the sections' declared purposes and read as budget gaming to a panel that sees the trick every year. The disclosure section carries disclosure; references carry references; nothing else.
Appendix design rules (for the pages you do spend)
When appendix pages are budget pages, they must earn their cost like body pages:
- Order appendix sections by reference order from the body, and give each a one-line purpose header ("A. Prompt templates used in §4.3") — a reviewer arriving from a citation should land exactly where sent.
- No orphans: appendix content never referenced from the body is invisible at best and budget-waste at worst; either reference it or move it to the artifact.
- Keep appendix tables to the same caption standard as body tables (protocol, variance, source); reviewers do not lower standards past the references.
- Never let the appendix complete an incomplete body argument. The body must stand alone; the appendix substantiates, it does not continue.
Reviewer-attention economics
Plan the material split around how attention actually distributes:
| Material location | Realistic reviewer attention | Design consequence |
|---|---|---|
| Body pages 1-2 | Everyone, carefully | The whole argument, compressed |
| Body middle | Everyone, once | Protocol and headline evidence |
| In-PDF appendix | Sometimes, when following a specific doubt | Doubt-resolving material only |
| Artifact README | A minority, for ~10 minutes | Orientation + strongest single script |
| Artifact internals | Rarely, and only if the README earned trust | Completeness, not persuasion |
The split fails when persuasion-critical material sits below its audience's attention line — which is why "nothing decision-critical lives only in the artifact" recurs across this pack.
Worked split: a 13-page full-paper draft
A draft holds 9 pages of core, 2 pages of extra per-dataset tables, 1 page of prompts, 1 page of proofs. Target: 10 including appendix. Moves: per-dataset tables → artifact, keeping one summary row per dataset in the body (saves ~1.7 pages); prompts → appendix at half a page by showing one template plus deltas, full set → artifact; proof → appendix, compressed to the lemma statements with the full derivation in the artifact. Result: 10 pages with the review-critical material still in the PDF, and nothing decision-relevant living solely in the artifact — because artifact inspection is at reviewer discretion, never guaranteed.
Version-consistency discipline
Body, appendix, and artifact are three copies of one experimental truth, and they drift independently under deadline pressure. Locks that prevent it: generate paper tables from the same logged results the artifact ships (one source of numbers, two renderings); when a result updates, update body summary, appendix table, and artifact data in the same commit; and record in the artifact README which commit corresponds to the submitted PDF, so a reviewer comparing the two sees agreement, not archaeology. A number that differs between the PDF and the artifact is read as sloppiness at best — and at a venue that reserves automated compliance checking, consistency across submitted surfaces is the cheap kind of integrity to maintain.
Pre-upload verification
[ ] Page count with appendix ≤ track budget (measure, don't estimate)
[ ] Artifact URL anonymous end-to-end (open it logged out; check git log)
[ ] Every artifact pointer in the PDF resolves to real content
[ ] GenAI/reference sections carry only their declared content
[ ] Body survives a reviewer who never opens the artifact
Output format
[Budget state] <measured pages incl. appendix vs. track limit>
[Split map] <item → body / appendix / artifact / cut>
[Artifact health] anonymity / README / per-table scripts
[Gaming check] uncounted sections clean?
[Residual risk] <what a no-artifact reviewer would still miss>