brycewang-stanford ISCA Supplementary Strategy
ISCA's verified 2026 format was spare: a printable PDF of at most 11 pages of text plus unlimited references, uploaded to HotCRP. The public guidelines described no separate supplementary-file channel and no appendix regime — the only sanctioned overflow surface they mention is a fully anonymized artifact link inside the paper. Strategy at this venue is therefore not "what goes in the supplement" but "what earns a place in the only document reviewers are given" — and what waits, staged, for the revision window and the camera-ready.
Triage: the four destinations for every piece of material
| Material | Destination | Why |
|---|---|---|
| Anything a claim depends on (mechanism detail, methodology, headline evidence, overheads) | The 11 pages | Reviewers judge the submission PDF; content elsewhere is invisible at decision time |
| Full config dumps, per-workload result tables, extra sweeps | Anonymized artifact repo, referenced once | Available to the curious without spending pages; not assumed read |
| Experiments anticipating likely objections | Staged, unpublished — ready for the rebuttal/revision window | The 2026 process let authors add evidence in February (isca-author-response) |
| Tutorials, extended lineage, code documentation | Post-acceptance artifact and camera-ready pointers | Serves readers and evaluators, not reviewers |
The discipline mirrors tape-out: the submission PDF is the mask set. Everything else is either supporting collateral or inventory for a later window.
What the anonymized link may carry — and what it may not
The 2026 guidelines explicitly required artifact links (e.g., repositories) to be fully anonymized. Practical protocol:
- Host on an anonymization-friendly mirror (anonymous review services or a scrubbed repo under a neutral account), never the lab's GitHub organization.
- Scrub the contents, not just the URL: commit author fields, README affiliations, hardcoded home-directory paths, cluster hostnames, license headers with names, CI badges pointing at the real org.
- Reference it once, matter-of-factly ("configurations and per-workload data:
# Identity scrub of the anonymous artifact mirror before submission
git -C mirror log --format='%an %ae' | sort -u # commit identities
grep -rniE 'university|institute|\.edu|corp|@[a-z-]+\.(com|org)' \
mirror/ --include='*.md' --include='*.txt' | head
grep -rn '/home/\|/Users/\|hostname\|slurm' mirror/ | head
zip -r artifact-mirror.zip mirror/ -x '*.git*' # if a file drop is used
Fitting evidence into 11 pages without hiding it
When the draft overflows, resist the instinct to exile results tables to the repo. Compression moves that keep evidence reviewable:
- Collapse per-workload bars into a compact table + one distribution figure;
keep the outliers visible and named (per
isca-experiments, outlier honesty is credibility). - Replace a second mechanism diagram with a walk-through that reuses the first.
- Cut background prose before cutting any measurement — architecture reviewers forgive terseness; they do not forgive invisible evidence.
- Summarize a sweep with its break points ("benefit holds for T ∈ [4K, 32K], collapses below") and put the full curve in the repo.
The staging shelf: supplementary material in time, not space
ISCA's process gives material a second entry point: the rebuttal/revision window (three weeks in the 2026 cycle). Maintain a staging shelf from day one:
staging/
obj-baseline-storage/ # equal-storage comparison vs <rival>, ready to run
obj-threshold-sweep/ # sensitivity sweep, plotted, unpublished
obj-os-effects/ # full-system spot-check, half-built
notes.md # for each: trigger objection, cost to finish, owner
Rules for the shelf: each item names the objection that would trigger it; each
is runnable inside a week (window math); none contains new-contribution
material — window additions must repair the submitted claim set, not extend it
(isca-author-response explains why extension backfires).
After acceptance: the release ladder
Camera-ready (isca-camera-ready) de-anonymizes the artifact link and the AE
process (isca-artifact-evaluation) formalizes the package. Sequence the
release: swap the anonymous mirror for the canonical repo; add an archival
snapshot (DOI-issuing service) for the version evaluated; move the shelf's
run-but-unpublished sweeps into the repo's extended-results directory, clearly
versioned against the paper.
Failure modes specific to this venue's spareness
- The phantom appendix: authors trained at appendix-friendly venues put the proof-of-overhead in "Appendix A" past page 11 — at ISCA that page either violates the format or does not exist; the overhead analysis must be in the 11 (or its space earned by cuts elsewhere).
- The revealing repo: an immaculate PDF and a mirror whose commit log names the lab. Sweep contents, not just URLs.
- The overloaded link: a reference like "see repository for evaluation details" attached to a headline claim — reads as evidence dodging and draws exactly the review it fears.
- The empty promise: "code will be released upon acceptance" with no link at all is weaker than either a clean anonymous mirror or silence; write only what is true and checkable.
Quick triage output
Record the decision for every displaced item so deadline week doesn't relitigate:
[Item] <table/figure/analysis/derivation>
[Claim it supports] <claim id, or "none">
[Destination] 11-pages / anonymized-repo / staging-shelf / post-acceptance
[If repo] identity-swept? yes/no [If shelf] trigger objection + cost-to-run
[Reviewer impact if invisible] fatal / weakens / none
Any item marked "fatal if invisible" that is not in the 11 pages is a layout bug — fix the budget, not the claim.
Verified format and blinding facts: ../../resources/official-source-map.md
(2026-07-08). Whether 2027 adds an appendix allowance, a supplement upload, or
different link rules is 待核实 — reread the new guidelines before triaging.