eurosys-supplementary

EuroSys Supplementary

Use this when a EuroSys draft overflows its 12 pages of technical content. The

governing fact (EuroSys 2027 CFP, rendered 2026-07-08): the budget is 12 pages

of technical content plus unlimited reference pages. Whether the current

round accepts appendices beyond the references, or a separate supplementary

upload in HotCRP, is per-cycle policy that was not verifiable on 2026-07-08 —

待核实 in the live CFP and the HotCRP submission form before relying on either.

The four shelves

| Shelf | Review status | What belongs there |

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

| 12 technical pages | Fully reviewed, the paper | Design, key mechanisms, headline evaluation, honest limitations |

| Reference pages | Free, always allowed | Bibliography only — never technical content disguised as notes |

| Anonymized artifact repo | Reviewer-optional | Code, configs, full result sets, extra graphs, raw logs |

| Extended tech report / appendix | Only if the round permits (待核实) | Proofs, protocol details, exhaustive parameter sweeps |

The iron rule across all shelves: **a reviewer who reads only the 12 pages must

be able to accept the paper.** Anything reviewer-optional is corroboration, not

argument. If the claim's only support sits on shelf 3 or 4, the claim is

unsupported at decision time.

What systems reviewers push outward — and what they don't

Safe to externalize:

• Full configuration dumps and tuning grids behind a summarized sensitivity plot.
• Per-benchmark breakdowns when the body shows the aggregate plus the outliers.
• Formal-model details for a mechanism whose intuition the body already conveys.
• Additional workloads that repeat, rather than extend, the body's conclusion.

Dangerous to externalize at EuroSys:

• The baseline-fairness evidence (versions, tuning) — its absence from the body

reads as concealment, the single most corrosive suspicion in systems review.

• The one experiment that tests the design's stated weakness.
• Failure-handling behavior for a system whose pitch is reliability.

Anonymized repository mechanics

When linking supporting material from a double-blind submission:

# Build an identity-clean mirror rather than sanitizing in place
git clone --depth 1 file:///lab/systemX systemX-anon && cd systemX-anon
rm -rf .git && grep -rInE 'university|lab-name|@|githubusercontent' . | less
# fix hits, then publish to an anonymous host (e.g., an anonymized-repo service)

• Strip commit history entirely; author fields in old commits are the classic leak.
• Check config files and comments for hostnames, cluster names, and usernames.
• Keep the mirror frozen during review so reviewers see one consistent state.

Reference-page hygiene

The unlimited reference allowance is generous, not lawless:

• Bibliography pages hold citations — footnote essays, extended tables, or

"notes" sections smuggled after the references invite a compliance read

of the whole paper.

• Cite the version that supports the claim: a versioned release or paper,

not a bare project homepage that will drift.

• Under double-blind, reference entries for your own artifacts must not

resolve to identifying URLs; cite the anonymized mirror during review.

• Long URLs and dataset citations belong in the bibliography, where they

are free, rather than inline where they spend technical-page budget.

Micro-example: a 13.5-page draft

A distributed-tracing paper is 1.5 pages over. The cuts that worked, in

order of pain: the second motivating scenario compressed to two sentences

(0.4 pages); three per-workload latency plots collapsed into one CDF panel

with the full set moved to the artifact repository (0.5 pages); the

formal-consistency argument reduced to its statement plus intuition, full

treatment in the repo's design note (0.4 pages); prose narration of Figure 4

deleted because the caption already said it (0.2 pages). Nothing reviewers

needed for the accept/reject call left the body — that is the test each cut

must pass.

Overflow triage, in order

1. Compress prose before cutting content — systems drafts carry 10–15% slack in

restated motivations and figure narration.

1. Merge figures that make one point; delete figures that make none.
2. Move corroborating (not load-bearing) material to the artifact repo with a

one-line pointer in the body.

1. Only then consider an appendix or tech report, after confirming the round's

policy (待核实) and remembering reviewers owe it nothing.

Coordination with the review timeline

• Whatever optional material accompanies the submission must be final at

the paper gate; there is no mid-review update channel, and a stale repo

contradicting the PDF is worse than no repo.

• If a rebuttal phase exists, it cannot introduce a new supplement —

externalized material must already be in place to be pointed at.

• A one-shot revision, by contrast, is a full resubmission: the revised

round is the moment to promote material inward (e.g., a boundary

experiment the conditions demanded) rather than to grow the repo.

Output format

[Page pressure] <current overflow in pages>
[Shelf plan] <item -> body / references / artifact repo / appendix-if-allowed>
[Decision-critical check] <claims whose evidence would leave the body: none / list>
[Anonymity status of repo] <clean / leaks found>
[待核实] <does this round allow appendices or supplementary uploads?>