brycewang-stanford EuroSys Artifact Evaluation
Use this when an accepted (or optimistically, submitted) EuroSys paper has code,
data, or hardware behind its claims. EuroSys artifact evaluation is organized
with the sysartifacts community (sysartifacts.github.io/eurosys2026/,
rendered 2026-07-08); it is optional, happens after acceptance notification,
and awards up to three badges that appear on the published paper. The EuroSys
2027 call for artifacts was not yet verifiable on 2026-07-08 — 待核实 before
committing to dates or badge criteria.
The three-badge menu
Authors declare which badges they want; the AEC grants zero to three.
| Badge | EuroSys 2026 criterion (rendered) | Typical blocker |
|---|---|---|
| Available | Artifact retrievable from a public, DOI-backed archive | "It's on our GitHub" — mutable, not archival |
| Functional | Complete, documented, exercisable as described | Undocumented cluster assumptions; missing inputs |
| Reproduced | Paper's main claims independently re-obtained | Experiments needing hardware the AEC cannot access |
Sensible combinations tracked by the process: all three for ordinary software artifacts; Available + Functional for datasets or custom-environment systems; Functional + Reproduced when the artifact itself cannot be published. Pick the target set before packaging — each badge changes what you build.
Design for the evaluator you will actually get
The AEC member is a systems student on hardware you did not choose, with a fixed time budget and no context beyond your paper. EuroSys evaluations have run a kick-the-tires phase with author interaction in past cycles (待核实 for the current one), but the artifact should survive with zero interaction:
- One top-level document mapping each paper claim to a command and an expected output range — claims first, directory tour second.
- Pin the environment: container image or lockfile, exact kernel/library expectations stated, hardware floor declared honestly.
- Scale knobs: a minutes-long subset run for functionality, a full run for reproduction, clearly separated.
- Where results are hardware-sensitive (NUMA layout, NIC model, device speeds), state the expected deviation instead of letting the evaluator guess.
artifact/
├── README.md # claim -> experiment -> command -> expected result
├── Dockerfile # or environment.lock + install script
├── run_functional.sh # ~15 min, exercises every component
├── run_full.sh # regenerates Figs 5-9, prints runtime estimate
├── data/ # inputs or fetch script with checksums
└── expected/ # reference outputs + tolerance notes
Archival step for the Available badge
Push the evaluated snapshot — not a later fix — to a DOI-minting archive (Zenodo is the norm in this community), record the DOI in the camera-ready availability section, and keep the GitHub repository as the living mirror. A tag on a mutable host does not satisfy the badge's permanence requirement.
Why aim above the minimum
EuroSys names an artifact prize — the Gilles Muller Best Artifact Award (EuroSys 2023 awards page; the award honors the late Gilles Muller, whose EuroSys '08 Coccinelle work won a Test-of-Time award). Committees remember artifacts that reproduce headline numbers unattended; that reputation follows the system into follow-up papers and adoption.
Timeline coupling with the paper lanes
- Start the artifact during review, not after notification: the gap between acceptance and artifact submission is weeks, and it overlaps camera-ready work that wants the same people.
- Spring-cohort papers (notified in late August for 2027) have a long runway to the April conference; fall-cohort papers (notified late January) run artifact evaluation and camera-ready nearly concurrently — 待核实 the edition's actual AE dates, and assign separate owners if in the fall cohort.
- Freeze the artifact snapshot when the camera-ready evidence freezes; divergence between the published numbers and the archived artifact is the most embarrassing discrepancy the process can surface.
Documentation pyramid
Write three layers, in this order of effort:
- The claims map (one page): paper claim → command → expected result → tolerated deviation. This is what gets read.
- The environment contract (half a page): image or lockfile, hardware floor, network assumptions, run times. This is what prevents failed setups.
- Everything else (as needed): design notes, extension hooks, dataset documentation. This is what gets skimmed, if opened at all.
An artifact with only layer 3 — however thorough — evaluates as undocumented, because the evaluator's clock runs out before the claims map they needed emerges from the prose.
Failure patterns the AEC reports
- Hard-coded paths and hostnames from the authors' cluster.
- The "works in our CI" artifact whose dataset download link needs a VPN.
- Reproduction scripts that emit numbers but never say which figure they rebuild.
- A 40-page wiki where a two-page claims-to-commands map was needed.
Output format
[Badge target] Available / +Functional / +Reproduced (with rationale)
[Evaluator path] <first command an AEC member runs, and its runtime>
[Claim coverage] <paper claim -> script -> expected output>
[Hardware honesty] <what the AEC cannot reproduce, and the declared fallback>
[Archival plan] <DOI host, snapshot timing relative to camera-ready>
[待核实] <current call-for-artifacts dates, badge wording, kick-the-tires>