thedixitjain CHI Artifact Evaluation
CHI's Papers track posted no separate artifact-evaluation committee or badge track for the cycles checked (2026, 2027; confirm each year — 待核实 as a standing item). That absence cuts both ways: nobody certifies your artifacts, and nobody but you ensures a skeptical reviewer can inspect them. At CHI the "artifact" is rarely just code — it is the prototype, the study instruments, the codebook, the dataset, and the video evidence of the system working. Package for two audiences with opposite needs: the anonymous reviewer with five minutes, and the future researcher with a real use for your materials.
The CHI artifact inventory
| Artifact | Reviewer's question | Archival value after acceptance |
|---|---|---|
| Prototype / system code | "Does the claimed interaction actually exist?" | Others extend or compare against it |
| Video figure of real use | "Does it work outside the authors' hands?" | Permanent DL evidence (chi-supplementary) |
| Study instruments (guides, questionnaires, tasks) | "Was the study what the paper says?" | Direct reuse in replications |
| Codebook / analysis audit trail | "Do the themes live in the data?" | Methods teaching material |
| Dataset (de-identified) | "Do the numbers re-derive?" | Secondary analysis |
| Design files (3D prints, schematics, figma) | "Could this be rebuilt?" | Fabrication replication |
| Prompts / model configs for AI conditions | "What system did participants face?" | The only record once the API moves |
Inventory first, then decide per artifact: reviewed now, released later, or honestly
withheld with a reason (chi-reproducibility has the data-sharing ladder).
Reviewer-facing packaging (anonymous, five minutes)
The review-phase archive rides the single September deadline with everything else. Design it so the first five minutes land:
- Root README with a claims map: "Claim in §5.1 →
analysis/h1_test.R→ Table 2." Three such lines are worth thirty pages of appendix. - One command per claim where code is involved; vendor the environment
(
requirements.txt,renv.lock) because a reviewer will not debug pip. - A demo path that does not require your hardware. If the contribution is a physical device, the video figure is the reviewable artifact; the archive adds schematics and firmware so the claim is auditable in principle.
- Anonymity end to end: no
.git, no metadata, no named accounts, anonymized platform views only, usernames scrubbed from notebook outputs and file paths.
# Cold-simulate the reviewer on the exact ZIP you will upload
rm -rf /tmp/ae && unzip -q supplement.zip -d /tmp/ae && cd /tmp/ae
cat README* | head -30 # does the claims map appear immediately?
grep -rEil 'author|university|(^|[^a-z])lab' --include='*.md' . | head
time bash run_minimal_demo.sh # the five-minute budget is literal
Post-acceptance release is a separate product
At the publication-ready stage (February for CHI 2027), rebuild the artifact set under real names for permanence, not for review:
- Deposit in a persistent home — the DL supplemental record, OSF, Zenodo, an institutional archive — with a DOI; a lab URL is a dead link in five years.
- License deliberately: code (MIT/Apache/GPL), data and instruments (CC BY or CC BY-NC), and record third-party constraints (stimuli copyrights, model terms of service for cached AI outputs).
- Re-check de-identification harder than at review: public release is forever, and participant re-identification harms real people. Where consent was narrow, release the instruments and codebook instead of the data — respected practice.
- Tag the released version to match the camera-ready ("as-published"), then develop onward in a separate branch; future readers need the paper's version, not HEAD.
Honest holes beat cosmetic completeness
A packaging note that says "the deployment used partner infrastructure we cannot ship; this archive contains the full client, the API contract, and a mock server reproducing the study conditions" earns more trust than a repo padded with dead code that hides the same gap. Reviewers at CHI read many partial artifacts; what they punish is discovering the gap themselves after the README implied completeness.
Timing inside the CHI year
- Build the claims-map README while writing §4–5 of the paper — it doubles as your
own claim-evidence audit (
chi-experiments). - Freeze the review archive at T−1 week; sweep it with
chi-submission's checks. - Diary the February release rebuild at acceptance time; rushed public releases in the TAPS window are where consent violations happen.
Output format
[Inventory] <artifact: reviewed / released-later / withheld+reason, per row>
[Claims map] present in README: yes/no · claims covered: <n>/<n>
[Five-minute test] cold demo ran in <time> / failed at <step>
[Anonymity] archive clean: yes/no — <channels checked>
[Release plan] home: <DL/OSF/Zenodo> · license: <code/data> · consent re-check owner: <name>
Source: brycewang-stanford/Awesome-Journal-Skills → CHI-Skills/skills/chi-artifact-evaluation/SKILL.md