Ijcai artifact evaluation

IJCAI Artifact Evaluation

Use this for artifact packaging around IJCAI. Treat the current reproducibility guidelines and supplementary-material rules as controlling; do not assume a separate formal artifact evaluation track unless the current cycle announces one.

Package design

• Decide what reviewers need to classify the results as convincing or credible: proofs, pseudocode, datasets, code, model cards, logs, environment details, or ablation notebooks.
• Keep essential evidence in the paper whenever space allows because reviewers are not required to read supplementary material.
• Put optional evidence in the Technical Appendix or ZIP, respecting the current size and format limit. IJCAI-ECAI 2026 allowed up to 50MB in PDF or ZIP form.
• Anonymize repository paths, user names, institutions, license headers, model checkpoints, data provenance, and notebook metadata.
• Include a minimal run map: environment, dependencies, hardware, commands, expected outputs, runtime, seeds, and known limitations.
• For proprietary or restricted data/code, explain why it cannot be shared and provide enough detail for in-principle reproduction.

Evidence by claim type

IJCAI usually has no separate artifact-evaluation badge, so the artifact's job is to move the reproducibility rating toward convincing and to pre-empt a broad PC's doubt. Match the package to the claim.

Claim	Artifact that convinces an IJCAI reviewer	Weak substitute to avoid
New search/planning algorithm	Runnable solver, instance generator, seeds, time/memory limits	A results CSV with no way to regenerate it
Theoretical guarantee	Full proofs, assumption list, citations to formal tools	"Proof omitted for space" with no appendix
Multi-agent protocol	Simulator, opponent policies, randomization seeds	Screenshots of one run
Learning result	Code, environment file, configs, compute and runtime	Final-number table only
Dataset	Datasheet, license, controlled-access path	Vague "available on request"

Worked vignette: packaging a SAT-solver paper

A SAT/CSP paper claims a new restart heuristic wins on hard industrial instances. Strong package: the solver binary or source, the exact instance set or a deterministic generator, solver and compiler versions, per-instance runtimes with the timeout stated, and a run map showing one command that reproduces the cactus plot. Anonymize the repository name, license headers, and any cluster paths. This lets a skeptical constraint-reasoning reviewer re-run a sample and raises the rating from credible to convincing without needing a badge track.

Reviewer pushback and the venue-specific fix

• "Cannot regenerate the benchmark instances." Ship a generator plus seeds, not just outputs.
• "Artifact leaks author identity." Scrub paths, license headers, checkpoint names, and notebook metadata before the full-paper deadline; there is no late re-upload.
• "Restricted data blocks reproduction." Document the legal barrier and give enough protocol for in-principle reproduction rather than implying a release you cannot deliver.

Output format

[Artifact role] paper evidence / supplement / post-acceptance release
[Contents] <proofs/code/data/models/logs/docs>
[Anonymity risks] <paths/licenses/metadata/URLs>
[Reproducibility claim] convincing / credible / weak
[Fixes before upload] <ordered list>