vis-reproducibility
GitHub用于强化IEEE VIS论文的可重复性与开放实践，涵盖材料声明、匿名可运行代码、研究预注册、数据与渲染管线溯源、图-结论映射及一致性检查，确保读者能复现结果。
触发场景
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill vis-reproducibility -g -y
SKILL.md
Frontmatter
{
"name": "vis-reproducibility",
"description": "Use when strengthening IEEE VIS reproducibility and open-practices evidence, covering the open-materials statement, anonymized-but-runnable code and stimuli, preregistration of perceptual and user studies, provenance for datasets and rendering pipelines, claim-to-figure mapping, honest degrees of reproducibility, and consistency between what the TVCG paper says and what the supplemental archive contains."
}
VIS Reproducibility
Use this before submission and again before camera-ready. IEEE VIS's Open Practices posture and the Graphics Replicability Stamp make reproducibility a visible dimension, not a courtesy: reviewers routinely open the supplemental code, data, and video, and the TVCG camera-ready collects open-practices disclosures. The goal is that a competent reader could rebuild your figures, rerun your study analysis, and reach your conclusions.
Evidence map
- Map each figure, quantitative result, and study finding to a verifiable location — a section, a figure generated from logged data, or a script in the supplemental archive.
- For techniques and rendering, give enough of the algorithm, parameters, and environment (including GPU/driver assumptions and tolerances) that a reader could re-implement or re-run.
- For empirical and perceptual studies, report participants and recruitment, apparatus/stimuli, the task, the design (within/between), measures, statistics, and the analysis scripts.
- Keep the open-materials statement truthful and specific: what is shared, where it lives, and — if something cannot be shared — exactly why.
- Keep the paper and the archive consistent: a number in the PDF that no script reproduces is the contradiction reviewers read as carelessness.
Open-materials statement audit
| Claim in the paper | Weak availability answer | VIS-ready answer |
|---|---|---|
| "We render/lay out X" | "Code available on request" | Public, licensed repo with a build path that regenerates the teaser figure |
| "Our system supports task Y" | "Demo will be released" | Runnable build (or Docker) with bundled sample data and a demo script |
| "N participants judged Z" | Nothing (privacy cited vaguely) | Anonymized responses, stimuli, the analysis notebook, and the ethics/consent note |
| "We evaluated on dataset D" | Named but not shared | The dataset or documented access + the preprocessing scripts |
"Available on request" reads as not available; convert every such line into a concrete, anonymized archive or an explicit, justified exception.
Preregistration for studies (a distinctly VIS-valued move)
Perceptual experiments and controlled user studies benefit from preregistration (e.g., on OSF): locking hypotheses, design, sample size, and the analysis plan before data collection separates confirmatory from exploratory findings and blunts the "you fished for that result" objection.
[Preregister] hypotheses, conditions, planned N + power analysis, primary DV, analysis plan
[Cite it] reference the (anonymized) preregistration in the paper; report deviations honestly
[Separate] label confirmatory vs. exploratory results; do not present post-hoc as planned
Provenance pinning
[Datasets] record source and version; archive the actual data or stimuli, not just a query/URL;
document any cleaning/filtering with the script
[Rendering] pin toolchain and library versions; provide reference images and a comparison
tolerance for GPU-dependent or non-deterministic output
[Studies] store raw per-participant responses (anonymized), the exact stimuli, and timing
[Compute] state hardware and runtime so a reader can size a reproduction
[Randomness] log seeds; say what is and is not deterministic
Degrees of reproducibility (state the one you achieved)
- Turnkey: one documented command regenerates each figure/result from logged data.
- Scripted: scripts exist but need documented manual steps, large data, or specific hardware.
- Descriptive: prose detailed enough that a competent reader could rebuild the pipeline.
For VIS, aim turnkey for anything an evaluator might rerun quickly (a figure from logged benchmark data, a study's statistics from anonymized responses); a large rendering benchmark or a proprietary dataset may stay scripted with access clearly documented. Stating the achieved level honestly beats promising turnkey behavior that fails on a clean machine — the GRSI stamp is decided exactly there.
Vignette: a technique-plus-study paper
A paper contributing a new encoding and a controlled study evaluating it. Its reproducibility spine: the encoding code with a script that regenerates each figure; the study's stimuli and anonymized per-participant responses; the preregistration for the confirmatory hypotheses; the analysis notebook that turns responses into the reported effect sizes and CIs; and one honest sentence about anything (identifiable video, proprietary data) that cannot be shared and why.
Consistency and camera-ready pass
- Before submission: every scored number and figure traces to the archive; the open-materials statement matches reality; if double-blind, the archive is anonymized (no owner strings, lab names, or institutional URLs).
- Before camera-ready: swap anonymized links for permanent, DOI-issuing archives, complete the Open
Practices form, and align the package with what you submit to GRSI (
vis-artifact-evaluation).
Output format
[Claim inventory] <figure/result/finding -> evidence location>
[Open materials] concrete / vague / missing
[Preregistration] present / not applicable / should have (for studies)
[Provenance gaps] <dataset versions / rendering references / study raw data / seeds>
[Reproducibility level] turnkey / scripted / descriptive, stated honestly
[Paper fixes] <must appear in the PDF>
[Archive fixes] <additions before upload>
版本历史
- 9f86f09 当前 2026-07-19 17:53
