prepare-artifacts

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Prepare Artifacts

准备可复现性工件

Turn a research codebase into a submittable, badge-ready reproducibility artifact. Artifact evaluation is a separate, post-acceptance track at most systems/PL/ML venues with its own deadline, its own appendix, and badges that change per venue per year — this skill builds the package (README, appendix, run instructions, anonymized repo, archival deposit guidance), produces an artifact-readiness checklist and a packaging plan, and lints the artifact directory for the bars reviewers actually check.

It does not run the author's experiments or claim a result reproduces — it prepares and checks the package, and tells the author exactly what reviewers will verify by hand.

将研究代码库转换为可提交、符合徽章要求的可复现性工件。工件评估是大多数系统/编程语言/机器学习会议中独立于论文评审的、在论文录用后的环节，有单独的截止日期、专属附录，且徽章要求每年随会议不同而变化——本工具可构建工件包（README、附录、运行说明、匿名仓库、归档存储指导），生成工件就绪检查清单和打包方案，并针对评审人员实际检查的标准对工件目录进行校验。

本工具不会运行作者的实验或声称结果可复现——它仅负责准备和检查工件包，并明确告知作者评审人员会人工验证的内容。

When to use

使用场景

"My paper was accepted — how do I do the artifact evaluation / get a badge?"
"Package / clean up my code for submission." / "anonymize my repo for review."
"What's an artifact appendix / Artifacts Available / Functional / Reusable?"
"Do I need a Zenodo DOI? concept vs version?" / "Software Heritage?"
"Fill out the NeurIPS code/reproducibility or ACL repro checklist."
"What does Reproduced vs Replicated mean for this badge?"
Alongside
```
prepare-camera-ready
```
(de-anonymization + final deposit overlap).

"我的论文录用了——如何进行工件评估/获取徽章？"
"打包/清理我的代码用于提交。" / "匿名化我的仓库用于评审。"
"什么是工件附录/Artifacts Available/Functional/Reusable？"
"我需要Zenodo DOI吗？概念DOI和版本DOI有什么区别？" / "什么是Software Heritage？"
"填写NeurIPS代码/可复现性清单或ACL可复现性清单。"
"对于该徽章，Reproduced和Replicated分别是什么意思？"
与
```
prepare-camera-ready
```
配合使用（去匿名化+最终归档环节重叠）。

Inputs

输入信息

The artifact directory — the code/data repo to be packaged (path).
The target venue + track, and ideally
```
venues/conferences/<v>-<year>.yml
```
(supplies the review blind level; create with
```
parse-cfp
```
if missing). The venue profile does NOT encode the artifact track's badge offering or its separate deadline — those are fetched live (step 1).
The paper's major claims (for a per-claim reproduction plan) and whether the artifact is for review-phase (often double-blind) or the final deposit. These change everything (anonymized ZIP vs version DOI).

工件目录——待打包的代码/数据仓库（路径）。
目标会议+环节，理想情况下提供
```
venues/conferences/<v>-<year>.yml
```
文件（提供评审盲审级别；若缺失可使用
```
parse-cfp
```
生成）。会议配置文件不包含工件环节的徽章类型或单独截止日期——这些信息会实时获取（步骤1）。
论文的核心结论（用于制定逐结论的复现方案），以及工件是用于评审阶段（通常为双盲）还是最终归档。这会影响所有流程（匿名ZIP包 vs 版本DOI）。

Process

流程步骤

Fetch the venue's CURRENT Call for Artifacts — mandatory, live. Badge offerings vary per venue per year (OSDI '26 evaluates ONLY "Artifacts Available"; SOSP '26 offers all three). Memory and last year are stale by construction; verify live. From the live CFA confirm: which badges are offered this cycle, the separate artifact deadline, the archival-hosting requirement, the appendix template/length, and the blind model. Snapshots to start from (re-verify, don't trust): references/venue-artifact-rails.md. Record the chosen badge target + artifact deadline in
```
.paper-memory/decisions.md
```
.
Resolve the badge taxonomy and the era trap. Use
```
python3 scripts/badge_advisor.py --badge <name>
```
to print the ACM v1.1 families/tiers and, critically, the Reproduced/Replicated swap: ACM inverted these terms on 2020-05-14, so a pre-2020 badge means the inverse (
```
--era pre-2020
```
). Reproduction is never bit-exact — it must agree within a tolerance that does not change the paper's claims. Background: references/badging-standards.md.
Lint the artifact directory against the bars reviewers check:
```
python3 scripts/check_artifact.py <artifact_dir> \
    --venue venues/conferences/<v>-<year>.yml [--blind double]
```
It reports, with file paths: the ML Code Completeness 5 items (dependency spec, training code, evaluation code, pre-trained models or a documented way to get them, a README with a results table + the exact reproduce command); archival readiness (GitHub-only vs a DOI/SWHID); double-blind anonymization (author names/emails, identifying URLs, a
```
.git
```
directory, PDF/appendix metadata) — driven by the venue's blind level or
```
--blind
```
; and hygiene (a LICENSE, upload-size cap). Flags:
```
--json
```
,
```
--strict
```
,
```
--zip-cap-mb N
```
,
```
--venues-dir
```
. Exit codes: 0 clean, 1 errors, 2 usage. The lint covers FILES only — it cannot prove the build runs, that results reproduce, or that a DOI resolves.
Build the package the venue asks for (with the author, not for them):
- README — overview, exact dependency install, the precise command to reproduce each result, a results table, hardware/runtime expectations, and the license. (ML Code Completeness item 5.)
- Artifact appendix — for USENIX-family Phase 2, a ≤3-page PDF (their LaTeX template): hardware/software/config, the paper's major claims, and a per-claim reproduction procedure + result-comparison method ("agrees if within X%"). For SIGMOD ARI, include experiment scripts AND graph-generation scripts ("similar behavior", not exact numbers).
- Checklists — fill the NeurIPS Paper Checklist / Code policy or the ACL "Responsible NLP Research" checklist accurately: an honest "no"/"n/a" with justification is safe; a missing or misleading filing is the desk-reject (ARR desk-rejects misleading filings since Dec 2024). Do not game boxes to "yes."
Anonymize for double-blind review (if review-phase). Ship an anonymized ZIP without
.git
, or proxy through Anonymous GitHub (anonymous.4open.science), listing every identifying term to scrub. Cover PDF/appendix metadata, acknowledgments, funding, and self-citation phrasing — same rules as the paper (
```
anonymize-paper
```
). Details: references/archival-hosting.md.
Plan the archival deposit. For "Artifacts Available," the permanent copy must be on an archival host — USENIX-family rejects GitHub/personal sites. Use a Zenodo version DOI for the final (a concept DOI is OK only during evaluation) and/or a Software Heritage SWHID (intrinsic, ISO/IEC 18670); they are complementary. Add CITATION.cff/codemeta so the archive emits citation metadata. De-anonymize and deposit the FINAL version at camera-ready (
```
prepare-camera-ready
```
).
Write the artifact-readiness checklist + packaging plan to
```
paper-workspace/submission/artifact-readiness.md
```
and append a line to
```
paper-workspace/INDEX.md
```
. Order by severity; cite each finding's source (the lint, the live CFA, the badge taxonomy). Re-run the lint until the file-level bars pass.

实时获取目标会议当前的工件征集要求——强制步骤。徽章类型每年随会议不同而变化（OSDI '26仅评估「Artifacts Available」；SOSP '26提供全部三种徽章）。仅凭记忆或去年的信息必然过时；必须实时验证。从实时征集要求中确认：本届会议提供哪些徽章、工件环节的单独截止日期、归档存储要求、附录模板/篇幅限制，以及盲审模式。可参考快照（需重新验证，不可直接信任）：references/venue-artifact-rails.md。将选定的目标徽章+工件截止日期记录到
```
.paper-memory/decisions.md
```
中。
解析徽章分类体系与术语陷阱。使用
```
python3 scripts/badge_advisor.py --badge <name>
```
命令打印ACM v1.1的徽章分类层级，关键是要注意Reproduced/Replicated术语的变更：ACM在2020-05-14反转了这两个术语的含义，因此2020年前的徽章含义与现在相反（使用
```
--era pre-2020
```
参数）。复现绝非完全精确的比特级匹配——只需在不改变论文结论的误差范围内达成一致。背景参考：references/badging-standards.md。
针对评审标准校验工件目录：
```
python3 scripts/check_artifact.py <artifact_dir> \
    --venue venues/conferences/<v>-<year>.yml [--blind double]
```
该脚本会带文件路径报告以下内容：ML代码完整性的5项要求（依赖说明、训练代码、评估代码、预训练模型或获取方式说明、包含结果表+精确复现命令的README）；归档就绪性（仅GitHub存储 vs DOI/SWHID存储）；双盲匿名化（作者姓名/邮箱、可识别URL、
```
.git
```
目录、PDF/附录元数据）——由会议盲审级别或
```
--blind
```
参数驱动；以及规范性检查（LICENSE文件、上传大小限制）。支持参数：
```
--json
```
、
```
--strict
```
、
```
--zip-cap-mb N
```
、
```
--venues-dir
```
。退出码：0表示无问题，1表示存在错误，2表示使用不当。校验仅针对文件层面——无法证明构建可运行、结果可复现或DOI可访问。
按照会议要求构建工件包（与作者协作，而非代劳）：
- README——概述、精确的依赖安装步骤、复现每个结果的具体命令、结果表、硬件/运行时预期，以及许可证信息。（ML代码完整性第5项要求）
- 工件附录——对于USENIX系列的第二阶段，需提交**≤3页的PDF**（使用其LaTeX模板）：硬件/软件/配置信息、论文的核心结论，以及逐结论的复现流程+结果对比方法（如“误差在X%以内即视为一致”）。对于SIGMOD ARI，需包含实验脚本和图表生成脚本（要求“行为相似”，而非精确数值）。
- 清单填写——准确填写NeurIPS论文清单/代码政策或ACL“负责任NLP研究”清单：诚实填写“否”/“不适用”并说明理由是安全的；遗漏或误导性填写会导致直接拒稿（自2024年12月起，ARR会因误导性填写直接拒稿）。不要为了勾选“是”而弄虚作假。
为双盲评审进行匿名化处理（若处于评审阶段）。提供不含
```
.git
```
的匿名ZIP包，或通过Anonymous GitHub（anonymous.4open.science）代理，列出所有需要清理的可识别内容。清理PDF/附录元数据、致谢、资助信息和自引用表述——遵循与论文匿名化相同的规则（
```
anonymize-paper
```
工具）。详细说明：references/archival-hosting.md。
规划归档存储方案。对于“Artifacts Available”徽章，永久副本必须存储在归档级平台——USENIX系列拒绝GitHub/个人站点。最终版本使用Zenodo版本DOI（评审阶段可使用概念DOI）和/或Software Heritage SWHID（符合ISO/IEC 18670标准）；两者可互补。添加CITATION.cff/codemeta文件，以便归档平台生成引用元数据。在终稿提交阶段（
```
prepare-camera-ready
```
）进行去匿名化并完成最终归档。
编写工件就绪检查清单+打包方案，保存到
```
paper-workspace/submission/artifact-readiness.md
```
，并在
```
paper-workspace/INDEX.md
```
中添加一行记录。按问题严重程度排序；注明每个发现的来源（校验结果、实时征集要求、徽章分类体系）。重新运行校验工具，直到文件层面的要求全部达标。

Output

输出内容

An artifact-readiness checklist (PASS / PASS-WITH-WARNINGS / FAIL with file paths) plus a packaging plan: target badges (from the live CFA), hosting (anonymized review copy + final version DOI/SWHID), the completeness gaps to close, the appendix/checklist to fill, and the separate artifact deadline. Written to
```
paper-workspace/submission/
```
.
Draft README / appendix / checklist content the author edits and owns.

工件就绪检查清单（通过/带警告通过/失败并附文件路径）和打包方案：目标徽章（来自实时征集要求）、存储方案（匿名评审副本+最终版本DOI/SWHID）、需填补的完整性缺口、需填写的附录/清单，以及工件环节的单独截止日期。保存到
```
paper-workspace/submission/
```
目录。
供作者编辑和确认的README/附录/清单草稿内容。

Adapt to your discipline

适配不同学科

The badge taxonomy here is ACM/USENIX/SIGMOD/ETAPS/ML-venue specific. For other fields, swap in your venue's artifact/data-availability rules (e.g. journal "data availability statements", FAIR data deposits) — the completeness and anonymization lints read the directory, not a discipline, so they still apply.

此处的徽章分类体系针对ACM/USENIX/SIGMOD/ETAPS/机器学习会议。对于其他领域，可替换为对应会议的工件/数据可用性规则（如期刊的“数据可用性声明”、FAIR数据归档要求）——完整性和匿名化校验仅针对目录内容，不局限于特定学科，因此仍适用。

Guardrails

注意事项

Re-verify the venue's CURRENT artifact rules live (step 1 is not optional). Badge offerings change per venue per year; never assume from memory or last year. Overconfidence is highest right after a fetch — re-check the primary CFA.
Never claim a result reproduces, and never demand bit-exact reproduction. ACM/SIGMOD/ETAPS require agreement within a tolerance that doesn't change the paper's claims. This skill prepares and checks the package; it does not run the experiments or judge the science.
The Reproduced/Replicated terms were swapped in 2020 — check the badge era (
```
badge_advisor.py --era
```
) before interpreting them.
Archival hosting is specific: a GitHub URL is not "Available" for the USENIX family — direct authors to a Zenodo version DOI / SWHID.
Anonymization-aware: for double-blind, scrub
```
.git
```
, names, emails, URLs, and metadata before any review-phase upload.
Accurate checklists, not gamed ones: honest "no"/"n/a" with justification is safe; misleading filings get desk-rejected.
Copilot, not pilot: never deposit, never submit to an artifact-evaluation system, never complete a checklist form on the author's behalf. Prepare, lint, explain — the author clicks.
Quote at most the flagged line/path; never bundle the author's artifact into this repo.

必须实时验证目标会议当前的工件规则（步骤1不可省略）。徽章类型每年随会议不同而变化；绝不要仅凭记忆或去年的信息判断。实时获取信息后仍需再次确认——务必核对官方征集要求原文。
绝不声称结果可复现，也绝不要求比特级精确复现。ACM/SIGMOD/ETAPS要求结果在不改变论文结论的误差范围内一致。本工具仅负责准备和检查工件包；不运行实验或评判科研内容。
Reproduced/Replicated术语在2020年发生了含义反转——解读前需确认徽章的时代（使用
```
badge_advisor.py --era
```
参数）。
归档存储有明确要求：对于USENIX系列，GitHub链接不满足“Available”徽章要求——需引导作者使用Zenodo版本DOI/SWHID。
支持匿名化处理：双盲评审阶段，上传前需清理
```
.git
```
目录、姓名、邮箱、URL和元数据。
准确填写清单，拒绝弄虚作假：诚实填写“否”/“不适用”并说明理由是安全的；误导性填写会导致直接拒稿。
仅做辅助，不代劳：绝不自动归档、绝不向工件评估系统提交、绝不代作者填写清单表单。仅负责准备、校验、解释——最终操作由作者完成。
最多引用标记的行/路径；绝不将作者的工件打包到本仓库。

Memory

记忆机制

Uses the shared

.paper-memory/

convention (full spec:

paper-memory-convention.md

At start: read
```
lessons.md
```
(skip re-flagging fixed items; lead with any
```
recurring
```
packaging habits, e.g. "you tend to ship a
```
.git
```
directory") and
```
decisions.md
```
(the chosen venue/badge target + artifact deadline).
At end: append the target badge + artifact deadline to
```
decisions.md
```
, and one dated entry per finding worth remembering to
```
lessons.md
```
in the shared
```
- [YYYY-MM-DD] (prepare-artifacts | <scope>) issue -> recommendation
```
format (use
```
reflect-and-improve
```
's
```
reflect_log.py append
```
, which dedupes/dates).
Create
```
.paper-memory/
```
on demand and offer to add it to the project
```
.gitignore
```
. Local-only; never upload it or copy it into this repo.

遵循共享的

.paper-memory/

约定（完整规范：

paper-memory-convention.md

）。

开始时：读取
```
lessons.md
```
（跳过已修复的问题；优先提示重复出现的打包习惯，如“你通常会包含
```
.git
```
目录”）和
```
decisions.md
```
（选定的会议/目标徽章+工件截止日期）。
结束时：将目标徽章+工件截止日期追加到
```
decisions.md
```
，并将每个值得记录的发现按
```
- [YYYY-MM-DD] (prepare-artifacts | <范围>) 问题 -> 建议
```
格式追加到
```
lessons.md
```
（可使用
```
reflect-and-improve
```
的
```
reflect_log.py append
```
命令，该命令会自动去重并添加日期）。
按需创建
```
.paper-memory/
```
目录，并建议将其添加到项目的
```
.gitignore
```
中。该目录仅本地存储；绝不上传或复制到本仓库。