REPRO · Reproducibility · 可复现性
↑ Matrix Hub ↑ magalia.wiki
pinned inputs · auditable pipeline锚定输入 · 可审计管线

Reproducing the Findings复现本站结论

Every analytical finding regenerates from a build script over a pinned raw snapshot. The dumps are not shipped — they are multi-gigabyte, and EDCS grants no redistribution license — but this page pins the exact filename, size and SHA-256 of every input, tells you where to acquire the same data, and lists the script chain from raw dump to each published dataset. The ML benchmark has its own harness documentation in eval/REPRODUCIBILITY.md. 本站每一项分析结论均由构建脚本自锚定的原始快照再生成。原始转储不随站发布 —— 体积以 GB 计，且 EDCS 不授予再分发许可 —— 但本页锚定每个输入的确切文件名、大小与 SHA-256，注明同一数据的获取途径，并列出自原始转储至每个已发布数据集的脚本链。ML 基准另有 eval/REPRODUCIBILITY.md 之评测装置文档。

1 · Raw snapshots (the inputs)一 · 原始快照（输入）

Sizes and hashes are computed from the files on the build machine — if your acquired copy hashes differently, you have a different snapshot and counts will differ.大小与哈希均由构建机上的文件计算 —— 若你获取的副本哈希不同，则快照不同，计数亦将不同。

2 · Regeneration chains (raw → dataset → finding)二 · 再生成链（原始 → 数据集 → 结论）

Each published dataset's checksum is in the chain — regenerate and compare.每个已发布数据集的校验和均在链中 —— 再生成后即可比对。

3 · Honest limits三 · 如实声明的边界

⚠ Reproducibility here is acquire-and-verify, not one-click: the EDCS snapshot cannot be redistributed, so you must harvest your own (and a later harvest WILL differ — that is why every magalia count carries its snapshot date). Build scripts live in the repository (scripts/build_*.py); they read the snapshots at the pinned filenames. Model-assisted outputs are additionally benchmarked with abstention-honesty metrics (eval/). ⚠ 此处的可复现性是获取后验证，而非一键复现：EDCS 快照不可再分发，须自行采集（且后来的采集必然有异 —— 故 magalia 的每个计数都注明快照日期）。构建脚本在仓库中（scripts/build_*.py），按锚定文件名读取快照。模型辅助的输出另以弃答诚实度指标评测（eval/）。