Reference Check / 参考文献核验

Verify that every reference in a manuscript actually exists — and that the claimed title, year, and venue match the official record. / 核验论文参考文献是否真实存在，标题、年份、 venue 是否与官方登记一致。

1. Workflow / 工作流

EN

1. Locate the reference section in the extracted full text (first numbered entry [1] ... or author-year entries; end boundary = next section heading).
2. Split entries with re.split(r'\n\[(\d+)\]\s', ...). For every NODOI entry, re-check the original text for a line-wrapped DOI (10.1145/\n3613904... is common).
   • Line-wrapped DOIs — always look at the NEXT line: grep shows only the matched line, so the continuation is invisible. Real-world miss: one entry ended with doi:10.1016/j.tics. and the continuation 2021.02.007 sat on the next line — checking only the matched line caused a false "incomplete DOI" verdict. Safer: join each entry's internal newlines into spaces (or append the next line) before regex-matching DOIs.
   • An "incomplete-looking" DOI is a wrap signal, not a citation error: a captured DOI ending in . or / (e.g. 10.1016/j.tics., 10.1145/) should be assumed line-wrapped — stitch the original text and verify before judging; only then FAIL.
3. Verify each entry:
   • DOI → https://api.crossref.org/works/{doi} (User-Agent with a mailto; sleep 0.3–0.5 s between calls); retrieve title / container-title / issued.
   • 10.48550/arXiv.* DOIs are DataCite-registered → route to the arXiv API directly; a Crossref 404 on a repaired DOI falls back to https://api.datacite.org/dois/{doi}.
   • arXiv → https://export.arxiv.org/api/query?id_list={id}; parse <entry><title>.
   • Line-wrapped DOIs are stitched by repair_doi (see step 2); network errors retry with backoff — if an entry still FAILs after retries, re-run the batch before judging: transient network outages typically produce clusters of consecutive FAILs.
   • Entries without a DOI (book chapters, monographs, web pages) are judged manually — never FAIL.
4. Title comparison (v2): compare the FULL entry (whitespace-collapsed) against the official title via best-window partial ratio — slide a window of len(official) over the entry, take the max SequenceMatcher.ratio(); exact normalized substring = 1.0. Bands: ≥ 0.85 OK / 0.60–0.85 REVIEW / < 0.60 MISMATCH; official-year absent from the entry forces at least REVIEW. Do NOT use plain ratio on entry[:80]: author names dilute it until genuine citations score 0.3–0.6, overlapping fabricated ones (measured).
5. Report: one line per entry — No | DOI | score | claimed title | official title | years | venue | OK/REVIEW/MISMATCH/FAIL/ARXIV/NODOI — plus an overall verdict; review all REVIEW/MISMATCH/FAIL lines manually.

中文

1. 在提取文本中定位参考文献段（首个编号条目起，下一章节标题止）。
2. 按编号正则切分条目；每个 NO-DOI 条目回原文二次检查跨行 DOI。
   • 跨行 DOI 必须看命中行的下一行：grep 只显示命中行，续行会漏（实例：某条目命中行 结尾是 doi:10.1016/j.tics.，续行 2021.02.007 在下一行；只看命中行会误判"残缺 DOI"）。更稳妥的方式：切分时把条目内部换行合并为空格（或拼接下一行）后再匹配 DOI。
   • "DOI 看似残缺"是续行信号而非引用错误：正则抓到以 . 或 / 结尾的 DOI （如 10.1016/j.tics.、10.1145/），先假定跨行折断，回原文拼接验证；确认拼接后仍无法解析才可判 FAIL。
3. 逐条核验：
   • DOI → Crossref API（https://api.crossref.org/works/{doi}，User-Agent 带 mailto，请求间隔 0.3–0.5 秒），取回 title / container-title / issued；
   • 10.48550/arXiv.* 是 DataCite 前缀 → 直接送 arXiv API；Crossref 404 且 DOI 已修复 → 回退 DataCite API；
   • arXiv ID → arXiv API，解析 <entry><title>；
   • 跨行 DOI 由 repair_doi 拼接（见第 2 步）；网络错误带退避重试，重试后仍成段 FAIL 的先重跑再下结论（本机网络偶发断连会连续多条目 FAIL）；
   • 无 DOI 条目（书章/专著/网页）人工判断，列出来，不判 FAIL。
4. 标题比对（v2）：整条条目（合并空白）vs 官方标题，用最优滑窗 partial ratio （归一化后精确子串 = 1.0）。分档：≥0.85 OK / 0.60–0.85 REVIEW / <0.60 MISMATCH； 官方年份在条目中找不到 → 强制至少 REVIEW。
5. 输出：每条一行——编号 | DOI | 分数 | 引文标题 | 官方标题 | 年份 | venue | 状态，末尾给 总体结论；所有 REVIEW / MISMATCH / FAIL 逐条人工复核；核验结论写入交付文档的复核/附录小节。

2. Known false positives / 可能的检查失误

• Official metadata title truncation — Crossref often stores short versions of ACM/IMWUT titles (real cases: Gaver 1999 → "Design: Cultural probes"; Geeng & Roesner → "Who's In Control?"; Sasha → "Sasha"). If the DOI resolves to the very paper cited, it is OK.
• Author names in the entry no longer distort v2 best-window scoring, but very short official titles (book chapters, "Implicit personalization"-style truncations) can still fall below the OK band — that is what REVIEW is for.
• Subtitle casing / hyphenation differences.

中文：官方元数据标题截断（DOI 解析正确即为 OK，v2 下通常落入 REVIEW 灰区，需人工确认）；极短官方 标题可能低于 OK 档；副标题大小写与连字符差异。

3. Genuine mismatch signals / 引用错误

• DOI resolves 404; official title unrelated to the claim; year/venue mismatch; the same DOI reused by multiple entries; a "citation" pointing to a non-existent software release.

中文：DOI 404；官方标题与引文无关；年份/venue 对不上；同一 DOI 被多条引文共用；指向不存在的软件发布页。

4. Related skills / 相关技能

• paper-citation-tracer — orthogonal: traces a quoted passage INTO the cited paper's PDF and highlights the original wording (answers "is it cited correctly?"; requires the cited PDF). This skill answers "does the entry exist?" (requires only network access). In a review, run this skill first, then switch to the tracer for any suspicious entry.
• chi-review — the full CHI review workflow; calls this skill in its reference-verification step.

中文：paper-citation-tracer 与本技能正交互补——tracer 溯源内容（引文段落 → 被引 PDF 高亮）， 本技能核验条目（参考文献列表 → 官方元数据）。审稿先跑本技能，可疑引用再切 tracer。 https://github.com/Imad7777/paper-citation-tracer