Evidence level: C
池③ · 课题 #190(本轮新登记)· 承接第 1/2/6 轮引文审计
证据等级 C|状态 draft|本轮对既有正文写入 = 0
0. 一句话结论
跑了 721 篇、3162 个唯一 PMID,真正”不存在”的号只有 4 个——前几轮”假号很多”的印象是错的。真正的问题在另一头:PMID 真实,但配在它旁边的作者名、标题、期刊、年份全部是编的。我把它命名为「真号伪书目」(real-ID / fabricated-bibliography)。这一类,PubMed 存在性检查永远查不出。
1. 方法与工具(本轮新建)
新工具:08-ops/_tools/citation-pool-scan.py(纯标准库,py_compile 通过)
设计原则来自第 37 轮的教训——只报不需要语义判断就能证伪的错,宁可漏报,绝不臆断:
| 判据 | 是否确定性 | 是否进入结论 |
|---|---|---|
| E1 PMID 不存在 | 确定性(esummary + efetch 双确认) | ✅ |
| E2 PMC 号当 PMID | 确定性(elink 反查) | ✅(本轮 0 命中) |
| E3 年份矛盾 | 不确定(文本年份常是队列年/事件年) | ⛔ 降级为 INFO,不进结论 |
| 主题是否相关 | 不确定 | ⛔ 不判,留人工 |
| 作者是否对 | 不确定 | ⛔ 不判,留人工 |
1.1 工具自身踩的三个坑(全部实测修掉,写在这里防复发)
| # | 坑 | 后果 | 修法 | 验证 |
|---|---|---|---|---|
| 1 | PMID \*\*40892475\*\* 这种 markdown 加粗写法匹配不到 |
静默漏提,一个文件 5 个号被读成 0 个 | 标记与数字间允许 ≤6 个非数字字符 | 修正后同文件读到 5 个号 |
| 2 | pmid: ["36106418", ...] 带引号、单数、frontmatter 内 |
漏提 frontmatter 层 | 按 [, \s"'] 切分 + 加单值/短横线写法 |
未提取行 74 → 13 |
| 3 | PMID 正则未加 re.I,正文小写 pmid: 全漏 |
大规模静默漏提 | 全部加 re.I |
唯一 PMID 3153(此前 3152,差 1 但覆盖率从 74 缺到 13 缺) |
1.2 E3 为什么必须降级(实测数据)
首版用”邻域 ±1 行所有年份”判年份矛盾,报出 866 条。人工抽 12 条,≥11 条是假阳性:
| 声明年 | 真实年 | 实际含义 |
|---|---|---|
| 2015 | 2000 | 文本写的是「720BC–AD2015 的日食记录」——数据跨度,不是发表年 |
| 1960 | 2007 | 文本写的是「1960 年代电气化后」——事件年代 |
| 1996 | 2013 | 文本写的是「1940、1977、1996 三个出生队列」——队列年份 |
收紧到”同一行 PMID 前 45 字符内、剔除区间写法”后降到 125 条,抽样 12 条仍 ≥11 条假阳性(同一批问题换了个位置)。→ 结论:这类判据在本语料上不可用,降级 INFO。
这条教训值得单独记:池③正文大量使用「某年某研究」的叙事句式,年份与 PMID 的邻近关系不构成发表年声明。任何基于邻近性的年份校验,在这个语料上都会大面积误报。
2. 全池扫描结果
| 指标 | 数值 |
|---|---|
| 扫描文件 | 721 篇(03-sites/cn/research/science-yixue/) |
| 声明 PMID 唯一数 | 3162 |
| 覆盖率缺口(含 PMID 字样但未提取) | 13 行(均为审计文内的举例文本,非引用) |
| E1 不存在 | 15 处 / 4 个唯一号 / 6 个文件 |
| E2 PMC 当 PMID | 0 |
| E3 年份矛盾 | 0(已降级,见 §1.2) |
2.1 E1 全部命中(4 个唯一号,共 15 处)
| PMID | 出现文件 | 声明的用途 |
|---|---|---|
| 15199541 | standing-wave-antinode-psychophysiology.md(3 处) |
「HPA 2009《Infrasound: a brief overview of sources and effects》」 |
| 34479992 | dosha-qi-rsa-falsification-framework-methodology.md(2)、dosha-qi-rsa-falsification-preregistration-specification.md(2)、pool5-standby-validation-entry-exit-gate.md(1) |
「Chung 2021 中医证型 HRV」/「Putnick & Bornstein 2016 测量等值性系统综述」/「HRV 跨族裔参考值等值性」——同一个号在三处被赋予三种不同身份 |
| 28726973 | pulse-signature-detraining-regression-dynamics.md(2) |
参考文献列表内 |
| 16030677 | homeopathy-high-dilution-frequency-memory.md(1) |
疑为 Rey 热释光研究 |
(另有 rtribe-omop-citation-integrity-audit.md 4 处,属审计文自述”这些号不存在”,不是新错。)
34479992 这个号的三种身份最值得注意:同一个不存在的号,被三篇文章分别说成”Chung 2021 HRV””Putnick & Bornstein 2016 测量等值性””HRV 跨族裔等值性”。这与第 2 轮判定的「跨子线挪用」失效模式完全一致——只不过这次被挪用的号本身就是假的。
3. 真正的主问题:真号伪书目(本轮新命名)
3.1 什么是”真号伪书目”
pulse-pressure-normalization-hri-algorithm.md 第 105–106 行:
3. Zhang Y, et al. "Real-time pressure feedback in pulse acquisition: impact on
signal stability." Biomed Signal Process. 2021; PMID: 31730401
4. Chen X, et al. "Pressure-harmonic gain curves for six primary pulse types."
Chin J Integr Med. 2022; PMID: 33901427
两个 PMID 都真实存在。但:
| PMID | 真实文献 | 文档声称 |
|---|---|---|
| 31730401 | Schnyer RN 等,J Altern Complement Med 2019,”Can Reliability of the Chinese Medicine Diagnostic Process Be Improved? Results of a Prospective RCT” | 「Zhang Y, Biomed Signal Process 2021,实时压力反馈」 |
| 33901427 | Weaver NA 等,Lancet Neurol 2021,”Strategic infarct locations for post-stroke cognitive impairment”(卒中) | 「Chen X, Chin J Integr Med 2022,六种脉象压力-谐波增益曲线」 |
→ 作者、标题、期刊、年份、主题,五项全错;只有那串数字是真的。
3.2 为什么这条最危险
| 层面 | 后果 |
|---|---|
| 存在性检查 | 查不出——PMID 是真的,E1 全绿 |
| 引用数统计 | 不受影响——看起来引了 5 篇,实际可用 0 篇 |
| 读者抽查 | 点进 PubMed 会看到一篇真实论文,反而增强可信感 |
| 后果 | 「弦脉 G(P) 斜率是浮脉 1.7 倍」这种具体数值论断,挂在一篇完全无关的卒中影像论文上 |
第 1 轮把这类判为「张冠李戴(B 类)」,当时判断是”不需重新找文献,只需归位“。本轮修正:这个判断只在”号本身有正确出处”时成立。 本例中,被挪用的号在池内没有任何一处正确引用(33901427 例外,见 §3.3),所以”归位”无门——必须重新找文献。
3.3 同一个号,一处对一处错(第 2 轮判据仍成立)
33901427 在 tcm-pulse-hri-inter-operator-reliability-training-standardization.md 里被正确使用:
「虽然 Weaver 等人(2021)关于卒中认知损伤的研究并非脉诊研究,其核心方法论价值在于提供了多中心、大样本条件下评估操作者间信度的标准化范式」
→ 这是完全合规的引用:作者对、年份对、主题说清了、并明确声明了”不是脉诊研究”的边界。
同一个号在 pulse-pressure-normalization-hri-algorithm.md 里被伪装成”陈 2022 六脉增益曲线”。
→ 结论:判断错配必须看”这个号在这篇里被用来支撑什么”,全池扫描 + 上下文比对才能定性。第 2 轮的方法论正确,本轮只是把它的适用面从 OMOP 扩到全池。
3.4 伪书目的三个结构性特征(供后续自动检测用)
| 特征 | 表现 | 可否自动化 |
|---|---|---|
| 作者姓名为常见中文拼音(Zhang/Chen/Li/Wang + 单字母) | 池③伪条目高发 | 弱信号,单用误报高 |
| 期刊名与主题匹配但非该领域主刊 | 「Biomed Signal Process」「Chin J Integr Med」配脉诊 | 需领域词表 |
| 同一条目在不同文件里被赋予不同标题 | 34479992 三身份 | ✅ 可自动(同一 PMID 跨文件标题不一致 → 强信号) |
→ 唯一可低成本自动化的是第三条。已建议加入扫描器(见 §6)。
4. 本轮额外确认:谐波主线”域外文献”41 个,但多数是合法的
对 53 篇谐波/HRI 文件做”标题不含脉搏/心血管/中医词”的粗筛,得 41 个域外文献。逐一核对后,绝大多数是合规引用:
| 类型 | 例 | 判定 |
|---|---|---|
| 方法学借用(明确说明) | 26653405(中介分析指南)、27330520(ICC 指南)、24240655(NRI 评价) | ✅ 合规 |
| 运动生理学(去训练课题) | 10999420、2692122、7219129、8440543、33629972 | ✅ 合规(主题就是去训练) |
| 联邦学习(隐私课题) | 37044052、38340728、39766014、39081567 | ✅ 合规 |
| 完全无关 | 34166332(HIV 检测趋势)、33081257(铸铁柱 CFRP 加固)、34637011(LGBTQ 青少年干预)、31787578(口香糖黏附性)、35262535(牦牛奶乳酸菌)、35343188(肺癌 EGFR 液体活检)、39157688(癫痫手术)、32451380(风速预测)、35127882(马牙修复) | ⛔ 错配 |
→ 错配集中在 pulse-harmonic-gut-microbiome-organ-axis.md 与 pulse-harmonic-spectrum-individualized-zangfu-functional-scoring.md 两个文件,其余为合规引用。这说明错配不是全池弥漫,而是文件级爆发。
5. 处置建议(分级,全部待 JOE 决策,本轮不动手)
| 级 | 对象 | 动作 | 成本 | 是否需 JOE |
|---|---|---|---|---|
| P0 | 2 个”肠-心轴”文件(gut-microbiome-organ-axis、spectrum-individualized-zangfu-scoring) |
降级 frontmatter evidence_level 至 C,并在正文加”引文待重建”横幅(不改论断,只加免责) |
低 | ✅ 需批准(改既有正文) |
| P1 | 4 个不存在号(E1) | 删除或替换为真实文献;34479992 三处必须分别处理(三身份) |
中 | ✅ 需批准 |
| P2 | pulse-pressure-normalization-hri-algorithm.md 第 105–106 条伪书目 |
重新检索真实文献(实时压力反馈 / 六脉压力-谐波增益,确有其题但号错) | 中 | ✅ 需批准 |
| P3 | OMOP 9 篇(第 1/2 轮遗留) | 原 A/B 方案,范围须含健康主线(第 6 轮已提) | 高 | ✅ 已挂 6 轮未决 |
| P4 | 全池 3162 个 PMID 的语义比对 | 工具化 + 分批人工 | 极高 | 建议只做”跨文件同号异题”自动检测 |
5.1 本轮最想请 JOE 拍的一句话
是否接受”先用横幅标注 + 降级 evidence_level,暂不逐条重检文献”这个过渡方案?
理由:全量重检 3162 个 PMID 的语义,成本远超本池季度产能;而”横幅 + 降级”是可逆、低成本、且能立刻止住”读者被误导”这个最大风险。文献重建可以按 P1→P2→P3 分批做。
6. 工具侧待办(下轮可自主)
- 加”跨文件同号异题”检测(§3.4 第三条,唯一低成本可自动化的伪书目信号)——输入两个文件、同一 PMID、两处声明的标题首词不同 → 报警。
citation-verify.sh升级:现有版本只回查”号是否真实 + 真实文献是什么”,不比对文中的声明。建议增加--claim参数,让写作者粘贴自己的书目行,脚本做词元重叠度打分(阈值 <0.2 即报”疑伪书目”)。- 覆盖率闸门:任何扫描跑完,”未提取行”必须为 0 才允许出结论(本轮首版漏提 74 行,差点得出错误结论)。
7. 证据边界(必读)
- 本轮只对谐波主线 53 篇做了全量域外文献筛查 + 17 篇人工逐条比对;其余 668 篇只过了 E1/E2 确定性判据。
- 因此「错配总数」是下界,不是全量。真实的真号伪书目数量可能远高于本文列出的 9 个文件。
- 本文所有”真实文献”信息来自 NCBI E-utilities
esummary/efetch,未做全文精读。 - 本文未修改任何既有正文(含错误的 9 个文件一字未动)。
- 本文不构成任何健康/医学结论。
7.1 方法学锚点(3 个 PMID,E-utilities 逐条回查 title/首作者/期刊一致)
本文所用审计方法并非自创,医学文献学界对”引用不准确”已有系统性量化研究:
| # | PMID | 首作者 | 载体 | 年 | 在本文中的角色 |
|---|---|---|---|---|---|
| M1 | 40696441 | Baethge C | Res Integr Peer Rev | 2025 | 引用不准确性的系统综述与 meta 分析——本文的”错配是系统性而非偶发”这一判断的上游方法学依据 |
| M2 | 28910404 | Mogull SA | PLoS One | 2017 | 医学论文中”被引事实”准确性的方法学与报告规范综述——支撑本文”必须比对声明而非只查存在性”的方法论选择 |
| M3 | 32977312 | Montenegro TS | J Neurosurg Spine | 2021 | 单个专科的参考文献准确性实测——说明这类问题在各领域普遍存在,非本池特有 |
三点共同的支撑作用:把本池的引文问题从「agent 出错的个案」重新定位为「文献学上已知的系统性现象」,从而支持 §3.2 的核心判断——存在性检查对这类失效无效,必须比对声明内容。
免责声明
本文为引文完整性审计,不涉及任何医学结论,也不构成对既有文章观点的支持或否定。所列错配仅为书目层面的定位,不代表原文的学术论断必然错误——但在文献重建完成前,相关文章的论断不得被引用为证据。
池③ 课题 #190|勘探员:WB-FIVEPOOLS|2026-10-18|证据等级:C|状态:draft(未标 accepted,未写 779/782)
免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。
证据与局限
本文证据等级为 C,属 measurable,结论限于研究层面的推断,不构成对个体情形的判定。不是假号(只有 4 个),而是”PMID 真实 + 书目信息整体虚构”这一类此前没被命名的失效模式。
局限:本文未覆盖全部相关文献,效应量与结论随样本与方法变化;传统体系与现代研究的对照属结构类比而非等价证明,请结合专业判断参考。
Pool 3: Full Localization of Citation Mismatch – Upgrade in Determination from \”False ID\” to \”True ID with Fake Bibliography\
Evidence level: C
Pool ③ · Project #190 (Newly Registered in This Round) · Building upon Citation Audits from Rounds 1/2/6
Evidence Level C | Status draft | Writings to existing text in this round = 0
0. One-Sentence Conclusion
Across 721 articles and 3,162 unique PMIDs, only four IDs are genuinely non-existent; the impression from earlier rounds that “many fake IDs exist” is incorrect. The real issue lies elsewhere: the PMIDs are authentic, but the author names, titles, journals, and years listed alongside them are fabricated. I term this phenomenon “Real-ID / Fabricated-Bibliography.” For this category, PubMed’s existence verification will never flag a discrepancy.
1. Methods and Tools (New in this Round)
New tool: 08-ops/_tools/citation-pool-scan.py (pure standard library, passes py_compile).
Design principles are derived from the lessons of Round 37: only report errors that can be falsified without semantic judgment, preferring false negatives over false positives:
| Criterion | Certainty | Enters Conclusion |
|---|---|---|
| E1 PMID missing | Certain (confirmed via esummary + efetch) |
✅ |
| E2 PMC ID treated as PMID | Certain (verified via elink reverse lookup) |
✅ (0 hits in this round) |
| E3 Year contradiction | Uncertain (text years often refer to cohort/event years) | ⛔ Downgrade to INFO; exclude from conclusion |
| Topic relevance | Uncertain | ⛔ No judgment; leave for human review |
| Author validity | Uncertain | ⛔ No judgment; leave for human review |
1.1 Three Pitfalls Encountered by the Tool Itself (All Fixed and Logged Here to Prevent Recurrence)
| # | Pitfall | Consequence | Fix | Verification |
|---|---|---|---|---|
| 1 | Markdown bolding syntax like PMID \*\*40892475\*\* fails to match |
Silent omission: reads 5 IDs in a file as 0 | Allow ≤6 non-digit characters between marker and number | File now reads all 5 IDs after fix |
| 2 | pmid: ["36106418", ...] with quotes, singular form, or in frontmatter |
Misses frontmatter layer | Split by [,\s"'] + support single-value/dash notation |
Line 74 now extracted (was missing) |
| 3 | PMID regex lacks re.I, causing all lowercase pmid: in text to be missed |
Large-scale silent omission | Added re.I globally |
Unique PMID 3153 now captured (previously 3152; coverage improved from 74 missing to 13 missing) |
1.2 Why E3 Must Be Downgraded (Empirical Data)
The initial version used “neighborhood ±1 lines” to judge year contradictions, reporting 866 cases. Manual sampling of 12 cases showed ≥11 were false positives:
| Stated Year | Actual Year | Actual Meaning |
|---|---|---|
| 2015 | 2000 | Text refers to “solar eclipse records from 720 BC–AD 2015” — data span, not publication year |
| 1960 | 2007 | Text refers to “after electrification in the 1960s” — event era |
| 1996 | 2013 | Text refers to “three birth cohorts: 1940, 1977, 1996” — cohort year |
Restricting to “within the first 45 characters preceding each PMID on the same line, excluding interval notation” reduced cases to 125. Sampling 12 again showed ≥11 false positives (same issues, just repositioned). → Conclusion: This criterion is unreliable for this corpus; downgrade to INFO.
This lesson warrants separate documentation: The pool corpus heavily employs narrative phrasing like “Year X Study.” The proximity between year and PMID does not constitute a declaration of publication year. Any year validation based on proximity will yield widespread false positives in this corpus.
2. Full Pool Scan Results
| Indicator | Value |
|---|---|
| Scanned Files | 721 (03-sites/cn/research/science-yixue/) |
| Unique Declared PMID Count | 3,162 |
| Coverage Gap (PMID present but not extracted) | 13 lines (all illustrative text within audit documents, not citations) |
| E1 Non-Existent | 15 instances / 4 unique IDs / 6 files |
| E2 PMC Match for PMID | 0 |
| E3 Year Contradiction | 0 (downgraded, see §1.2) |
2.1 E1 Full Hits (4 unique IDs, 15 instances total)
| PMID | Affected Files | Stated Purpose |
|---|---|---|
| 15199541 | standing-wave-antinode-psychophysiology.md (3 instances) |
«HPA 2009 Infrasound: a brief overview of sources and effects» |
| 34479992 | dosha-qi-rsa-falsification-framework-methodology.md (2), dosha-qi-rsa-falsification-preregistration-specification.md (2), pool5-standby-validation-entry-exit-gate.md (1) |
«Chung 2021 TCM Syndrome HRV» / «Putnick & Bornstein 2016 Measurement Equivalence Systematic Review» / «HRV Cross-Ethnic Reference Value Equivalence» — the same ID assigned three distinct identities across three locations |
| 28726973 | pulse-signature-detraining-regression-dynamics.md (2) |
Listed in reference section |
| 16030677 | homeopathy-high-dilution-frequency-memory.md (1) |
Suspected Rey luminescence study |
(Additionally, rtribe-omop-citation-integrity-audit.md contains 4 instances; these belong to the audit document’s own statement that “these IDs do not exist,” not new errors.)
The three identities assigned to 34479992 are particularly notable: A non-existent ID is described by three articles as “Chung 2021 HRV,” “Putnick & Bornstein 2016 Measurement Equivalence,” and “HRV Cross-Ethnic Equivalence.” This aligns perfectly with the failure mode identified in Round 2 as «cross-subline appropriation» — except here, the appropriated ID itself is fabricated.
3. The Core Question: Real IDs, Fake Bibliographies (New Naming for This Round)
3.1 What is a “Real ID, Fake Bibliography”
Lines 105–106 in pulse-pressure-normalization-hri-algorithm.md:
3. Zhang Y, et al. "Real-time pressure feedback in pulse acquisition: impact on
signal stability." Biomed Signal Process. 2021; PMID: 31730401
4. Chen X, et al. "Pressure-harmonic gain curves for six primary pulse types."
Chin J Integr Med. 2022; PMID: 33901427
Both PMIDs do exist. However:
| PMID | Actual Literature | Document Claim |
|---|---|---|
| 31730401 | Schnyer RN et al., J Altern Complement Med 2019, “Can Reliability of the Chinese Medicine Diagnostic Process Be Improved? Results of a Prospective RCT” | 「Zhang Y, Biomed Signal Process 2021, real-time pressure feedback」 |
| 33901427 | Weaver NA et al., Lancet Neurol 2021, “Strategic infarct locations for post-stroke cognitive impairment” (Stroke) | 「Chen X, Chin J Integr Med 2022, six pulse types pressure-harmonic gain curves」 |
→ Author, title, journal, year, and topic are all incorrect; only the number is authentic.
3.2 Why This Is Most Dangerous
| Aspect | Consequence |
|---|---|
| Existence Check | Undetectable — PMID is valid, Evidence Level 1 shows all green |
| Citation Count | Unaffected — Appears to cite 5 papers, but 0 are usable |
| Reader Verification | Clicking into PubMed reveals a real paper, paradoxically enhancing credibility |
| Consequence | Specific numerical claims like “the G(P) slope of the ‘Xian pulse’ is 1.7 times that of the ‘Fu pulse’” are attached to a completely unrelated stroke imaging paper |
In Round 1, this was classified as “Wrong Attribution (Level B),” judged as “no need to re-find literature, just reposition.” Correction for this round: This judgment holds only when the ID itself has a correct source. In this case, the misappropriated IDs have no correct citations anywhere in the pool (except 33901427, see §3.3), so “repositioning” is impossible — new literature must be sought.
3.3 One ID, One Correct and One Incorrect Usage (Round 2 Criteria Still Valid)
In tcm-pulse-hri-inter-operator-reliability-training-standardization.md, 33901427 is used correctly:
“Although Weaver et al. (2021) regarding stroke cognitive impairment is not a pulse diagnosis study, its core methodological value lies in providing a standardized paradigm for assessing inter-operator reliability under multi-center, large-sample conditions”
→ This is fully compliant: Author matches, year matches, topic clarified, and the boundary explicitly states it is “not a pulse diagnosis study.”
The same ID is disguised in pulse-pressure-normalization-hri-algorithm.md as “Chen 2022 Six Pulse Gain Curves.”
→ Conclusion: Misalignment judgment must examine “what this ID is used to support in this context.” Full-pool scanning + contextual comparison is required for qualitative determination. Round 2 methodology is correct; this round simply expands its applicability from OMOP to the entire pool.
3.4 Three Structural Features of Fake Bibliographies (For Future Automated Detection)
| Feature | Manifestation | Automatable? |
|---|---|---|
| Author name is common Chinese Pinyin (Zhang/Chen/Li/Wang + single letter) | High frequency in Pool 3 fake entries | Weak signal; high false positive rate if used alone |
| Journal name matches topic but is not the primary journal for that field | Biomed Signal Process or Chin J Integr Med paired with pulse diagnosis | Requires domain term list |
| Same entry assigned different titles across files | 34479992 with three identities | ✅ Yes (inconsistent titles for same PMID across files → strong signal) |
→ Only the third feature can be automated at low cost. Adding a scanner for this is already suggested (see §6).
4. Additional Verification: Harmonic Mainline “Foreign Literature” (41 Items), Most Legitimate
A preliminary screening was conducted on 53 harmonic/HRI documents, filtering out those with titles containing terms related to pulse, cardiovascular systems, or Traditional Chinese Medicine. This yielded 41 foreign literature items. Upon individual verification, the vast majority constitute compliant citations:
| Type | Examples | Verdict |
|---|---|---|
| Methodological Borrowing (Explicitly Stated) | 26653405 (Mediation Analysis Guidelines), 27330520 (ICC Guidelines), 24240655 (NRI Evaluation) | ✅ Compliant |
| Exercise Physiology (De-training Focus) | 10999420, 2692122, 7219129, 8440543, 33629972 | ✅ Compliant (Topic is de-training) |
| Federated Learning (Privacy Focus) | 37044052, 38340728, 39766014, 39081567 | ✅ Compliant |
| Completely Irrelevant | 34166332 (HIV testing trends), 33081257 (CFRP reinforcement of cast iron columns), 34637011 (LGBTQ+ adolescent intervention), 31787578 (gum adhesion), 35262535 (yogurt lactobacillus), 35343188 (lung cancer EGFR liquid biopsy), 39157688 (epilepsy surgery), 32451380 (wind speed prediction), 35127882 (natal tooth repair) | ⛔ Mismatch |
→ The mismatches are concentrated in two files: pulse-harmonic-gut-microbiome-organ-axis.md and pulse-harmonic-spectrum-individualized-zangfu-functional-scoring.md. All other citations are compliant. This indicates the mismatch is not a pervasive issue across the entire pool, but rather a file-level outbreak.
5. Actionable Recommendations (Tiered; Pending JOE Decision; No Action This Round)
| Tier | Target | Action | Cost | Requires JOE Approval |
|---|---|---|---|---|
| P0 | 2 “gut-heart axis” files (gut-microbiome-organ-axis, spectrum-individualized-zangfu-scoring) |
Downgrade frontmatter evidence_level to C and add a “citation pending reconstruction” banner in the body (do not alter assertions; only add disclaimers) |
Low | ✅ Yes (modifies existing body text) |
| P1 | 4 non-existent citations (E1) | Delete or replace with valid references; handle 34479992 in three instances separately (three distinct identities) |
Medium | ✅ Yes |
| P2 | pulse-pressure-normalization-hri-algorithm.md lines 105–106 (pseudo-bibliography) |
Re-search for valid literature (real-time pressure feedback / six-pulse pressure-harmonic gain; title exists but ID is incorrect) | Medium | ✅ Yes |
| P3 | OMOP 9 papers (carried over from rounds 1/2) | Retain original A/B approach; scope must include the health mainline (raised in round 6) | High | ✅ Flagged as unresolved since round 6 |
| P4 | Semantic comparison of all 3,162 PMIDs in the pool | Tool-assisted + batched manual verification | Extreme | Recommend limiting to automatic detection of “same ID, different title across files” |
5.1 One-Sentence Request for JOE This Round
Should we accept the transitional approach of “first annotating with banners and downgrading
evidence_level, then postponing individual literature re-verification”?
Rationale: Conducting a full semantic re-verification of 3,162 PMIDs exceeds the quarterly production capacity of this repository. Conversely, the “banner + downgrade” strategy is reversible, low-cost, and immediately mitigates the greatest risk: misleading readers. Literature reconstruction can proceed incrementally following the P1 → P2 → P3 sequence.
6. Tool-Side To-Dos (Autonomous for Next Iteration)
- Implement “Cross-File Same Number, Different Title” Detection (§3.4, Item 3; the only low-cost, fully automated pseudo-bibliographic signal) — Trigger an alert when two files share the same PMID but declare different first words in their titles.
- Upgrade
citation-verify.sh: The current version only verifies “whether the number is real and what the actual literature is,” but does not compare textual declarations. Recommend adding a--claimparameter allowing authors to paste their bibliography lines, enabling the script to calculate token overlap scores (flagging “suspected pseudo-bibliography” if threshold <0.2). - Coverage Gate: No conclusion may be drawn unless the count of “unextracted lines” is zero after any scan run (the initial draft missed 74 lines, nearly leading to erroneous conclusions).
7. Evidence Boundaries (Mandatory Reading)
- This round conducted full-domain foreign literature screening for the 53 harmonic mainline articles and manual line-by-line comparison for 17 articles; the remaining 668 articles passed only the E1/E2 determinative criteria.
- Therefore, the “total mismatch count” represents a lower bound, not a comprehensive total. The actual number of true false-positive bibliographies may far exceed the 9 files listed in this article.
- All “verified literature” information in this text originates from NCBI E-utilities
esummary/efetch; full-text reading was not performed. - This text has not modified any existing body text (including the 9 erroneous files, which remain unchanged).
- This text constitutes no health or medical conclusions.
7.1 Methodological Anchors (3 PMID, E-utilities verification of title/first author/journal consistency)
The audit methodology employed here is not novel; the medical literature community has conducted systematic quantitative research on “citation inaccuracies”:
| # | PMID | First Author | Vehicle | Year | Role in This Text |
|---|---|---|---|---|---|
| M1 | 40696441 | Baethge C | Res Integr Peer Rev | 2025 | Systematic review and meta-analysis of citation inaccuracies — providing the upstream methodological basis for this text’s judgment that mismatches are systemic rather than sporadic |
| M2 | 28910404 | Mogull SA | PLoS One | 2017 | Methodological review and reporting standards for the accuracy of “citation facts” in medical papers — supporting the methodological choice to compare statements rather than merely verify existence |
| M3 | 32977312 | Montenegro TS | J Neurosurg Spine | 2021 | Empirical measurement of reference accuracy within a single specialty — demonstrating that such issues are prevalent across fields and not unique to this corpus |
Three shared supporting roles: Re-positioning the citation issues in this corpus from “isolated agent errors” to a “systematic phenomenon known in bibliometrics,” thereby supporting the core judgment in §3.2 — existence checks are ineffective for such failures; statement comparison is required.
Disclaimer
This text serves as a citation integrity audit and does not involve any medical conclusions, nor does it constitute support or rejection of existing viewpoints in the referenced articles. The listed mismatches represent bibliographic positioning only and do not imply that the original authors’ academic assertions are necessarily incorrect; however, until the literature reconstruction is completed, assertions from related articles shall not be cited as evidence.
Pool ③ Project #190 | Explorer: WB-FIVEPOOLS | 2026-10-18 | Evidence Level: C | Status: draft (not yet accepted, count not yet finalized at 779/782)
Leave a Reply