lit-review

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

lit-review — no shortcuts

lit-review — 不走捷径

A literature review fails silently. Every other check in this harness fails loudly — a control that doesn't recover the planted answer, digits that disagree with the independent route. A missed paper produces no error message: the review reads as complete, the claim goes out, and the miss surfaces later, from a referee or from the missed author.

Examined afterward, every miss traces to a shortcut that felt reasonable at the time: the searchers verified the papers they already knew instead of hunting the ones they didn't; queries stopped a year or two short of the present; abstracts were read where full texts were owed; a paywall quietly downgraded a paper from "read" to "skimmed". None of these announce themselves. The protocol's answer is to make each one visible: every form of thoroughness claimed here is backed by a receipt a second reader can check, or it does not count.

Refereeing runs the same protocol in reverse — the submission's novelty claims become C1..Cn and the verdicts belong to someone else's paper. When refereeing, follow the venue's rules on AI assistance and manuscript confidentiality; many prohibit sharing a manuscript under review with AI tools. Where that is the rule, run this protocol on the public literature around the submission's claims, never on the confidential text itself, and disclose the assistance to the editor if the venue asks.

文献综述会悄无声息地失败。这个检查框架中的其他每项检查都会大声报错——无法找回预埋答案的对照项、与独立路径所得数字不一致的指标。漏掉一篇论文不会产生任何错误信息：综述看起来已经完整，声明就此对外发布，而遗漏可能要到日后才由审稿人或被漏掉的作者暴露出来。

事后审视，每次遗漏都可追溯到当时看似合理的捷径：检索人员只核验了自己已经知道的论文，而没有寻找尚不知道的论文；检索词停在一两年前，未覆盖当下；本应阅读全文时却只读了摘要；付费墙悄然把一篇论文从“已读”降级为“略读”。这些情况都不会主动暴露。协议的答案是让每一种情况都变得可见：这里声称的每一种彻底性都有另一位读者可核验的回执作为支撑，否则就不予采信。

同行评审以反向方式运行同一协议——投稿中的新颖性声明变成 C1..Cn，而裁决对象是别人的论文。进行同行评审时，应遵守目标 venue 关于 AI 辅助和稿件保密性的规则；许多 venue 禁止使用 AI 工具处理正在评审的稿件。如果规则如此规定，只能针对投稿声明周围的公开文献运行本协议，绝不能处理保密文本本身；如果 venue 要求，还应向编辑披露 AI 辅助情况。

The proximity scale

接近度等级

Grade every candidate work as you go; the heavy obligations below key off the grade.

P0–P1 — same broad area, different question. Note or discard.
P2 — shares a component (a technique, a dataset, a subresult) but not the question. Goes in the credit list.
P3 — attacks the same question with different tools, or a neighboring question with the same tools. Full-text read required.
P4 — overlaps part of a claim: a special case proved, or the same result under different assumptions. Full-text read, verbatim quote, chains both directions.
P5 — contains the claim. The review's reason for existing.

边处理边为每篇候选文献评级；下面较重的义务要求均以评级为依据。

P0–P1 — 属于同一宽泛领域，但研究问题不同。记录或丢弃。
P2 — 共享某个组成部分（一种技术、一个数据集或一个子结果），但研究问题不同。进入已有工作认可列表（credit list）。
P3 — 使用不同工具研究同一个问题，或使用相同工具研究相邻问题。必须阅读全文。
P4 — 与某项声明的一部分重叠：证明了某个特殊情形，或在不同假设下得到相同结果。必须阅读全文、逐字引用，并双向追踪引用链。
P5 — 包含该声明。这是综述存在的根本理由。

The three iron rules (violations invalidate the review)

三条铁律（违反任一条即导致综述无效）

1. Read the papers. Nothing from memory.

Every work graded P3 or above is read in full text — the PDF fetched and read, not the abstract, not another paper's summary of it, not what the model remembers. The relevant theorem, definition, or construction is quoted verbatim with a page or section number. If the full text cannot be obtained (paywall without institutional access, no author-posted copy), the work is marked FULL-TEXT-UNREAD and its grade is a ceiling estimate, flagged as such — never silently downgraded to what the abstract suggests.
No citation, date, venue, or claim about what a paper shows comes from memory. Memory generates the search query; the source generates the sentence. Verification means a live fetch of the primary record (publisher page, preprint server, author PDF), with the route recorded.
Gated or scanned PDFs: use legitimate routes only — an open-access or preprint version, the author's own posted copy, your institution's subscription, interlibrary loan, or a request to the author; read a scanned PDF page by page with vision if that is what it takes. Never bypass access controls or use unauthorized repositories. If no legitimate full text is available, mark the work FULL-TEXT-UNREAD (grade = ceiling) and list it as owed.

2. Thoroughness is falsifiable, not asserted. Every review carries receipts:

A search ledger — every query actually run, on which engine or index, with the year range. A query that never named the current year is a defect.
A read ledger — which works were read full-text, which abstract-only, which flagged unread. A review whose read ledger is empty at P3+ is not a review.
A gap statement — what was NOT searched: fields, languages, venues, year ranges, any channel that died on a budget or access wall. Implied completeness is the exact lie this skill exists to prevent.
Iterate to dry — rounds continue until a fresh round (a new angle, a recency re-sweep, chains run on newly found works) returns nothing at P3+. Report the rounds run and what each added. One pass has never been enough.

3. Track citation chains, both directions. For every work at P3+:

Backward — mine its reference list for ancestors no search named. Record the chain (A cites B cites C): the chain is itself a deliverable, because it is how the next reviewer verifies coverage.
Forward — find who cites it: cited-by listings sorted newest first, the most recent two or three years read closely. This is how you find the paper whose name you don't know — it cites the ancestor you do.
Per-author recency — every author appearing at P3+ gets their last few years swept through the current month: personal website (publications and preprint pages, including linked working papers — drafts are often posted there long before journals), profile pages sorted by date, preprint and working-paper listings. Publicly posted drafts and working papers count fully.

1. 阅读论文。不得依赖记忆。

每篇评级为 P3 或以上的文献都须通读全文——获取并阅读 PDF，而不是摘要、另一篇论文对它的概述，或模型记忆中的内容。相关的定理、定义或构造必须逐字引用，并注明页码或章节编号。如果无法取得全文（付费墙且无机构访问权限、没有作者发布的副本），则将文献标记为 FULL-TEXT-UNREAD，其等级仅作为上限估计并明确标注这一点——绝不能悄然降级为按摘要推断的等级。
任何引用、日期、venue，或关于论文证明内容的陈述，都不得来自记忆。记忆可以用来生成检索式；句子必须来自原始来源。核验意味着实时获取主要记录（出版商页面、预印本服务器或作者 PDF），并记录获取路径。
受限或扫描版 PDF：只能使用合法途径——开放获取版本或预印本版本、作者自行发布的副本、所在机构的订阅、馆际互借，或向作者索取；必要时，应使用 vision 逐页阅读扫描版 PDF。绝不绕过访问控制，也不使用未经授权的存储库。如果无法通过合法途径取得全文，将文献标记为 FULL-TEXT-UNREAD（等级 = 上限），并列为待完成事项。

2. 彻底性必须可证伪，而不是口头断言。 每篇综述都必须附带回执：

检索账本 — 记录实际执行过的每条检索式、使用了哪个搜索引擎或索引，以及年份范围。一次从未包含当前年份的检索式就是缺陷。
阅读账本 — 记录哪些文献已读全文、哪些只读了摘要、哪些被标记为未读。如果 P3+ 级别的阅读账本为空，那就不是一篇综述。
缺口声明 — 说明没有检索什么：领域、语言、venue、年份范围，以及任何因预算或访问限制而中断的渠道。暗示覆盖完整，恰恰是本 skill 要防止的谎言。
迭代至无新增结果 — 持续执行多轮检索，直至新一轮检索（采用新角度、重新扫查近期文献，或基于新发现文献追踪引用链）在 P3+ 级别不再返回任何结果。报告执行了哪些轮次，以及每轮新增了什么。单轮检索从来都不够。

3. 双向追踪引用链。 对每篇 P3+ 文献：

向后追踪 — 从其参考文献列表中找出搜索未点名的祖先文献。记录引用链（A 引用 B，B 引用 C）：引用链本身也是一项交付物，因为下一位审阅者正是借此核验覆盖范围。
向前追踪 — 找出引用该文献的作品：将被引列表按最新优先排序，并仔细阅读最近两至三年的内容。这是找到那些名称尚不知道的论文的方法，因为它会引用你已经知道的祖先文献。
按作者扫查近期成果 — 对 P3+ 中出现的每位作者，扫查其最近几年截至当前月份的产出：个人网站（出版物和预印本页面，包括链接的工作论文——草稿通常早在期刊发表前就已发布）、按日期排序的个人资料页、预印本和工作论文列表。公开发布的草稿和工作论文完全计入。

Places to look

需要查找的渠道

Use every service within its terms. Prefer documented APIs (Semantic Scholar, OpenAlex, Crossref, arXiv, INSPIRE, PubMed, ADS); query interactive-only services such as Google Scholar by hand or at human pace, never from a fan-out of agents; rate-limit every automated fetch (pause ~2 s between calls) and honor robots.txt; list any service you could not lawfully query in the gap statement.

Tick each, or put it in the gap statement:

Google Scholar — keyword and quoted-phrase queries; cited-by chains; author profiles sorted by year.
Semantic Scholar (site and API) — citations and references programmatically; good recall on preprints.
The field's preprint server — arXiv by category plus full-text search, bioRxiv/medRxiv, SSRN, OpenReview for machine-learning-adjacent claims. The frontier lives here and on authors' pages, not in journals.
Working-paper series, where the field has them — institutional series, author listings, new-paper feeds.
Author websites — read the publications page, CV and any preprints or drafts the author has publicly linked; the newest work is often posted there before any index has crawled it. Do not guess URLs or enumerate directories.
Review articles, surveys, handbook chapters in the claim's territory — a field's own synthesis is a pre-built citation chain; read the newest one you can find.
Journal tables of contents, the last couple of years, for the three or four venues where the nearest ancestors published.
Course syllabi and lecture notes (site:edu filetype:pdf) — instructors track frontier papers faster than journals, and usually link the preprint version.
Software registries (CRAN, PyPI, the package archives of statistical environments) — a rival method often exists as a package before or beside its paper.
Dissertations — when a research line's students are active, the newest statement of the line is a thesis.
Field review databases where they exist (zbMATH, MathSciNet, INSPIRE, PubMed).
The claim's own vocabulary, and every synonym. Search the project's coined terms in quotes, plain paraphrases, and the whole synonym family — the same object wears a different name in every field that touched it. Build the synonym list as you go and re-run every search under each new name found. The phrase hunt does two jobs: it protects your naming, and it finds the rival wearing another field's name for your object.

在各自服务条款允许的范围内使用每项服务。优先使用有文档说明的 API（Semantic Scholar、OpenAlex、Crossref、arXiv、INSPIRE、PubMed、ADS）；对于仅支持交互的服务（如 Google Scholar），应手动或以人工速度查询，绝不能通过 Agent 批量展开；每次自动抓取都要限速（调用之间暂停约 2 秒），并遵守 robots.txt；任何无法合法查询的服务都应列入缺口声明。

逐项勾选，或写入缺口声明：

Google Scholar — 关键词和精确短语检索；被引引用链；按年份排序的作者主页。
Semantic Scholar（网站和 API）— 以编程方式获取引用和参考文献；对预印本有很好的召回率。
该领域的预印本服务器 — 按类别浏览 arXiv 并进行全文检索；bioRxiv/medRxiv、SSRN，以及用于机器学习相关声明的 OpenReview。领域前沿存在于这里以及作者主页上，而不在期刊中。
工作论文系列（如果该领域有此类系列）— 机构系列、作者列表、新论文信息流。
作者网站 — 阅读出版物页面、CV，以及作者公开链接的任何预印本或草稿；最新作品往往会在任何索引抓取之前发布在这里。不要猜测 URL，也不要枚举目录。
声明领域内的综述文章、survey 和 handbook 章节 — 该领域自己的综合研究就是一条预先构建好的引用链；找到能够获得的最新一篇并阅读。
期刊目录 — 查看最近两三年的目录，聚焦最接近的祖先文献发表过的三到四个 venue。
课程 syllabus 和讲义（site:edu filetype:pdf）— 教师追踪领域前沿的速度通常快于期刊，而且通常会链接预印本版本。
软件注册表（CRAN、PyPI，以及统计环境的软件包归档）— 竞争方法往往会先以软件包形式出现，随后或同时出现对应论文。
学位论文 — 如果某条研究路线仍有活跃学生，该路线最新的论述可能是一篇 thesis。
领域评审数据库（如存在）— zbMATH、MathSciNet、INSPIRE、PubMed。
声明自身的词汇及其每个同义词。 使用引号检索项目自造术语、普通释义以及整个同义词族——同一对象在每个接触过的领域中都有不同名称。在检索过程中不断构建同义词列表，并使用每个新发现的名称重新执行所有检索。短语检索有两个作用：保护你的命名，同时找到使用另一个领域名称称呼同一对象的竞争工作。

Procedure

执行流程

0. Claims first. Before any search, write the numbered novelty claims C1..Cn, plus the exact headline sentences the paper or announcement will carry. Each headline gets a verdict at the end: survives as worded / falsified (quote the falsifier) / survives only reworded (propose the wording). A review run without pinned claims drifts into confirming a mood.

1. Fan out by lineage. One searcher per lineage or subfield that could own the result — six to twelve angles for a serious claim (parallel subagents if your framework supports them; sequential passes otherwise). Any known-relevant names in a brief are a floor, never the hunt list: a searcher told "check Smith and Jones" will verify Smith and Jones and under-search everyone else. Each angle keeps its own ledgers.

2. Chains and recency (iron rule 3) on everything the angles surface at P3+. This is where the papers no query can name get found.

3. Synthesize. A deduplicated table of the closest works — P4/P5 entries with verbatim quotes and locations — plus a credit list (P2–P3), diffed against the paper's actual bibliography. Then per-claim novelty verdicts, stated against interest: (a) new as far as this review can see; (b) new with required credit — name the ancestor; (c) at risk as worded — name the work, quote the overlapping passage, propose the wording that survives. No softening. The verdicts are the product; a review that returns only a reading list has not finished.

4. Iterate to dry (iron rule 2), then carry the findings into the paper.

0. 先确定声明。 在检索前，写下编号的新颖性声明 C1..Cn，以及论文或公告将采用的准确标题句。每条标题句最后都要给出裁决：按原措辞成立 / 已被证伪（引用证伪依据）/ 仅在改写后成立（提出改写方案）。没有固定声明就开始的综述会偏离目标，转而确认某种主观感觉。

1. 按研究脉络展开。 为每个可能拥有该结果的脉络或子领域安排一名检索者；对于严肃的新颖性声明，应覆盖六到十二个角度（如果框架支持，可使用并行 subagent；否则顺序执行）。简报中已知的相关名称只是下限，绝不是完整检索清单：一名被要求“核查 Smith 和 Jones”的检索者只会验证 Smith 和 Jones，从而漏搜所有其他人。每个角度保留自己的账本。

2. 对所有角度发现的 P3+ 文献执行引用链与近期性追踪（铁律 3）。这正是找到那些任何检索式都无法点名的论文之处。

3. 综合。 生成去重后的最接近文献表——P4/P5 条目附逐字引文和位置——以及已有工作认可列表（P2–P3），并与论文实际参考文献表进行差异核对。然后针对每项声明给出新颖性裁决，且必须从最不利方向出发： (a) 就本次综述所能看到的范围而言为新；(b) 在必须认可已有工作的前提下为新——点名祖先文献；(c) 按当前措辞存在风险——点名相关工作，引用重叠段落，并提出可以成立的措辞。绝不能淡化问题。裁决才是最终产品；只返回阅读清单的综述尚未完成。

4. 迭代至无新增结果（铁律 2），然后将结果落实到论文中。

Findings into the paper

将发现写入论文

Prior work is motivation, never a rival. The closest ancestor goes in the introduction as the point of departure, credited generously BEFORE the delta is stated; the delta is then stated flat. Cut protective register on sight ("a referee will note", "to be clear about priority") — prose that argues with an imagined referee reads as exactly what it is.
Every citation added to the paper is copied from the review's verified entries only, with the verification route noted (the ref-check skill governs entry format and bibliography audits).
At-risk claims are reworded before anything goes out. A headline that failed its verdict never survives on "approximately true".
P2–P3 works land in the credit list even if never cited — the next revision reads the list, not the reviewer's memory.
Full-text debts that remain (a gated text, a book checked only at page level) are listed, not forgotten.

已有工作用于提供动机，绝不作为对手。 最接近的祖先文献应作为引言中的出发点，在陈述差异之前充分认可；随后再平实地陈述差异。见到防御性措辞立即删除（“审稿人会注意到”“为免关于优先权产生歧义”）——这种与想象中的审稿人争辩的行文，恰恰会暴露其真实意图。
论文新增的每条引用都只能从综述中已核验的条目复制，并注明核验路径（ref-check skill 负责条目格式和参考文献审计）。
存在风险的声明必须先改写，再对外发布。未通过裁决的标题句绝不能以“大致正确”的形式保留。
P2–P3 文献即使从未被引用，也应进入已有工作认可列表（credit list）——下一轮修改应阅读这份列表，而不是依赖审阅者的记忆。
尚未清偿的全文债务（例如受访问限制的文本、只查到页码层面的书籍）应列出，不能遗忘。

Failure-mode catalogue

失败模式目录

Each of these is a class observed in real reviews. When a review misses something, find the miss's class here — or add it; a miss without a new entry is a repeat waiting.

Canon anchoring. The brief names the famous papers; the searchers verify the famous papers; the miss is the recent work by an author nobody named. Verification effort concentrates on what you already know — hunting effort must be budgeted separately.
Recency blindness. Every query implicitly ended a year or two ago because the searcher's sense of the field did. A central author's latest papers sat unfound because nothing swept their output through the current month.
Venue bias. Searches favored journals; the field's actual frontier was on preprint servers and authors' own sites, months to years ahead. The decisive miss was a working paper posted on an author's own website that no index had crawled.
The abstract-grade read. An abstract-level skim reported a prior paper as claiming the opposite of the project's result; the full text showed the same sign and a different magnitude. A false confrontation was minted and repeated before anyone read the paper. Full-text reads are not overhead — they are where the review's claims come from.
Memory-sourced citations. An entry written from recall carries a paraphrase of the title, a plausible-but-wrong year, a truncated author list — and every field of it reads fine. The cure is upstream of any bibliography audit: nothing enters the review from memory.
Synonym blindness. The rival existed for years under another field's name for the same object. Every query used your field's term; every one returned nothing; "nothing found" was true and worthless.
The silent dead channel. A search or fetch budget ran out mid-review and a whole channel — the one that finds author drafts, say — went dark without any step failing. The angles reported normally; coverage had silently shrunk. Any searcher that hits a budget or access wall says so in its gap statement (which channel, which queries never ran), and the synthesis counts a dead channel as an owed round, never as covered. A review is not dry while a named channel died on budget.
One-pass certainty. The first pass found nothing close, and the review stopped there. Distance from the literature on pass one usually measures the queries, not the literature.
Confirmation search. Queries phrased to confirm novelty ("first proof of X") rather than to destroy it ("proof of X", each synonym, each ancestor's forward citations). The reviewer's job is to kill the claim; a review that wants the claim to survive will let it.
The convenience downgrade. A paywalled P4 candidate quietly became "checked" on the strength of its abstract because the PDF was hard to get. The grade kept its authority; the reading behind it was gone.
Implied completeness. The review reported everything it found and nothing it skipped. The reader took the union to be the whole field; the gap statement existed to prevent exactly that inference, and wasn't written.

以下每一种都是实际综述中出现过的类别。当综述遗漏某项内容时，先在此找到遗漏所属类别；如果没有，就新增一个；没有新条目的遗漏只会重演。

经典锚定。 简报点名了著名论文；检索人员便去核验这些著名论文；最终漏掉了某位无人提及作者的新近工作。核验精力集中在已知内容上；搜寻未知内容的精力必须单独分配。
近期性盲区。 每次检索实际上都停在一两年前，因为检索者对领域最新进展的感知滞后；某位核心作者的最新论文始终未被发现，因为无人将其产出扫查到当前月份。
venue 偏差。 检索偏向期刊，而该领域真正的前沿在预印本服务器和作者主页上，提前数月至数年；决定性的遗漏是一篇发布在作者个人网站、尚未被任何索引收录的工作论文。
摘要级阅读。 仅浏览摘要后，便将一篇先前论文报告为与项目结果结论相反；全文却显示符号相同、量级不同。于是一个虚假的冲突被制造出来并反复传播，直到有人真正阅读论文。全文阅读并非额外开销——综述中的论断正来源于此。
记忆来源的引用。 由回忆写出的条目带着标题的改写、看似合理但错误的年份、截断的作者名单——而每个字段看起来都没问题。解决方案必须前置到参考文献审计之前：任何内容都不得凭记忆进入综述。
同义词盲区。 竞争工作多年来一直使用另一个领域对同一对象的称呼。每次检索都只使用你的领域术语，自然一无所获；“没有找到”或许属实，却毫无价值。
静默失效的渠道。 检索或抓取预算在综述中途耗尽，导致整个渠道——例如用于寻找作者草稿的渠道——悄然消失，却没有任何步骤报错。各角度仍正常汇报，覆盖范围却已无声缩小。任何检索者一旦遇到预算或访问障碍，都必须在缺口声明中说明（是哪个渠道、哪些检索式从未执行）；综合阶段必须把失效渠道计为尚待完成的一轮，绝不能算作已覆盖。只要有名义渠道因预算耗尽而失效，综述就还没有达到“干涸”状态。
单轮即自信。 第一轮没找到接近的工作，综述便就此停止。第一轮与文献的距离通常反映的是检索式，而不是文献本身。
确认式检索。 检索式被写成用来确认新颖性（“X 的首个证明”），而不是用来推翻它（“X 的证明”、每个同义词，以及每个祖先文献的前向引用链）。审阅者的职责是击破声明；希望声明成立的综述只会放过它。
图便利的降级。 因难以获取 PDF，一篇付费墙后的 P4 候选文献被悄悄视为“已核查”，依据却只是其摘要。等级保留了权威性，支撑等级的阅读却没有了。
暗示式完整性。 综述报告了自己找到的一切，却不提自己跳过了什么。读者会把这片并集当成整个领域；缺口声明的存在正是为了防止这种推断，但它没有被写出来。

Two micro-examples

两个微型示例

The synonym family. A project coins "anchored window" for its object. The same object, in neighboring literatures, goes by tolerance region, feasibility band, and admissible set — each in a literature that never cites the others. A review that searches only the coined term and its plain paraphrase returns clean and is wrong. The working method: every time a P3+ paper names the object differently, that name becomes a fresh round of queries across every engine already used. The synonym list at the end of the review is a deliverable.

The chain that finds the unnameable paper. You cannot query for a paper you have no words for. But if it exists, it almost certainly cites the ancestor you DO know. Take the strongest known ancestor, open its forward citations sorted newest first, and read the last two years of titles closely. This is the highest-yield single move in the protocol, and it is exactly the move a keyword-only review never makes.

Sources and acknowledgments. The search ledger, read ledger and gap statement are a lightweight form of the PRISMA 2020 reporting items (Page et al., BMJ 372 (2021) n71); citation chaining in both directions is the snowballing procedure of Wohlin (EASE 2014), whose value over database search alone Greenhalgh and Peacock measured (BMJ 331 (2005) 1064). We thank the teams behind Semantic Scholar, OpenAlex, Crossref, arXiv, INSPIRE-HEP, PubMed and NASA ADS, whose open APIs make the ledgers checkable.

同义词族。 某项目为其对象自造了“anchored window”一词。同一对象在相邻文献中分别被称为“tolerance region”“feasibility band”和“admissible set”，而这些文献彼此从不引用。只检索自造术语及其普通释义的综述会得到“干净”的结果，却恰恰是错的。可行方法是：每当一篇 P3+ 文献用不同名称称呼该对象，就把这个名称变成一轮新检索，覆盖此前使用过的每个搜索引擎。综述末尾的同义词列表是一项交付物。

找到无法命名论文的引用链。 你无法检索一篇连名称都没有的论文。但如果它存在，它几乎肯定会引用你已经知道的祖先文献。取最强有力的已知祖先文献，打开其按最新优先排序的前向引用列表，仔细阅读最近两年的标题。这是本协议中收益最高的单项动作，也正是仅靠关键词的综述永远不会执行的动作。

来源与致谢。 检索账本、阅读账本和缺口声明是 PRISMA 2020 报告条目（Page et al., BMJ 372 (2021) n71）的一种轻量实现；双向引用链追踪就是 Wohlin（EASE 2014）的 snowballing 流程，其相较于仅使用数据库搜索的价值由 Greenhalgh and Peacock 测量（BMJ 331 (2005) 1064）。感谢 Semantic Scholar、OpenAlex、Crossref、arXiv、INSPIRE-HEP、PubMed 和 NASA ADS 背后的团队；正是这些团队的开放 API 让各项账本可被核验。