verify-citations

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Verify Citations

验证引用

The anti-hallucination gate for bibliographies. Takes a

.bib

file, validates every entry against live scholarly indexes (Crossref, DBLP, Semantic Scholar, arXiv, DataCite), and produces a per-entry verdict: VERIFIED, MISMATCH, UNRESOLVED (possible fabrication), or RETRACTED. One fabricated citation in a submitted paper can end a review — or a reputation. Be rigorous here.

Resolving is necessary but not sufficient. An identifier that resolves only proves the entry points at a real record — not that it is the right instance of the named work, nor that an independently-added reference is even on-topic. The gate therefore goes beyond resolution with two non-fabricating heuristics: a canonical-instance check (when a title exists as both, say, a conference paper and a later RFC/tech-report/preprint, surface the alternatives with citation counts so you cite the artifact the field cites) and an opt-in relevance gate (score an added reference's topical fit and flag low-fit ones for human review). And when an authoritative index is unreachable, the run reports PARTIAL-PASS, never a clean PASS, naming the checks that did not run — so "could not check" is never mistaken for "checked and clean".

文献目录的防幻觉校验工具。接收一个

.bib

文件，通过实时学术索引（Crossref、DBLP、Semantic Scholar、arXiv、DataCite）验证每一条目，并为每条目生成判定结果：VERIFIED（已验证）、MISMATCH（不匹配）、UNRESOLVED（无法解析，可能为伪造）或RETRACTED（已撤回）。提交的论文中只要有一条伪造引用，就可能导致评审失败，甚至损害个人声誉。请务必严谨对待。

解析成功是必要条件，但并非充分条件。标识符可解析仅能证明该条目指向某一真实记录，但无法证明它是对应作品的正确版本，也无法证明新增的参考文献与主题相关。因此，本工具在解析之外还采用了两种非伪造性启发式检查：标准版本校验（当同一标题同时存在会议论文、后续RFC/技术报告/预印本等版本时，展示不同版本及其引用量，以便引用该领域普遍认可的版本），以及可选的相关性校验（对新增参考文献的主题匹配度打分，标记匹配度低的条目供人工审核）。当权威索引无法访问时，工具会返回PARTIAL-PASS（部分通过），而非完全通过，并列出未执行的检查项——避免将“无法检查”误判为“检查通过且无问题”。

When to use

使用场景

The user asks to verify, check, validate, or audit citations / references / a bibliography / a
```
.bib
```
file.
Any writing skill (
```
draft-related-work
```
,
```
literature-review
```
,
```
write-rebuttal
```
, ...) added or edited citations — run this gate before declaring that work done.
Before
```
preflight-check
```
/ submission, camera-ready, or arXiv upload.
The user suspects AI-generated references ("are these real papers?").

用户要求验证、检查、校验或审核引用/参考文献/文献目录/.bib文件时
任何涉及编写引用的工具（
```
draft-related-work
```
、
```
literature-review
```
、
```
write-rebuttal
```
等）新增或编辑了引用——在完成该工具的工作前，需运行本校验工具
在
```
preflight-check
```
（预提交检查）、提交、终稿或arXiv上传之前
用户怀疑AI生成的参考文献（如“这些论文是真实存在的吗？”）

Inputs

输入项

A
```
.bib
```
file (path from the user, or find it:
```
*.bib
```
next to the main
```
.tex
```
, or the file named in
```
\bibliography{...}
```
/
```
\addbibresource{...}
```
).
```
CONTACT_EMAIL
```
environment variable — a real email, sent in the User-Agent so API providers can contact instead of block. Ask the user for it if unset; never invent one.
Optional:
```
S2_API_KEY
```
for a dedicated Semantic Scholar rate allowance (the shared anonymous pool 429s under load; the script backs off and falls back to other providers automatically).

一个
```
.bib
```
文件（用户提供路径，或自动查找：主
```
.tex
```
文件旁的
```
*.bib
```
文件，或
```
\bibliography{...}
```
/
```
\addbibresource{...}
```
中指定的文件）
```
CONTACT_EMAIL
```
环境变量——真实邮箱，会包含在User-Agent中，以便API提供商联系而非封禁。若未设置需向用户索要，切勿编造
可选：
```
S2_API_KEY
```
，用于Semantic Scholar的专属调用额度（共享匿名池在负载过高时会返回429错误；脚本会自动退避并 fallback 到其他提供商）

Process

处理流程

Run the checker. Deterministic work belongs to the script — do not verify entries by hand or from memory:
bash
```
export CONTACT_EMAIL=user@university.edu   # ask the user
python3 scripts/check_bibtex.py path/to/refs.bib --json /tmp/citecheck.json
```
Useful variants:
- ```
--offline
```
  — parse + duplicate/static checks only (no network; use when the user has no connectivity or only wants structural checks).
- ```
--key smith2024
```
  — re-check a single entry after a fix (repeatable).
- ```
--no-retraction-check
```
  — halves request count for very large files; keep retraction checks ON for any final pre-submission run.
- ```
--strict
```
  — exit nonzero on warnings and on PARTIAL-PASS (CI gate mode: a gate run that could not reach an index has not run).
- ```
--refresh
```
  — bypass the 24 h response cache in
```
.cache/
```
  .
- ```
--thesis-file thesis.txt --core-key dean2008 --core-key vaswani2017
```
  — turn on the relevance gate for independently-gathered additions: pass a plain-text thesis/abstract and the keys you have already confirmed are core; low-topical-fit entries get a
```
LOW_RELEVANCE
```
  flag for review. Use this when another skill (
```
draft-related-work
```
  ,
```
literature-review
```
  ) added references and you need to tell good additions from off-topic ones.
- ```
--no-canonical-instance
```
  — skip the extra same-title lookup that powers the
```
CANONICAL_INSTANCE
```
  wrong-artifact check (on by default).
- ```
--no-soft-fail
```
  — abort (exit 1) the instant any index is unreachable, instead of degrading to PARTIAL-PASS.
- ```
--self-test
```
  — run the bundled offline unit tests for the deterministic
```
ENTRY_TYPE_MISMATCH
```
  check (no network, no
```
.bib
```
  needed); exits 0 if they pass. Use this to confirm the script is intact after editing it.
Exit codes:
```
0
```
clean,
```
2
```
problems found,
```
1
```
operational failure (bad file, no network, missing CONTACT_EMAIL — fix the cause, do not skip the gate). For a 50-entry file expect ~2–4 minutes: the script is rate-limited to at most 1 request/second per host by design. Do not parallelize it and do not work around its politeness limits.

Triage every flag. Read references/triage-guide.md for what each flag means, known false positives (online-vs-print year off-by-one, venue aliasing, arXiv-vs-published versions, truncated author lists), and the exact fix for each. Severity at a glance:

Flag	Severity	Meaning
UNRESOLVED / DOI_NOT_FOUND / ARXIV_NOT_FOUND	ERROR	Not found in any index — possible fabrication
TITLE_MISMATCH / AUTHOR_MISMATCH / YEAR_MISMATCH	ERROR	Identifier points at a different paper, or metadata is wrong
RETRACTED	ERROR	A retraction notice exists for this DOI
DUPLICATE_KEY / DUPLICATE_DOI / DUPLICATE_TITLE	ERROR	Same paper or key twice
MALFORMED_DOI / MALFORMED_ARXIV_ID	ERROR	Identifier cannot be valid
POSSIBLE_ID_TYPO / MISSING_DOI / VENUE_MISMATCH / AUTHOR_LIST_DIFFERS / TITLE_PARTIAL_MATCH / NOT_IN_INDEXES / EXPRESSION_OF_CONCERN	WARN	Real paper, imperfect entry — fix or justify
ENTRY_TYPE_MISMATCH	WARN	BibTeX type contradicts the resolved record (e.g. a journal article or monograph typed `@inproceedings` , a `booktitle` naming a journal/publisher, or `@inproceedings` carrying `journal=` ) — set the type from the canonical `type` field
CANONICAL_INSTANCE	WARN	Resolves, but a different artifact of the same work is what the field cites — pick the canonical instance
LOW_RELEVANCE	WARN	Resolves, but scored low topical fit to the paper — confirm it is load-bearing, never auto-remove
UNVERIFIABLE_TYPE / RESOLVED_VIA_SEARCH / HAS_CORRECTION / RELEVANCE_OK / CHECK_SKIPPED	INFO	Context for manual judgment

Fix only from canonical records. Replace broken entries with BibTeX fetched from the authoritative source (DBLP
```
.bib
```
endpoint or doi.org content negotiation — exact commands in references/triage-guide.md). Never retype metadata from memory; that is how hallucinations get laundered into "fixes". Provider details, what each index is authoritative for, and the venue-alias problem are in references/verification-sources.md.
Escalate what cannot be fixed. For each UNRESOLVED entry, present the evidence to the user and ask: keep (with a manual source they vouch for), fix (they supply the real reference), or remove (also remove the
```
\cite
```
and adjust surrounding text). Never decide silently, never delete silently, and never substitute a different paper that merely sounds similar.
For
```
CANONICAL_INSTANCE
```
and
```
LOW_RELEVANCE
```
(both WARN), see references/triage-guide.md and references/relevance-gate.md. These are copilot prompts, not autopilot actions: surface the alternative artifact / the low-fit score and let the user decide. The script's lexical relevance score is a deterministic proxy — if you have an embedding model available, compute abstract-embedding similarity yourself for a stronger signal before advising, as the relevance-gate reference explains. Never auto-swap an instance or auto-delete a low-fit reference.
Re-run until exit code 0 (or until remaining flags are explicitly accepted by the user). Re-check just the fixed entries with
```
--key
```
, then do one full final pass. If a run came back PARTIAL-PASS because an index was unreachable, that is not done — re-run the skipped checks once connectivity returns before declaring the gate clean.
Report. The script prints a canonical
```
VERDICT-LINE:
```
(verdict + raw counts — N verified, errors, warnings, K skipped — plus the WARN breakdown by flag) and writes the same string to the JSON as
```
verdict_line
```
. Copy that line verbatim into your summary and into any README or status line you write — do not recompose it from memory and do not round it off. Around it, add: the list of fixes applied (old → new), user decisions taken, flags the user accepted as-is, and any wrong-artifact / low-relevance items raised for the user's judgment. If this run gates another skill's output, pass the same
```
verdict_line
```
through unchanged; never upgrade a PARTIAL-PASS to "passed", never collapse "PARTIAL-PASS, 11 WARN, 1 skipped" into "verified, 0 errors", and never promise acceptance or a clean review.

运行校验器。确定性工作由脚本完成——切勿手动或凭记忆验证条目：
bash
```
export CONTACT_EMAIL=user@university.edu   # 向用户索要
python3 scripts/check_bibtex.py path/to/refs.bib --json /tmp/citecheck.json
```
实用变体：
- ```
--offline
```
  ——仅执行解析+重复/静态检查（无需网络；适用于用户无网络连接或仅需结构检查的场景）
- ```
--key smith2024
```
  ——修正后重新检查单个条目（可重复执行）
- ```
--no-retraction-check
```
  ——大幅减少大型文件的请求量；但在最终预提交运行时请保持撤回检查开启
- ```
--strict
```
  ——若出现警告或PARTIAL-PASS则返回非零退出码（CI校验模式：若无法访问索引，则校验未完成）
- ```
--refresh
```
  ——绕过
```
.cache/
```
  中24小时的响应缓存
- ```
--thesis-file thesis.txt --core-key dean2008 --core-key vaswani2017
```
  ——为独立收集的新增引用开启相关性校验：传入纯文本论文/摘要，以及已确认的核心引用键；主题匹配度低的条目会被标记为
```
LOW_RELEVANCE
```
  供审核。适用于其他工具（
```
draft-related-work
```
  、
```
literature-review
```
  ）新增了引用，需要区分有效新增和偏离主题的引用时
- ```
--no-canonical-instance
```
  ——跳过用于
```
CANONICAL_INSTANCE
```
  （错误版本）检查的同名标题额外查找（默认开启）
- ```
--no-soft-fail
```
  ——一旦任何索引无法访问则立即终止（退出码1），而非降级为PARTIAL-PASS
- ```
--self-test
```
  ——运行内置的离线单元测试，用于确定性的
```
ENTRY_TYPE_MISMATCH
```
  检查（无需网络和
```
.bib
```
  文件）；测试通过则返回0。用于确认脚本在编辑后仍完整可用
退出码：
```
0
```
表示无问题，
```
2
```
表示发现问题，
```
1
```
表示操作失败（文件错误、无网络、缺少CONTACT_EMAIL——需修复问题，切勿跳过校验）。对于包含50条目的文件，预计耗时约2-4分钟：脚本设计为每个主机最多每秒1次请求，以遵守速率限制。请勿并行运行或绕过该礼貌限制。

分类处理所有标记。阅读references/triage-guide.md了解每个标记的含义、已知误报（线上与印刷版年份差1、出版地别名、arXiv与正式出版版本、作者列表截断）以及每个问题的具体修正方法。严重程度概览：

标记	严重程度	含义
UNRESOLVED / DOI_NOT_FOUND / ARXIV_NOT_FOUND	错误	在任何索引中均未找到——可能为伪造
TITLE_MISMATCH / AUTHOR_MISMATCH / YEAR_MISMATCH	错误	标识符指向另一篇论文，或元数据错误
RETRACTED	错误	该DOI存在撤回通知
DUPLICATE_KEY / DUPLICATE_DOI / DUPLICATE_TITLE	错误	同一论文或引用键重复出现
MALFORMED_DOI / MALFORMED_ARXIV_ID	错误	标识符格式无效
POSSIBLE_ID_TYPO / MISSING_DOI / VENUE_MISMATCH / AUTHOR_LIST_DIFFERS / TITLE_PARTIAL_MATCH / NOT_IN_INDEXES / EXPRESSION_OF_CONCERN	警告	论文真实存在，但条目存在瑕疵——需修正或说明理由
ENTRY_TYPE_MISMATCH	警告	BibTeX类型与解析记录不符（例如：将期刊文章或专著标记为 `@inproceedings` ， `booktitle` 字段填写期刊/出版社名称，或 `@inproceedings` 包含 `journal=` 字段）——需根据标准 `type` 字段设置类型
CANONICAL_INSTANCE	警告	可解析，但该作品的另一版本才是领域内普遍引用的——需选择标准版本
LOW_RELEVANCE	警告	可解析，但与论文主题匹配度低——需确认其是否必要，切勿自动删除
UNVERIFIABLE_TYPE / RESOLVED_VIA_SEARCH / HAS_CORRECTION / RELEVANCE_OK / CHECK_SKIPPED	信息	供人工判断的上下文信息

仅基于标准记录修正。使用权威来源获取的BibTeX替换有问题的条目（DBLP的.bib端点或doi.org的内容协商——具体命令见references/triage-guide.md）。切勿凭记忆重新输入元数据；这正是幻觉被洗白为“修正”的方式。关于提供商详情、各索引的权威领域以及出版地别名问题，请参阅references/verification-sources.md。
无法修正的问题需提交用户决策。对于每个UNRESOLVED条目，向用户展示证据并询问：保留（需用户提供可靠的手动来源）、修正（用户提供真实参考文献）或删除（同时删除
```
\cite
```
命令并调整上下文文本）。切勿静默决策、静默删除，也切勿替换为仅名称相似的其他论文。
对于
```
CANONICAL_INSTANCE
```
和
```
LOW_RELEVANCE
```
（均为警告），请参阅references/triage-guide.md和references/relevance-gate.md。这些是辅助提示，而非自动操作：需展示替代版本/低匹配度分数，由用户决策。脚本的词汇相关性分数是确定性代理——若有嵌入模型可用，可自行计算摘要嵌入相似度以获得更可靠的信号，再给出建议，具体见相关性校验的参考文档。切勿自动替换版本或自动删除低匹配度引用。
重新运行直至退出码为0（或直至剩余标记被用户明确接受）。使用
```
--key
```
仅重新检查已修正的条目，然后进行一次完整的最终检查。若因索引无法访问导致返回PARTIAL-PASS，则校验未完成——待网络恢复后重新运行未执行的检查，再确认校验通过。
生成报告。脚本会在标准输出中打印一条标准的
```
VERDICT-LINE:
```
（判定结果+原始统计数据——已验证数量、错误数量、警告数量、跳过数量——以及按标记分类的警告明细），并将同一字符串写入JSON文件的
```
verdict_line
```
字段。请将该行原文复制到你的总结、README或状态行中——切勿凭记忆重新编写或四舍五入。在该行周围补充：已应用的修正列表（旧→新）、用户做出的决策、用户接受的标记，以及提交给用户判断的错误版本/低相关性条目。若本次校验为其他工具的输出把关，请原样传递
```
verdict_line
```
；切勿将PARTIAL-PASS升级为“通过”，切勿将“PARTIAL-PASS，11条警告，1项跳过”简化为“已验证，0错误”，切勿承诺会被接受或获得干净的评审结果。

Output

输出结果

The script's per-entry report on stdout, ending with a canonical
```
VERDICT-LINE:
```
you reproduce verbatim, and — with
```
--json
```
— a machine-readable report (per-entry statuses and flags, resolved DOIs/URLs, the same
```
verdict_line
```
, and a
```
summary.warnings_by_flag
```
count) the calling skill can act on.
A corrected
```
.bib
```
(edits applied from canonical records, with the user's approval) and a short human summary whose first line is the verbatim
```
VERDICT-LINE
```
, followed by what changed and what remains open.

脚本在标准输出中生成的条目级报告，结尾为需原样复制的标准
```
VERDICT-LINE:
```
；若使用
```
--json
```
参数，还会生成机器可读的报告（条目级状态和标记、解析后的DOI/URL、相同的
```
verdict_line
```
，以及
```
summary.warnings_by_flag
```
统计），供调用工具使用
修正后的
```
.bib
```
文件（基于标准记录进行编辑，需获得用户批准），以及一份简短的人工总结，第一行为原文复制的
```
VERDICT-LINE:
```
，后续内容为变更点和未解决问题

Hard rules

硬性规则

Never fabricate a citation, DOI, arXiv ID, or BibTeX field. If a reference cannot be verified, say so — an honest gap beats a confident fake.
Never "fix" an unresolved entry by guessing which real paper was meant. Search, show candidates, let the user choose.
Resolution is not endorsement. A resolving DOI/ID proves the record is real, not that it is the right instance or topically relevant. Treat
```
CANONICAL_INSTANCE
```
and
```
LOW_RELEVANCE
```
as copilot prompts: surface the evidence, never auto-swap the artifact or auto-remove the reference.
Never invent a citation count, an alternative artifact, or a relevance score. Every alternative the canonical-instance check shows must be a record an index actually returned; every relevance number comes from the script or from an embedding model you actually ran — not from memory.
A PARTIAL-PASS is not a PASS. If any authoritative index was unreachable, report PARTIAL-PASS, list the skipped checks, and re-run them before treating the bibliography as verified. Never let "could not check" read as "clean".
The human summary must echo the machine verdict verbatim — never paraphrase it cleaner. Any prose summary, README, or status line you write must carry the script's exact verdict (PASS / PARTIAL-PASS / FAIL) plus the raw counts: N verified, M warnings (by flag), K skipped checks. Do not collapse "PARTIAL-PASS, 11 WARN, 1 skipped check" into "verified, 0 errors"; a clean-looking summary that diverges from the report it summarizes is itself a failure. An entry is only "verified" once an actual index round-trip confirmed it this session and reconciled its type/venue/DOI to the canonical (not reprint) record — never mark verified from a stale or skipped check, and never leave a "not yet trustworthy" banner on a file you call verified.
A RETRACTED result must be surfaced to the user verbatim, with the retraction-notice DOI. Citing retracted work knowingly is sometimes legitimate (e.g., studying retractions) — that is the user's call, and the citation should then mark the retraction explicitly.
Metadata only: this skill fetches and compares titles, authors, years, venues, DOIs. Do not store fetched abstracts or paper text in the repo.
Retraction coverage is best-effort (Crossref/Retraction Watch data plus title markers); absence of a flag is not proof a paper stands. Say so when it matters.
When using
```
venues/
```
profiles to judge a VENUE_MISMATCH, treat the profile as a starting point — re-verify any venue fact you rely on against the live
```
cfp_url
```
in the profile before telling the user their entry is wrong.

切勿伪造引用、DOI、arXiv ID或BibTeX字段。若参考文献无法验证，请如实告知——诚实的空白比自信的伪造更可取
切勿通过猜测“修正”无法解析的条目。需进行搜索，展示候选结果，由用户选择
解析成功不代表认可。可解析的DOI/ID仅能证明记录真实存在，无法证明其是正确版本或与主题相关。将
```
CANONICAL_INSTANCE
```
和
```
LOW_RELEVANCE
```
视为辅助提示：展示证据，切勿自动替换版本或自动删除引用
切勿编造引用量、替代版本或相关性分数。标准版本检查展示的每个替代版本必须是索引实际返回的记录；每个相关性分数必须来自脚本或你实际运行的嵌入模型——而非记忆
PARTIAL-PASS不等于通过。若任何权威索引无法访问，需报告PARTIAL-PASS，列出跳过的检查项，并在将文献目录标记为已验证前重新运行这些检查。切勿让“无法检查”被解读为“无问题”
人工总结必须与机器判定原文一致——切勿简化表述。你编写的任何文字总结、README或状态行必须包含脚本的准确判定结果（PASS/PARTIAL-PASS/FAIL）以及原始统计数据：已验证数量、警告数量（按标记分类）、跳过的检查数量。切勿将“PARTIAL-PASS，11条警告，1项跳过检查”简化为“已验证，0错误”；与报告内容不符的简洁总结本身就是失败。只有当本次会话中实际通过索引往返确认，并将其类型/出版地/DOI与标准（非重印）记录一致后，条目才可标记为“已验证”——切勿基于过期或跳过的检查标记已验证，也切勿在标记为已验证的文件上留下“尚未可信”的提示
RETRACTED结果必须原样展示给用户，并附上撤回通知的DOI。明知故犯地引用已撤回的文献有时是合理的（例如：研究撤回案例）——这由用户决定，且引用需明确标记撤回信息
仅处理元数据：本工具仅获取和比较标题、作者、年份、出版地、DOI。请勿在仓库中存储获取的摘要或论文文本
撤回信息覆盖为尽力而为（Crossref/Retraction Watch数据加标题标记）；无标记不代表论文有效。必要时需告知用户这一点
使用
```
venues/
```
配置文件判断VENUE_MISMATCH时，需将配置文件作为起点——在告知用户其条目错误前，需根据配置文件中的实时
```
cfp_url
```
重新验证你依赖的任何出版地信息

Bundled resources

附带资源

```
scripts/check_bibtex.py
```
— the verifier. Run it; do not reimplement it.
references/triage-guide.md — flag-by-flag interpretation, false positives, exact remediation commands (includes
```
CANONICAL_INSTANCE
```
and the PARTIAL-PASS verdict).
references/relevance-gate.md — how the relevance gate scores topical fit, the embedding-similarity upgrade you should run when a model is available, and how to act on
```
LOW_RELEVANCE
```
.
references/verification-sources.md — provider APIs, authority order, venue aliasing, retraction data, rate limits and licensing.

```
scripts/check_bibtex.py
```
——校验器。请直接运行，切勿重新实现
references/triage-guide.md ——逐条标记的解释、误报说明、具体修复命令（包含CANONICAL_INSTANCE和PARTIAL-PASS判定）
references/relevance-gate.md ——相关性校验如何计算主题匹配度、当有模型可用时应运行的嵌入相似度升级方法，以及如何处理
```
LOW_RELEVANCE
```
标记
references/verification-sources.md ——提供商API、权威顺序、出版地别名、撤回数据、速率限制和许可信息

Memory

记忆机制

This skill uses the shared

.paper-memory/

convention in the user's paper directory, following

paper-memory-convention.md

At start: read
```
.paper-memory/lessons.md
```
to skip re-flagging entries the user already resolved this cycle, and lead with any
```
recurring
```
citation habits recorded for this author (e.g. "tends to cite arXiv preprints that are now published; prefer the published record").
At end: append durable findings in the shared format
```
- [YYYY-MM-DD] (verify-citations | <scope>) issue -> recommendation
```
(via
```
reflect-and-improve
```
's
```
reflect_log.py append
```
, which dedupes and dates). A pattern across the bibliography or across papers is
```
recurring
```
; a single fixed entry is
```
this-paper
```
. Never record fabricated metadata in memory, only the pattern and the canonical fix.
Create
```
.paper-memory/
```
on demand if absent and offer to add it to the project
```
.gitignore
```
. It is local-only; never upload it or copy it into this repo.

本工具遵循用户论文目录中的共享

.paper-memory/

约定，详见

paper-memory-convention.md

。

启动时：读取
```
.paper-memory/lessons.md
```
，跳过重新标记用户本次周期内已解决的条目，并优先处理为该作者记录的
```
recurring
```
（重复出现的）引用习惯（例如：“倾向于引用已正式出版的arXiv预印本；优先选择正式出版记录”）
结束时：以共享格式
```
- [YYYY-MM-DD] (verify-citations | <scope>) issue -> recommendation
```
追加持久化发现（通过
```
reflect-and-improve
```
的
```
reflect_log.py append
```
实现，会自动去重并添加日期）。若在文献目录或多篇论文中发现模式，则标记为
```
recurring
```
；单个已修正的条目标记为
```
this-paper
```
。切勿在记忆中存储伪造的元数据，仅记录模式和标准修正方法
若
```
.paper-memory/
```
不存在则按需创建，并建议用户将其添加到项目的
```
.gitignore
```
中。该目录仅本地存储；切勿上传或复制到本仓库