translation-quality-defect-families

译文质量问题族 / Translation Quality Defect Families

Boundary / 边界

Use this skill to turn recurring translation-quality problems into reusable correction knowledge. 使用本 skill，把反复出现或可能系统性复现的译文质量问题，沉淀成可复用的修正规则。

This skill focuses on the translation as a readable book: fidelity, literary force, target-language fluency, terminology, narrative rhythm, reader engagement, and source-supported expression. 本 skill 关注“译文作为一本书”的质量：忠实度、文学性、目标语顺读、术语、叙事节奏、读者吸引力，以及所有有原文依据的表达转换。

Do not use this skill for format-only issues such as EPUB manifest errors, cover policy, release metadata, local-path leaks, or CI failures unless they also expose a translation-quality pattern. 不要把本 skill 用于纯格式问题，例如 EPUB manifest、封面 policy、release metadata、本机路径泄漏或 CI 失败；除非这些问题同时暴露了译文质量模式。

Required Workflow / 必须流程

When a review finds a defect that may recur, do the following before declaring the work closed. 评审发现可能复现的问题后，必须先完成以下步骤，才能宣布关闭。

1. Classify the defect family. 先归纳问题族。
2. Record how it was found: chapter control, random sample, human note, rg scan, glossary scan, source comparison, generated XHTML inspection, or reader feedback. 记录发现方式：章节控制、随机样本、人工意见、rg 扫描、术语扫描、原文对照、生成 XHTML 检查或读者反馈。
3. Define the risk: what reader misunderstanding, fidelity loss, rhythm damage, terminology drift, or unauthorized invention it causes. 定义风险：它会造成什么误读、失真、节奏损伤、术语漂移或擅自加戏。
4. Audit similar cases with the cheapest reliable method first: glossary fields, rg, title maps, source-to-target term lists, chapter controls, or targeted small-context agent review. 先用最低 token 且可靠的方法查同类：术语表字段、rg、标题映射、源语-目标语词表、章节控制，或小上下文 agent 专项复核。
5. Fix confirmed matches and record justified exceptions. 修复确认命中，并记录合理例外。
6. Add or merge a reusable lesson in the register below. 在下方注册表追加或合并可复用经验。 This is mandatory in the same round as the fix: do not postpone skill backfill to a final summary, because later agents will otherwise rediscover the same family. 必须在同一轮修复中完成 skill 回填：不要拖到最终总结，否则后续 agent 会反复重新发现同一问题族。
7. Recheck with the appropriate gate: full-chapter recheck for chapter-stage defects, new-seed stratified random sample for post-EPUB defects, or a targeted family audit for book-wide risks. 用合适门禁复查：章节阶段问题做整章复查，EPUB 后抽检问题用新 seed 复抽，全书风险做问题族专项审计。

Update Discipline / 更新纪律

Do not append repetitive notes blindly. 不要盲目重复追加。

Before adding a new entry, search this file for the family, source-language mechanism, and target-language symptom. If an entry already exists, update that entry with a sharper detection method, better fix pattern, or a concrete example. 新增条目前，先搜索本文件中的问题族、源语机制和目标语症状。已有同族条目时，应合并更新：补更准的发现方法、更好的修复模式或具体例子。

Every entry should stay reusable across books. Book-specific names may appear as examples, but the rule must be phrased as a general lesson. 每个条目都应能跨书复用。可以用具体书名或人名做例子，但规则必须写成通用经验。

Use bilingual Chinese-English wording for new major sections. Chinese may lead when the expected contributor is Chinese. 新增主要段落使用中英文双语。面向中文贡献者时，中文可置前。

Validator Fields / 校验字段

For random spot-check rounds, book-local evidence belongs in reviews/random_spotcheck/round_XXX/fixes/fix_log.md and verification/closure_check.md, but reusable translation-quality lessons must be merged here in the same round. The pass validator checks this through machine-readable fields. 随机抽检轮次中，书内证据写入 reviews/random_spotcheck/round_XXX/fixes/fix_log.md 和 verification/closure_check.md，但可复用译文质量经验必须在同一轮合并回填到本 skill。强校验会通过机器可读字段检查这件事。

When a round found translation-quality defect families, set: 发现译文质量问题族时，必须填写：

defect_family_count: {total_family_count}
translation_quality_defect_family_count: {quality_family_count}
translation_quality_skill_backfill: "UPDATED" # or "MERGED"
translation_quality_skill_backfill_path: "skills/translation-quality-defect-families/SKILL.md"
translation_quality_skill_backfill_summary: "{what was added or merged: finding method, classification, low-token audit, fix pattern, recheck}"

And in closure_check.md: 并在 closure_check.md 写入：

translation_quality_skill_backfill_verified: true

If the round found only format, asset, path, EPUB structure, or other non-translation-quality issues, set translation_quality_defect_family_count: 0, translation_quality_skill_backfill: "NOT_APPLICABLE", and write a concrete translation_quality_skill_backfill_not_applicable_reason. 如果该轮只有格式、资产、路径、EPUB 结构或其他非译文质量问题，应写 translation_quality_defect_family_count: 0、translation_quality_skill_backfill: "NOT_APPLICABLE"，并填写具体的 translation_quality_skill_backfill_not_applicable_reason。

Low-Token Audit Pattern / 低 token 审计模式

• Start with machine-readable evidence: glossary/terms.csv, forbidden renderings, title maps, chapter controls, sample manifests, and review tables. 先用机器可读证据：glossary/terms.csv、禁用正文写法、标题映射、章节控制、抽样 manifest 和评审表。
• Use rg to collect candidates before asking an agent to read prose. 先用 rg 收集候选，再让 agent 阅读片段。
• Give the agent only the candidate passage, nearby source text, relevant glossary rows, and the current family rule. 只给 agent 候选段、邻近原文、相关术语行和当前问题族规则。
• Treat random sampling as discovery, not as the main polishing engine. 随机抽检用于发现盲点，不是主要润色引擎。
• Treat full-chapter review as the main place to catch fidelity, fluency, rhythm, and terminology issues. 忠实度、顺读、节奏和术语问题，主要应在每章全量复检阶段解决。

Defect Family Register / 问题族注册表

High-Impact Literal and Explanatory Opening Prose / 高影响开篇直译腔与解释腔

• Symptom: Author/source prefaces, introductions, openings, first chapters, or other reader-entry passages are accurate but sound stiff, flat, awkward, overly literal, or explanatory in the target language. The prose may use causal bridges such as 原因在于, conditional chains such as 要是只能...就必然, abstract noun stacks, or target-only analytical wording that explains the source instead of becoming a natural book sentence. 症状：原作者/原书序言、导言、开篇、首章或其他读者入口段落意思准确，但目标语读起来不顺口、怪、平、硬、直译感重或解释腔明显。常见表现包括 原因在于 一类因果桥、要是只能...就必然 一类条件链、抽象名词堆叠，或把原意说明出来而不是转成自然书面句。
• Find by: Treat opening files and headings matching preface|introduction|序言|前言|导言|开篇|首章|第一章 as high impact before random sampling. Use scripts/check_literary_style_gate.py --write-report plus targeted scans such as 原因在于|也就不能|就必然|换言之|也就是说|这说明|进行|具有|予以|加以|表达 意思是|原文意思是. These are candidates, not automatic defects; inspect small source-target contexts and prioritize the first pages of the book. 发现方式：在随机抽检前即把文件名/标题匹配 preface|introduction|序言|前言|导言|开篇|首章|第一章 的章节视为高影响区域。运行 scripts/check_literary_style_gate.py --write-report，并用 原因在于|也就不能|就必然|换言之|也就是说|这说明|进行|具有|予以|加以|表达 意思是|原文意思是 等扫描候选。这些只是候选，不是自动错误；必须小上下文对照源文，优先检查全书最前几页。
• Risk: Readers judge the translation before they reach the middle of the book. A weak opening can lose trust even when later chapters are accurate. Treat repeated opening style debt as P2/blocking for final release, not as a harmless P3 comment. 风险：读者会在进入正文前判断译本是否可信。开篇弱，即使后文准确，也会让读者流失。反复出现的开篇文学债务在最终发布前应按 P2/阻塞处理，不应作为无害 P3 建议。
• Fix by: First perform target-language-only rewriting: remove explanatory scaffolding, break source syntax, rebuild cadence, and choose idiomatic target-language verbs and clauses. Then compare against the source to ensure the polish did not add argument, motivation, or tone. For Gibbon-like argumentative prose, preserve balance and antithesis without turning the sentence into modern analysis; for biography or dialogue, preserve voice and character pressure instead of smoothing everything into commentary. 修复方式：先只看目标语重写，删去解释脚手架，打散源语句法，重建节奏，选择自然目标语动词和分句；再回到源文校准，确认润色没有新增论点、动机或语气。吉本式论辩散文要保留均衡和对照，不要改成现代分析说明；传记或对话要保留人物声口和叙述压力，不要全部抹成评论腔。
• Recheck: Final release requires qa/literary_style/literary_style_review.md PASS with target_only_reading_score: 5, read_aloud_awkward_sentence_count: 0, literal_explanatory_style_debt_count: 0, and high-impact sections reviewed. Random review agents must set high_impact_style_debt_count: 0 and style_debt_count: 0 for excellence PASS. 复查：最终发布必须有 qa/literary_style/literary_style_review.md PASS，并记录 target_only_reading_score: 5、read_aloud_awkward_sentence_count: 0、literal_explanatory_style_debt_count: 0，且高影响开篇已复审。随机评审 Agent 若要给优秀 PASS，必须写明 high_impact_style_debt_count: 0 和 style_debt_count: 0。

Reader-Facing Note and Quote Interface / 读者可见注号与外文引文接口

• Symptom: Notes or citation units expose source-side mechanics instead of a reader-stable interface: bare split markers such as 注18-1, raw bibliography such as Ludewig in Vit. Justiniani, unexplained foreign phrases such as Cqamalh edra or Dux... sedentem Imperatorem adorat, or long source quotations after an already adequate Chinese gist. 症状：注释或引文单元把源文侧机制直接暴露给读者，例如裸露的拆分注号 注18-1、未中文化的书目信息 Ludewig in Vit. Justiniani、未解释的外文短语 Cqamalh edra / Dux... sedentem Imperatorem adorat，或已经有中文大意后仍保留过长原文。
• Find by: Start with exact scans before rereading: rg -n "注[0-9]+-[0-9]+|Ludewig in Vit|Hist\\. de l'Acad|Cqamalh|Dux.*sedentem|Et ego, Deo" over chapters/final, chapters/translated, and generated XHTML. Then inspect only matched units against nearby source notes. 发现方式：先低 token 精确扫描，再读小上下文；可扫 注[0-9]+-[0-9]+|Ludewig in Vit|Hist\\. de l'Acad|Cqamalh|Dux.*sedentem|Et ego, Deo，范围限于 chapters/final、chapters/translated 和生成 XHTML。随后只复核命中单元及邻近源注。
• Risk: A note may be technically linked but still feel broken to readers. Split numbers look like corrupted references, raw bibliography or Latin/French/Greek strings interrupt the Chinese book surface, and long original quotes can overload a reader-facing audit unit. 风险：注释链接即使技术上存在，也可能在读者侧像损坏编号；未处理的书目、拉丁文、法文或希腊文会打断中文书面界面；过长原文会让单个 audit unit 失去可读性。
• Fix by: Make the interface explicit and stable. For split editor notes, use a reader label such as 补注18.1 consistently in both body references and note definitions. For foreign quotes, add immediate Chinese meaning in the same reader-facing unit, or shorten the original to an opening phrase plus bibliographic source when the Chinese gist already carries the meaning. Translate bibliographic titles enough for a Chinese reader, while preserving source identifiers only as supporting evidence. 修复方式：把接口改成明确、稳定的读者标签。拆分编注可统一写成 补注18.1，正文引用和注释定义必须同步。外文引文应在同一读者单元内立刻给出中文意思；若中文大意已经承载意义，长原文可缩为起首短语加出处。书目标题至少应中文化到读者能理解，源文标识只作为辅助证据保留。
• Production-placeholder guard: Phrases such as 外文原句略, 原文大意已译入上文, 中文接口, or bare scaffolding like 大意是 are private repair markers, not publishable prose. Scan chapters/final, chapters/translated, and generated XHTML for exact markers; replace each confirmed hit with concrete Chinese rendering and citation, or delete the orphan placeholder when neighboring translated prose already carries the source meaning.
• Finished-note interface guard: Phrases such as 由某某这样表达 意思是 or 原文意思是 often signal a draft bridge between source quotation and Chinese note prose. Scan exact sampled fragments plus 表达 意思是|原文意思是; convert confirmed hits to finished citation prose such as 某某这样概括... or 某某接着说..., while preserving acceptable exceptions where 拉丁原文意思是 directly introduces a necessary translation of a long source quotation. 制作占位语防护：外文原句略、原文大意已译入上文、中文接口，以及裸露的 大意是 脚手架，都只能作为私下修复标记，不得进入发布文本。应在 chapters/final、chapters/translated 和生成 XHTML 中精确扫描；确认命中后改成具体中文译意和出处，若邻近译文已承载源文意义，则删除孤立占位语。
• Recheck: Scan for the old marker or raw phrase, compare final and translated files for exact parity, then run random artifact validation. A repair round that changes note-interface conventions does not count as a clean random PASS; it needs a fresh seed. 复查：扫描旧标记和原触发短语，确认 final / translated 完全同步，再运行随机抽检 artifact 校验。凡修改注释接口约定的返工轮不计入 clean random PASS，必须新 seed 复抽。
• Citation-boundary and truncation guard: Finished notes can still fail through tiny boundaries: a citation glued to following prose (第十一章但), a sentence ending after an orphan comma (一个家庭的不幸，), a dangling close quote before a citation (（注57）”阿米阿努斯), a note ending after a connector (又$), or a Latin evidence phrase left without an immediate gloss (gloriæ salutis suique oblitus). Scan exact sampled strings plus patterns such as 第[一二三四五六七八九十百0-9]+章[但而这], ，$ or 又$ in sampled note units, and high-risk Latin phrases; repair with punctuation, complete the source sentence, and add a short Chinese gloss in the same reader-facing unit. 引文边界与截断防护：完稿注释仍可能因极小边界失败，例如书目和正文粘连（第十一章但）、句子停在孤立逗号（一个家庭的不幸，），或拉丁证据短语没有紧邻中文释义（gloriæ salutis suique oblitus）。应扫描抽样原串及 第[一二三四五六七八九十百0-9]+章[但而这]、抽样注释单元内行尾 ，、高风险拉丁短语；用标点断开、补全源句，并在同一读者单元给出简短中文释义。
• Bibliography and variant-word interface guard: Retained authority names, dates, or foreign variant words must still have a Chinese interface when they carry the sampled unit's meaning, for example Protonotarius Interpretum, Chandler's Travels into Asia Minor, ennenhkonta/ebdomhkonta, or legal phrases such as Reddimus ad præsens. Scan exact sampled strings plus A\\.D\\.|Travels into|Interpretum|[a-zæû]+mus\\b|ennenhkonta|ebdomhkonta; translate titles and offices, add compact glosses for variant words, and keep source forms in parentheses only as evidence. 书目与异文接口防护：保留权威姓名、年代或外文异文时，若它们承担抽样单元意义，仍必须给出中文接口，例如 Protonotarius Interpretum、Chandler's Travels into Asia Minor、ennenhkonta/ebdomhkonta，或 Reddimus ad præsens 这类法律拉丁短语。可扫抽样原串及 A\\.D\\.|Travels into|Interpretum|[a-zæû]+mus\\b|ennenhkonta|ebdomhkonta；书名和职衔应中文化，异文词给紧凑释义，源文形式只作括号证据保留。

Unsupported Note Commentary / 无源依据的注释解释性加戏

• Symptom: A translated note adds reader-facing commentary about what a note “continues,” “intends to show,” or “means for the argument,” even though the source note only states evidence or citation. 症状：译文注释额外加入“这条注释延续正文”“意在说明”“换言之”等解释性评语，而源注只提供证据或书目。

• Find by: Use exact phrase scans such as 这条.*注释|延续正文|意在说明|这段.*说明|换言之|也就是说, then compare the matched note with the source. Treat these as candidates, not automatic errors, because some source notes genuinely paraphrase their own point. 发现方式：先扫 这条.*注释|延续正文|意在说明|这段.*说明|换言之|也就是说，再对照源注。它们只是候选，不是自动错误，因为源注本身有时确会复述论点。

• Risk: The translation stops being a note translation and becomes an unmarked translator gloss, which can shift tone, fidelity, and reader trust. 风险：译文从“翻译注释”滑成未标明的译者讲解，改变语气、忠实度和读者信任。

• Fix by: Remove unsupported commentary, or mark it as an authorized translator note only if the project policy explicitly permits such notes. Preserve the source note's evidence chain without adding interpretive bridges. 修复方式：删除无源依据的解释；只有项目政策明确允许译者注时，才可另行标注。保留源注证据链，不补写解释性桥段。

• Round-derived extension: Unsupported commentary can appear as whole inserted recap paragraphs, not only as one extra note sentence. If a sampled unit reads like a modern analytical bridge, compare paragraph-by-paragraph with chapters/src; delete target-only recap paragraphs unless they are explicitly authorized translator notes.

• Historical synthesis guard: In long historical chapters, target-only recap paragraphs can sound plausible because they summarize nearby source-backed material, for example framing cities or dynastic divisions as modern analytical patterns. Treat triggers such as 代表了...另一种, 征服史之外, 因此，这场..., or 这种忠诚更多依附于... as audit candidates; delete only confirmed target-only synthesis and keep source-backed historical judgment. 轮次补充：无源解释不只会是注释里的一句译者评语，也可能是整个插入的现代分析总结段。若抽样单元读起来像现代分析桥段，应按段落对照 chapters/src；没有源文对应的 target-only recap 段落应删除，除非项目明确授权译者注。

• Low-token analytical-recap audit: After one sampled target-only recap is confirmed, scan the whole book for modern interpretive trigger phrases before asking for broad reread. Useful zh-Hans candidates include 制度自救, 罗马记忆, 主权归属, 战略位置, 心理优势, 不能替代.*证据, 说明.*而不是, and 显示.*可保存. Treat them as candidates, not automatic defects: delete only when nearby source lacks the analytical conclusion, and record source-supported military, legal, or historiographical statements as exceptions. 低 token 分析总结审计：抽样确认一处 target-only 总结后，先用现代解释触发词全书扫描，再请人复核候选。简体中文候选包括 制度自救、罗马记忆、主权归属、战略位置、心理优势、不能替代.*证据、说明.*而不是、显示.*可保存。这些只作为候选，不是自动缺陷；只有邻近原文没有相应分析结论时才删除，并把有源文支撑的军事、法律或史学判断记为例外。

• Quantitative evidence compression guard: If a source note gives concrete counts, dates, page references, or named authorities, do not replace it with a broad explanatory gloss such as "mortality tables help explain the point." Low character-ratio chapters should trigger targeted scans of their shortest or most summarized notes; preserve the exact numbers and citation chain unless a project-specific note policy explicitly permits abridgement. 数字证据压缩防护：若源注给出具体人数、年代、页码或具名出处，不得改写成“死亡率表可帮助理解”等泛泛解释。字符覆盖率偏低的章节应追加扫描最短、最概括的注释；除非本书另有明确删节政策，否则必须保留具体数字和出处链。

• Plausible-bridge trigger extension: Unsupported bridges may be a single polished sentence inside an otherwise source-backed paragraph, for example 这种答复把...并列起来, 不能靠本地土壤...把...消息吸入城中, 这段评论把...收束成..., 争端的表面理由...合并为一套政策, or 这些相互竞争的传说.... After a sampled hit, scan exact bridge verbs and summary nouns (并列起来|凸显.*重要|不能靠.*却能靠|消息吸入|反而凸显|收束成|正对应正文所说|分成三层|合并为一套政策|它既保留了节日|吉本称其为.*牧函|相互竞争的传说|讲道成功的标准|七眠人传说吸引人的地方) over chapters/final and chapters/translated, then compare only the matched small context with source before deleting confirmed target-only synthesis. 貌似顺滑的桥接句补充：无源桥接有时不是整段，而是嵌在有源段落中的一句顺滑总结，例如“这种答复把……并列起来”、“不能靠本地土壤……把……消息吸入城中”、“这段评论把……收束成……”、“争端的表面理由……合并为一套政策”或“这些相互竞争的传说……”。抽样命中后，应先扫精确桥接动词和总结词（并列起来|凸显.*重要|不能靠.*却能靠|消息吸入|反而凸显|收束成|正对应正文所说|分成三层|合并为一套政策|它既保留了节日|吉本称其为.*牧函|相互竞争的传说|讲道成功的标准|七眠人传说吸引人的地方），范围限于 chapters/final 与 chapters/translated，再只对照命中小上下文，删除确认无源的综合句。

Historical Legal-Status Term Drift / 历史法律身份与劳役术语漂移

• Symptom: In dense historical or legal notes, a recurring institution is translated with several near-synonyms in the same passage, such as 奴役, 服役, 劳役, 服务, or 束缚, even when the source is tracking one legal status such as servitude, bondage, tenure, or a statutory labour category. 症状：历史法制或法律注释中，同一个制度概念在同段或同章内被译成多个近义词，如“奴役”“服役”“劳役”“服务”“束缚”等；但原文其实是在追踪同一法律身份或劳动制度，例如 servitude、bondage、tenure 或成文法中的劳动身份。
• Find by: Start from the glossary row for the high-risk legal term, then scan the translated chapter and nearby final chapters for sibling renderings around the same source cue. For Chinese output, exact-string scans for 奴役|服役|劳役|服务|束缚 near the same names, statutes, courts, or labour groups are usually cheaper than rereading the whole chapter first. 发现方式：先从高风险法律术语的术语表行出发，再扫描本章译稿和相邻终稿中同一 source cue 周边的近义译法。中译本可先用 奴役|服役|劳役|服务|束缚 这类精确检索，限定在同一人物、法规、法院或劳动群体附近，通常比直接重读整章更省 token。
• Risk: The reader may think the source distinguishes different legal statuses when it does not, or may miss that a mild word such as service is being used inside a coercive legal regime. This is especially damaging in slavery, serfdom, apprenticeship, penal labour, and Scottish or colonial legal contexts. 风险：读者会误以为原文区分了多个法律身份，或看不出 service 这类温和词在强制性法律制度中使用。奴隶制、农奴制、学徒身份、刑罚劳动、苏格兰法和殖民地法语境中尤其危险。
• Fix by: Choose one controlled Chinese term for the legal status in that passage, then reserve softer verbs only for ordinary employment or domestic service when the source clearly shifts sense. If the source itself juxtaposes ordinary service with legal servitude, make the contrast explicit in Chinese without inventing a new category. 修复方式：为该段法律身份选定一个受控中文术语；只有当原文明确转入普通雇佣或家内服务时，才使用较软的动词。若原文同时出现普通服务与法律奴役，应在中文中保留这种对照，但不要凭空创造新类别。
• Low-token audit: Use glossary fields, rg scans over the chapter, and small source-context windows. Record exceptions where the source deliberately varies terms or where a later sentence clearly refers to ordinary employment rather than status. 低 token 审计：使用术语表字段、本章 rg 扫描和小上下文原文窗口。若原文本身有意变换术语，或后文已清楚转为普通雇佣意义，应记录为例外。
• Recheck: After normalizing the family, rerun the full-chapter fidelity and Chinese-only readability pass. A round that repaired term drift remains non-PASS and must be followed by a fresh zero-issue recheck. 复查：术语族统一后，重跑整章忠实度与中文独立顺读检查。凡修复过术语漂移的轮次仍不能 PASS，必须追加新一轮零问题复查。

Source Compound Role Literalization / 源语复合身份词直译误导

• Symptom: A source compound such as straw-man, glass-man, chair-man, post-boy, or another hyphenated occupational or performance label is translated as if it named a modern object, creature, office, or fixed title rather than the person or act described in context. 症状：straw-man、glass-man、chair-man、post-boy 等带连字符的职业、身份或表演称呼，被按现代物件、固定头衔或字面实体直译，而没有根据上下文还原为“做某事的人 / 表演某技艺的人 / 承担某角色的人”。
• Find by: During target-only reading, flag any compound rendering that creates a surprising concrete image not supported by the scene. Then scan exact Chinese candidates such as 稻草人, 玻璃人, 椅子人, or overly modern job titles near the same source terms. 发现方式：中文独立顺读时，凡复合词译法生成了场景不支持的具体形象，都要标记。随后用精确词如“稻草人”“玻璃人”“椅子人”或过现代的职称，在同一 source term 附近做检索。
• Risk: The reader may imagine the wrong object or social role, and comic or anecdotal evidence becomes absurd in a way the source did not intend. 风险：读者会想象出错误物件或身份，喜剧性、轶事性证据会变成原文并未打算的荒诞。
• Fix by: Rebuild the compound into a target-language role phrase anchored to the immediate action, such as 耍稻草者的灵巧表演 for a performer balancing a straw on his nose, rather than 稻草人的灵巧. 修复方式：根据近旁动作把复合词改写成目标语角色短语，例如鼻上平衡稻草的表演者，应写作“耍稻草者的灵巧表演”，而不是“稻草人的灵巧”。
• Recheck: Reread the whole anecdote in target language first, then compare the exact source sentence to confirm the repair does not add a new explanation beyond the scene. 复查：先只读目标语整则轶事，再对照源句，确认修复没有越出场景添加解释。

Historical Semantic Downgrading / 历史语义降格

• Symptom: A historically strong praise word or moral term is translated with its weaker modern sense, such as treating eighteenth-century respectable as merely 体面的, 还算不错的, or middle-class social respectability when the passage means 可敬的, 值得尊重的, or 庄严重要的. 症状：历史语境中很强的赞词或道德词，被按现代弱义翻译。例如十八世纪 respectable 在原文中表示“可敬的、值得尊重的、庄严重要的”，却被译成较弱的“体面的”“还算不错的”或中产社会身份感。
• Find by: Flag source words that the note itself defines, contrasts with later usage, or illustrates through many examples. Search the translated passage for modern downgraded renderings and compare each example sentence rather than relying on the dictionary's most familiar current meaning. 发现方式：凡原文注释亲自定义某词、对比其后世用法，或连举多例说明语义，都要标为高风险。检索译文中的现代弱化译法，并逐例对照例句，不要只靠现代最常见词义。
• Risk: The reader loses the historical argument of the note and may misunderstand why the editor quotes many examples. Semantic history notes become flat word glosses rather than evidence. 风险：读者会丢失注释中的历史语义论证，也看不出编者为何连举多例。语义史注释会变成扁平词义解释，而不是证据链。
• Fix by: Keep the source word visible when it is under discussion, choose one historically appropriate target rendering, and repeat that rendering consistently through the examples unless a quoted context requires a deliberate variant. 修复方式：当源词本身是讨论对象时保留源词；选择一个符合历史语义的中文译法，并在各例中一致使用，除非某条引文语境明确要求有意变体。
• Recheck: Do a target-only reading of the whole semantic note, then verify that every example still supports the same historical claim. 复查：先只读中文整条语义史注释，再确认每个例句仍共同支撑同一个历史语义判断。

Short Cross-Reference Note Coverage Loss / 短交叉参照注覆盖不足

• Symptom: Many source notes are short cross-references such as See ante, See post, or a bare title/page citation. A faithful but too-short Chinese rendering can fall below paragraph-coverage gates or look like a fragment in reader-facing notes. 症状：原文大量注释只是 See ante、See post 或裸标题/页码参照。中文若只机械译成“见前文第某页”，虽然忠实，却可能低于段落覆盖门槛，或在读者可见注释中显得像残片。
• Find by: Compare reader paragraph block counts after translation and scan for note blocks shorter than the source's function requires. Pay special attention to note-heavy units where many consecutive notes are cross-references. 发现方式：译后比较 reader paragraph block 数量，并扫描短到不能承担原文功能的注释块。大量连续交叉参照注的单元尤其高风险。
• Risk: The chapter can structurally under-cover the source even when no note number is missing, and readers get a choppy note apparatus with little contextual help. 风险：即使注号没有缺失，章节仍可能在结构上低覆盖源文；读者也会看到碎片化、缺少上下文帮助的注释系统。
• Fix by: Expand only enough to state the cross-reference function, such as 此处为同一话题在前文的交叉参照, 后文该日还有相关补充材料, or 出处用于核对正文引句. Do not invent substantive facts not present in the source. 修复方式：只扩到能说明交叉参照功能即可，例如“此处为同一话题在前文的交叉参照”“后文该日还有相关补充材料”“出处用于核对正文引句”。不得凭空添加源文没有的事实内容。
• Low-token audit: Use metrics first (reader_blocks and blank-line blocks), then inspect only the short note blocks and their source counterparts. 低 token 审计：先看 metrics（reader_blocks 与空行块），再只检查短注块及其源文对应块。
• Recheck: A coverage repair is a fix round, not a PASS round. Rerun full-chapter parity, residue scan, and target-only note readability before allowing the unit forward. 复查：覆盖率修复属于修复轮，不能直接 PASS。必须重跑整章结构、残留扫描和中文注释顺读，才能放行。

Historical Names and Currency Left in Source Form / 历史人名与币名裸留源语

• Symptom: A literary translation keeps ordinary source-language names, currency units, or social terms in Latin letters inside continuous target-language narration, such as Zaide, blanca, maravedí, or real, even though the passage is not discussing source-language form. 症状：文学译文在连续目标语叙事中裸留普通源语人名、币名或社会词，例如 Zaide、blanca、maravedí、real，但该处并不是在讨论源语词形。
• Find by: Start from random-sample hits, then scan final chapters, translated drafts, generated EPUB XHTML, and glossary forbidden renderings with exact strings and rg --no-ignore when ordinary directory scans are affected by ignore rules. 发现方式：先从随机抽样命中出发，再用精确字符串扫描终稿、译稿、生成 XHTML 和术语表禁用写法；若普通目录扫描受 ignore 规则影响，使用 rg --no-ignore。
• Risk: Readers are forced to decode unexplained source words, and the prose feels unfinished. Small currency words are especially risky because they carry poverty, price, and social-status evidence in the scene. 风险：读者被迫自行解码未说明的源语词，译文呈现未完成感。小额币名尤其危险，因为它们承担贫困、价格和阶层证据。
• Fix by: Translate or transliterate into target-language form and add a short class word when needed, such as 扎伊德, 布兰卡小钱, 马拉维迪铜钱, or 雷亚尔银币. Keep the source form in the glossary or a concise note only when useful; do not leave it bare in ordinary narrative. 修复方式：译成或音译成目标语形式，并在需要时加简短类别词，例如“扎伊德”“布兰卡小钱”“马拉维迪铜钱”“雷亚尔银币”。源语形式可留在术语表或简短说明中，不要裸留在普通叙事正文。
• Low-token audit: For Spanish-to-Chinese work, scan likely residue with rg --no-ignore -n "Zaide|blanca|maraved|real|reales" chapters/final chapters/translated output/epub_work/EPUB glossary, then classify frontmatter source-title/source-evidence mentions as documented exceptions. 低 token 审计：西译中可用 rg --no-ignore -n "Zaide|blanca|maraved|real|reales" chapters/final chapters/translated output/epub_work/EPUB glossary 扫描候选，再把前置页原名、来源证据等有记录的源语提及列为例外。
• Post-replacement spacing and measure-word audit: After converting bare source forms to target-language forms, scan for half-repaired Chinese such as 扎伊德 , 布兰卡小钱, 马拉维迪铜钱, 雷亚尔银币, and source-like quantifiers such as 个 布兰卡小钱 or 三十 马拉维迪铜钱. Repair them into target prose units such as 扎伊德的, 半枚布兰卡小钱, 一枚雷亚尔银币, and 三十枚马拉维迪铜钱. 替换后空格与量词审计：把裸露源语改成目标语形式后，还要扫描半修复中文，例如 扎伊德 、 布兰卡小钱、 马拉维迪铜钱、 雷亚尔银币，以及 个 布兰卡小钱、三十 马拉维迪铜钱 这类源语式量词。应修成目标语叙事单位，如“扎伊德的”“半枚布兰卡小钱”“一枚雷亚尔银币”“三十枚马拉维迪铜钱”。
• Denomination-interface audit: Historical currency fixes must also catch generic or unclear denomination phrases such as 一枚二面值的钱, because they hide the currency system from readers. When the source says pieza de a dos near real/reales, use a readable target phrase such as 一枚二雷亚尔银币, then scan for 二面值|个 雷亚尔|个 马拉维迪|半 布兰卡 in final chapters, translated drafts, generated XHTML, and random samples. 面值接口审计：历史币名修复还必须检查“一枚二面值的钱”这类泛化面值说法，因为它会让读者看不出币种体系。源文在 real/reales 语境中写 pieza de a dos 时，可译成“一枚二雷亚尔银币”这类清楚目标语；随后扫描 二面值|个 雷亚尔|个 马拉维迪|半 布兰卡，覆盖终稿、译稿、生成 XHTML 和随机样本。
• Recheck: Rebuild EPUB and rerun publication lint plus a fresh new-seed random round; the stale sample manifest from before the fix must not be counted as a final PASS round. 复查：重建 EPUB，并用新 seed 追加随机抽检；修复前生成的旧样本 manifest 不得计入最终 PASS 轮次。

Gate-Passing Style Debt / 过线但不优秀的顺读债务

• Symptom: A review says the passage is readable, understandable, or faithful, but also repeatedly notes stiffness, density, abstract wording, explanatory prose, source-language syntax residue, or "reads like analysis notes." Scores in the 80s are being treated as final excellence even though the prose is only above the hard failure line. 症状：评审说译文“可读、能懂、忠实”，但反复出现较硬、偏密、略抽象、解释化、源语句法残留、像分析笔记等描述。80 多分被误当成最终优秀，其实只是高于硬失败线。
• Find by: Random-review score distributions with many 80-91 rows, review keywords such as 可读, 略硬, 较硬, 偏密, 略抽象, 解释化, 翻译腔, 英式分析腔, and sample reasons that praise comprehension while withholding natural target-language quality. 发现方式：查看随机评审分布中 80-91 分占比过高的行；扫描评审关键词如 可读、略硬、较硬、偏密、略抽象、解释化、翻译腔、英式分析腔；重点看那些“能懂但不像自然目标语书”的理由。
• Risk: The book passes the hard gate but remains merely serviceable. Later samples keep rediscovering the same polish debt, and random spot-check becomes the main polishing engine instead of a final verification gate. 风险：书稿通过硬门槛，却只是勉强可用。后续抽样会反复发现同类润色债务，随机抽检被迫承担主要精修工作，而不是最终验证。
• Fix by: Route the family back to target-language independent reading and source-supported sentence rebuilding. For zh-Hans, treat 80-87 as rework candidates, 88-91 as not-yet-excellent, and require final artifact reviews to reach average_score >= 92 and lowest_score >= 88 with every sample scored. Do not solve it by raising summary scores; improve the prose or keep the round from final PASS. 修复方式：把该问题族回退到目标语独立阅读和有原文依据的句法重组。简体中文中，80-87 应作为返修候选，88-91 仍未达优秀；最终产物评审必须达到 average_score >= 92、lowest_score >= 88 且每个样本逐项评分。不要靠提高总评分解决；要修正文风，或阻止该轮作为最终 PASS。
• Low-token audit: Scan review files first with rg -n "可读|略硬|较硬|偏密|略抽象|解释化|翻译腔|英式分析腔|分析稍密|较机械|稍显" reviews/random_spotcheck, then inspect only the cited sample units and nearby source context before book-wide pattern searches. 低 token 审计：先扫评审文件：rg -n "可读|略硬|较硬|偏密|略抽象|解释化|翻译腔|英式分析腔|分析稍密|较机械|稍显" reviews/random_spotcheck，再只检查被点名样本及邻近原文，然后按确认的问题族扩大全书候选检索。
• Fiction-dialogue refinement guard: In novels, random samples often expose non-blocking but non-publishable stiffness through child speech, domestic banter, and emotional narration rather than through explicit factual mistakes. When a sample phrase is semantically correct but reads like a literal shell such as 正好吻合, 这是最好的地方, or 深而急地思考着, classify it under this family, audit the whole book with exact-string or near-string searches for the same residue type, and rebuild into target-language novel diction instead of patching only the sampled line. 小说对白精修防护：小说抽样里，很多不过线级而未达出版级的问题，不是事实错，而是孩童口吻、家常对白、情绪叙述里残留的半直译硬壳。若样本出现语义没错、但读感像直译外壳的句子，如“正好吻合”“这是最好的地方”“深而急地思考着”，应归入本问题族；随后用精确残留词或近似残留词做全书扫描，不得只修样本句，必须整体重写为目标语小说腔。
• Picaresque prose guard: In picaresque or early-modern satirical prose, random samples may surface stiff literal shells around hunger, poverty, tricks, and honor, such as 喉咙最能忍, 我指甲工作的那份, 没什么好处可长, 它的不幸作祟, 真实和善良, or 漂亮、剪裁得体的西班牙白话. Audit these exact strings and nearby variants with rg, compare the small source context, and recast them into natural target narrative (最经得住饿, 几片指甲的功劳, 再讨不到什么便宜, 是它晦气, 清白和善意, 漂亮利落的西班牙白话) without flattening the satire. 流浪汉小说散文防护：流浪汉小说或近代讽刺散文中，随机样本常暴露饥饿、贫困、骗术和体面主题周围的半直译硬壳，例如“喉咙最能忍”“我指甲工作的那份”“没什么好处可长”“它的不幸作祟”“真实和善良”“漂亮、剪裁得体的西班牙白话”。应用 rg 审计这些精确短语及邻近变体，对照小上下文原文，再改成自然目标语叙事，如“最经得住饿”“几片指甲的功劳”“再讨不到什么便宜”“是它晦气”“清白和善意”“漂亮利落的西班牙白话”，但不能抹平原文讽刺。
• Honor/hunger/interface extension: When later samples still mark 88-91 prose around honor, hunger, religious office, or classical allusion, do not dismiss the round as fully closed. Scan the cited exact phrases such as 黑东西，也就是所谓体面, 外衣下摆伺候自己, 狂暴而持续的死亡状态, 我死得很难看, 棋步上的错失, 那条蛇，或者说那条公蛇, 仁慈会修士, and 宣告酒事; then recast the confirmed body text into target-readable satire (黑漆漆的所谓体面, 拿外衣下摆凑合, 被饥饿逼得日日像在死, 眼看就要难看地饿死, 棋盘上的失着, 那条“蛇”，也就是我这个小子, 梅尔塞德会修士, 沿街报卖酒). Historical review files and old sample manifests may retain the bad strings as evidence; final chapters, translated drafts, generated XHTML, terminology controls, and current release-facing metadata must not endorse them. 体面/饥饿/接口扩展：后续样本若继续在体面、饥饿、宗教职衔或古典典故周围给出 88-91 分顺读债务，不要直接把该轮视为完全关闭。先扫描被点名的精确短语，如“黑东西，也就是所谓体面”“外衣下摆伺候自己”“狂暴而持续的死亡状态”“我死得很难看”“棋步上的错失”“那条蛇，或者说那条公蛇”“仁慈会修士”“宣告酒事”；再把确认命中的正文改成目标语可读讽刺，如“黑漆漆的所谓体面”“拿外衣下摆凑合”“被饥饿逼得日日像在死”“眼看就要难看地饿死”“棋盘上的失着”“那条‘蛇’，也就是我这个小子”“梅尔塞德会修士”“沿街报卖酒”。历史评审文件和旧抽样 manifest 可以保留坏例作证据；终稿、译稿、生成 XHTML、术语控制和当前发布面元数据不得继续认可这些写法。
• Classical-history analysis-insertion guard: In long historical works, random samples may expose paragraphs that summarize what the author is "doing" rather than translating the source, with phrases such as 吉本把...写成, 吉本关心的不只是, 这种反差更尖锐之处在于, 政治用途相当清楚, or 这条注释把...嵌入. Classify these as over-explanation or inserted-summary defects when the wording is not itself a source/editor note. Audit with exact rg strings over final chapters, translated drafts, and generated XHTML, compare the small source window, and either remove the added commentary or rebuild the passage from source details. 古代史分析腔插入防护：长篇历史书抽检可能暴露一种段落：它不是翻译源文，而是在正文里总结作者“如何写”“关心什么”，常见触发语包括“吉本把……写成”“吉本关心的不只是”“这种反差更尖锐之处在于”“政治用途相当清楚”“这条注释把……嵌入”。若这些话不是源文本或编者注本身，应归入过度解释/插入总结问题族。先用精确 rg 字符串覆盖终稿、译稿和生成 XHTML，再对照小源文窗口；确认后删除插入评论，或按源文细节重译。
• Late classical-history extension: If a fresh seed finds sentences such as 换言之, 这种取舍也解释, 这一路线本身就是, 这个选择把, 这些要求并非偶然, 狄奥菲卢斯的结尾, or source-compressed battle summaries, treat them as the same family even when they are fluent. Low-token audit nearby paragraphs and notes for companion "therefore/this means" bridges, remove confirmed additions, and restore coverage only with direct source facts or source-note material. 古代史后期扩展：新 seed 若发现“换言之”“这种取舍也解释”“这一路线本身就是”“这个选择把”“这些要求并非偶然”“狄奥菲卢斯的结尾”等句式，或把战役事实压缩成战术摘要，即使文字顺，也应归入同一问题族。低 token 审计相邻段落和注释中配套的“因此/意味着/说明”桥段；确认后删除增写，只用源文事实或源注材料补足覆盖率。
• Density-and-collocation follow-up: If a fresh seed no longer finds old exact strings but still clusters around 句群偏密, 搭配略生, or 可懂但略硬, treat the cited units as the same style-debt family when they share source-interface causes. Audit exact phrases like 落在多少贫苦、命运和灾祸里, 把这可怜储藏箱弄成那样, 脸色都被夺走, 在街上吞空气, 穷困已经高到眼前, 变成一张小牌, 把吃饭抵作服务, 告别赦罪符, 交好账, 好交换和好生意, and 到供面时给一对供面; then recast into idiomatic but source-faithful target prose such as 要遭多少贫苦、厄运和灾祸, 把这可怜储藏箱折腾成那副模样, 脸上血色全没了, 到街上拿空气充饥, 穷困已经逼到眉睫, 随手使唤的小牌, 拿吃饭抵工钱, 最后一次讲赦罪符，也算作告别, 交一笔漂亮账, 这笔好买卖、好生意, and 奉献面饼. Do not over-polish away period strangeness; remove only target-language friction that reviewers can point to as a repeatable pattern. 密度与搭配后续防护：如果新 seed 不再命中旧精确短语，却继续围绕“句群偏密”“搭配略生”“可懂但略硬”聚集，且原因仍是源语接口未完全转换，应归入同一顺读债务问题族。可审计“落在多少贫苦、命运和灾祸里”“把这可怜储藏箱弄成那样”“脸色都被夺走”“在街上吞空气”“穷困已经高到眼前”“变成一张小牌”“把吃饭抵作服务”“告别赦罪符”“交好账”“好交换和好生意”“到供面时给一对供面”等精确短语；再改成忠实而自然的目标语，如“要遭多少贫苦、厄运和灾祸”“把这可怜储藏箱折腾成那副模样”“脸上血色全没了”“到街上拿空气充饥”“穷困已经逼到眉睫”“随手使唤的小牌”“拿吃饭抵工钱”“最后一次讲赦罪符，也算作告别”“交一笔漂亮账”“这笔好买卖、好生意”“奉献面饼”。不要把时代怪味过度抹平；只移除评审能明确指认、可复现的目标语摩擦。
• Late-stage recurrence rule: If a third or later fresh seed still reports P3 clusters after prior fixes, the family has not been fully exhausted. Continue low-token exact audits on the newly cited strings, for example 神魂已迁入神圣本体, 看在您性命的分上, 两个两个月加起来, 关系到我的体面, 从我的劳苦中帮一帮他的劳苦, 我看起来较大的门前, 开口不能少于, 拿供养来塞我, and 为了剥夺在场众人如此大的恩惠而说恶言. Prefer small target recasts (仿佛神魂已经升入..., 求您行行好, 两项各两个月的租钱, 会伤我的体面, 靠我的苦处解一解他的苦处, 较富的人家门前, 开口至少要说, 拿供养这句话来堵我, 说恶言，要夺去...) and then force a new seed; do not start final consecutive PASS counting until the latest round has no such clustered style-debt findings. 后期复发规则：如果第三轮或更后面的新 seed 仍在前几轮修复后报告 P3 聚集，说明该问题族还没有追尽。继续对新点名短语做低 token 精确审计，例如“神魂已迁入神圣本体”“看在您性命的分上”“两个两个月加起来”“关系到我的体面”“从我的劳苦中帮一帮他的劳苦”“我看起来较大的门前”“开口不能少于”“拿供养来塞我”“为了剥夺在场众人如此大的恩惠而说恶言”。优先做小幅目标语重写，如“仿佛神魂已经升入……”“求您行行好”“两项各两个月的租钱”“会伤我的体面”“靠我的苦处解一解他的苦处”“较富的人家门前”“开口至少要说”“拿供养这句话来堵我”“说恶言，要夺去……”，随后必须新 seed 复抽；最新轮仍有这类风格债聚集时，不得开始最终连续 PASS 计数。
• Recheck: The next new-seed round should not merely remove blocking issues; the distribution should show excellence-level prose, with no summary-only review file and no cluster of "readable but stiff" comments. 复查：下一轮新 seed 不只看阻塞问题是否消失；分数分布应体现优秀级译文，不能有只写总评的评审文件，也不能继续集中出现“可读但硬”的评语。
• Russian literary syntax guard: Russian-to-Chinese fiction can pass hard fidelity gates while preserving too much source syntax in social narration, dialogue, and interior analysis, such as 用同样的硬币回敬我, 自从我没有见到您以来, 证明顺序, 在脑中推翻所有学校逻辑规则, 你的脑袋会很糟, 强盗的血也没沉默, 安排到他们那一边没有任何好处, or 睡眠也不会闭上我的眼睛. Treat sampled phrases like these as a style-debt family: scan exact residues and close variants, then rebuild into idiomatic Chinese narration (借这一笑还击我, 自从上次见您以来, 论证次序, 先把脑子里那套学校逻辑全推翻, 你的脑袋要遭殃, 强盗的血性还没哑火, 不让他们占任何便宜, 我也别想合眼) while preserving the narrator's irony and cadence. 俄语文学句法防护：俄译中小说可能已过忠实度硬门槛，却在社交叙事、对白和内心分析中保留过多源语句法，如“用同样的硬币回敬我”“自从我没有见到您以来”“证明顺序”“在脑中推翻所有学校逻辑规则”“你的脑袋会很糟”“强盗的血也没沉默”“安排到他们那一边没有任何好处”“睡眠也不会闭上我的眼睛”。这类样本应归入顺读债务问题族：先扫精确残留和近似变体，再改成自然中文叙事，如“借这一笑还击我”“自从上次见您以来”“论证次序”“先把脑子里那套学校逻辑全推翻”“你的脑袋要遭殃”“强盗的血性还没哑火”“不让他们占任何便宜”“我也别想合眼”，同时保留叙述者的反讽和节奏。

Short-Sentence Fragmentation / 短句切断

• Symptom: The translation breaks a coherent source movement into clipped fragments, making the prose read like notes, commands, or a plot outline. 症状：译文把连续动作或论述切成碎短句，读起来像提纲、指令或剧情摘要。
• Find by: Chinese-only read-aloud pass, high punctuation density, repeated comma-only action chains, or samples that feel breathless without source pressure. 发现方式：中文独立朗读、高标点密度、连续动作清单，或原文并无强压迫感但译文读起来喘不过气。
• Fix by: Rebuild sentence groups with natural connectors, subject recovery, rhythm variation, and source-supported pacing; do not flatten tension. 修复方式：用自然连接、主语恢复、长短变化和原文允许的节奏重组句群；不要把紧张感改平。
• Recheck: Read the revised paragraph aloud and compare with the source to ensure no action, order, or tone was changed. 复查：朗读修订段，并对照原文确认动作、顺序和语气没有被改错。

Metaphor Collision / 比喻自撞

• Symptom: A translated image contains mixed vehicles, incompatible physical logic, or extra imagery that sounds vivid but no longer follows the source. 症状：译文意象混用、物理逻辑互撞，或新增了好看但原文没有支撑的比喻。
• Find by: Reviewing high-image passages, chapter openings/endings, emotional turns, and any phrase whose appeal depends on an added object, sound, motion, or psychology. 发现方式：检查高画面段、开头结尾、情绪转折，以及任何魅力依赖新增物体、声音、动作或心理判断的表达。
• Fix by: Preserve the source effect rather than the source surface; remove unsupported added imagery and choose a target-language image grounded in source sensation, action, spatial relation, or symbol. 修复方式：保留原文效果，不机械保留表面；删除无依据新增意象，改用扎根于原文体感、动作、空间关系或象征功能的目标语表达。
• Craft-analysis direction guard: When a source spatial word is metaphorical or evaluative, do not turn it into a concrete room, body, or route unless the scene is literally spatial. For example, takes them to the basement in a value descent should be 带到谷底 or 跌到最底层, not 带到地下室. 写作分析方位防护：源语空间词若承担比喻或评价功能，不要把它译成真实房间、身体或路线，除非场景确实在写空间动作。例如价值下坠中的 takes them to the basement 应译为“带到谷底”或“跌到最底层”，不能写成“带到地下室”。
• Dialogue/craft-book guard: Do not intensify light source metaphors, compressed image syntax, or idioms into new physical, medical, or sound effects. flaws fester in the subtext should not become 缺陷在潜台词中化脓; dribbling basketball should not add clock sounds such as 滴答乱响; fire off words should normally become 说出狠话 or another speech action, not 发射词语; a compressed image sequence such as her shadow slashed rope furled sails... should not become a false action like 影子划开缆绳. 对白/技法书防护：不要把原文轻比喻、压缩意象句法或惯用表达强化成新的物理、病理或声音效果。flaws fester in the subtext 不应译成“缺陷在潜台词中化脓”；dribbling basketball 不应新增“滴答乱响”这类钟表声；fire off words 通常应处理成“说出狠话”等话语行动，不要写成“发射词语”；her shadow slashed rope furled sails... 这类压缩意象序列，不应误析成“影子划开缆绳”这类虚假动作。
• Conceptual-layer guard: In craft theory, source images such as sphere, level, inside, revolves, churns, or roils often teach a conceptual relation, not literal anatomy or mechanics. Preserve the model as a readable target-language relation (三层相套的内在空间, 未言这一层, 最深处, 暗自流转, 翻涌) and avoid false concrete phrases such as 第二层球体，未说之物，在人物内部旋转 or 不可说之物这一层球体. 概念层级防护：写作理论中的 sphere、level、inside、revolves、churns、roils 常是在教学上建立概念关系，不是在描写真实身体结构或机械运动。应转成目标语可读的关系模型（如“三层相套的内在空间”“未言这一层”“最深处”“暗自流转”“翻涌”），避免写成“第二层球体，未说之物，在人物内部旋转”或“不可说之物这一层球体”这类假实体。
• Sanskrit poetic interface guard: In Sanskrit poetry, compact mythic or erotic image syntax can become opaque if translated as a literal noun shell. Allusion phrases such as dohada flower-wish, fire-born divine light, sacrificial fame becoming a river, or wave-hands touching a god's hair need a target-readable relation (树木求花的愿望, 托付给祭火，使神子由火而生, 名声化作水流, 波浪之手攀住) rather than bare shells such as 花愿, 祭火之口, 祭牛之事化成水流, or 浪手抓住. 梵语诗歌接口防护：梵语诗歌的神话、情欲意象常被压缩在复合词和典故里；若直接译成名词硬壳，读者会看见词而看不见关系。dohada 花愿、火生神子、祭祀名声化河、波浪之手触及神发等，应改成目标语能读懂的关系，如“树木求花的愿望”“托付给祭火，使神子由火而生”“名声化作水流”“波浪之手攀住”，不要裸写“花愿”“祭火之口”“祭牛之事化成水流”“浪手抓住”。
• Sanskrit epithet/name interface guard: Sanskrit kavya often uses epithets that are semantically transparent to source readers but opaque if transliterated as a new proper name. Before leaving a target form such as 罗怙王, check whether the source epithet points to a known figure in context, such as रघुपति for Rama. Prefer the reader-recognizable target name in body prose (罗摩的足迹) and record the epithet in glossary or notes when needed. If an unfamiliar place, garden, or plant name must remain transliterated, add a target-language class word and relation, such as 名为室利毗舍罗、富丽宽广的城 or 水边尼俱罗树, not a bare name-like shell. 梵语称号/专名接口防护：梵语诗中常用对源语读者透明的称号；若在译文里直接音译，会像新增专名。正文留下“罗怙王”这类形式前，应先判断源语称号是否在语境中指向已知人物，例如 रघुपति 在本书中指罗摩。正文优先用读者能识别的目标语名，如“罗摩的足迹”；需要保留称号信息时放入术语表或注释。陌生地名、园名、植物名必须音译时，也要补目标语类别和关系，如“名为室利毗舍罗、富丽宽广的城”“水边尼俱罗树”，不要裸留像半成品的名词壳。
• Sanskrit body-image differentiation guard: Do not collapse nearby body-image words into one convenient target term when the poem uses them for different scenes. In Sanskrit poetry, अलका/alaka can be loose curls or hanging locks, while वेणी/veṇī, शिखा/śikhā, and एकवेणी/ekaveṇī can carry braid or single-braid imagery. If one scene shows women lifting loose hair to look at a cloud and another shows the separated wife wearing a single rough braid, translating both as 发辫 creates false continuity. Audit the source word family and use differentiated target terms such as 发绺, 垂发, 发辫, or 独辫. 梵语身体意象分化防护：不要把相邻身体意象词都压成同一个顺手译名。梵语诗中 अलका/alaka 可指垂散发绺/卷发，而 वेणी/veṇī、शिखा/śikhā、एकवेणी/ekaveṇī 可承担辫或独辫意象。若一个场景是远行人妻子拢起垂发望云，另一个场景是离别妻子结起粗涩独辫，两处都译成“发辫”会制造虚假的意象连续。应审计源词族，并区分为“发绺”“垂发”“发辫”“独辫”等。
• Sanskrit eye-glance differentiation guard: Eye and brow words in Sanskrit lyric are not interchangeable modern abstractions. अपाङ्ग/apāṅga is an eye-corner or sidelong glance, भ्रूविलास/bhrū-vilāsa is brow play or brow charm, and नयन/locana is the eye more generally. Do not translate an eye-corner image as 余光 when it carries erotic sidelong glancing, and do not flatten brow play into vague phrases such as 眉间风情. Prefer target phrases like 眼角的流盼, 侧目, 眉梢风情, or 眉梢一弯 according to the local image. 梵语眼波/眉态分化防护：梵语抒情诗里的眼、眉词不能互换成现代抽象词。अपाङ्ग/apāṅga 多是眼角、侧目流盼；भ्रूविलास/bhrū-vilāsa 是眉梢的风情/眉态；नयन/locana 才是更泛的眼。带有情欲侧目的 apāṅga 不宜译成“余光”，bhrū-vilāsa 也不宜泛化成“眉间风情”。应按局部意象译作“眼角的流盼”“侧目”“眉梢风情”“眉梢一弯”等。
• Low-token audit: Search for abstract nouns near hard physical or medical verbs, and for known round-derived image collisions: rg -n "(潜台词|缺陷|对白|词语|感知|意识|心智|内容|痛苦|反讽|兴趣).{0,12}(化脓|发射|固着|沉入|贴合|冲刷|啃噬|钉入|拖|拉|推|撞)|滴答乱响.*篮球|发射出.*词语|贴合耳朵|影子.*划开|划开.*缆绳|卷起的帆|带到地下室|第二层球体|三层球体|层球体|不可说之物这一层球体|slashed rope|furled sails|takes them to the basement" chapters/final chapters/translated output/epub_work/EPUB. 低 token 审计：扫描抽象名词附近的硬物理/病理动词，以及已由轮次发现的意象自撞：rg -n "(潜台词|缺陷|对白|词语|感知|意识|心智|内容|痛苦|反讽|兴趣).{0,12}(化脓|发射|固着|沉入|贴合|冲刷|啃噬|钉入|拖|拉|推|撞)|滴答乱响.*篮球|发射出.*词语|贴合耳朵|影子.*划开|划开.*缆绳|卷起的帆|带到地下室|第二层球体|三层球体|层球体|不可说之物这一层球体|slashed rope|furled sails|takes them to the basement" chapters/final chapters/translated output/epub_work/EPUB。
• Sanskrit low-token audit: For classical Sanskrit poetry, scan the confirmed target residues before rereading: rg -n "花愿|祭火之口|祭牛之事|浪手|触近|越出听觉所及|披湿象皮的念头|那条.*腿|饮入耳中|受爱的福分|脱去这份消瘦|罗怙王|尼拘罗|族名|藏在腹中|你并非寡居|有睡鸽|喷泉水室|举起发辫|太阳来平息泪水|喷泉戏水的云屋|余光|眉间风情|王雁|替爱侣献给你|醒着、睡在地上|幸福会永远来到|裸露的沙洲" chapters/final chapters/translated output/epub_work/EPUB glossary/terms.csv; then inspect only the matched verse units with the nearby Sanskrit source and glossary rows. 梵语低 token 审计：古典梵语诗歌先扫已确认的目标语残留，不要先整章重读：rg -n "花愿|祭火之口|祭牛之事|浪手|触近|越出听觉所及|披湿象皮的念头|那条.*腿|饮入耳中|受爱的福分|脱去这份消瘦|罗怙王|尼拘罗|族名|藏在腹中|你并非寡居|有睡鸽|喷泉水室|举起发辫|太阳来平息泪水|喷泉戏水的云屋|余光|眉间风情|王雁|替爱侣献给你|醒着、睡在地上|幸福会永远来到|裸露的沙洲" chapters/final chapters/translated output/epub_work/EPUB glossary/terms.csv；随后只对命中的诗节看邻近梵语原文和术语表。
• Recheck: Ask whether the sentence remains vivid after every unsupported image is removed. 复查：删除无依据意象后，句子是否仍然有画面；若全靠新增意象才好看，说明需要重译。

Concept Hard-Shell, Entityification, and Psychology Jargon / 概念硬壳、实体化与心理学滥词

• Symptom: A craft concept becomes a rigid target-language term, pseudo-psychology label, physical object, or mechanical spatial model. Typical zh-Hans symptoms include 真实人物 for technical true character, 潜意识心灵, 潜意识自我, 潜意识冲动, 进入人物的潜意识, 巨大领域, 球体, 同心层, and anatomical 内部/内里 where the source is only describing conceptual relation. 症状：写作技法概念被译成目标语里的硬壳术语、伪心理学标签、物理实体或机械空间模型。简体中文常见表现包括把术语 true character 译成“真实人物”，把内在动力写成“潜意识心灵”“潜意识自我”“潜意识冲动”“进入人物的潜意识”，把抽象 realm/sphere/level 写成“巨大领域”“球体”“同心层”，或在非实体语境里滥用“内部/内里”。
• Risk: Readers infer the wrong category: literal real people instead of character nature, pop psychology instead of craft analysis, or physical balls/rooms/body interiors instead of layers of dramatic meaning. This makes the prose sound comic, mechanical, and low-trust even when the source facts are present. 风险：读者会误判概念类别：把人物本性误读成现实真人，把技法分析误读成流行心理学，把戏剧意义层次误读成球体、房间或身体内部。即使事实没丢，译文也会显得滑稽、机械、可信度低。
• Fix by: Rebuild the Chinese relation instead of preserving the source noun shell. For true character, prefer 人物本性 as the terminology anchor, but use 本性, 本性的揭示, 真实是谁, or 到底是什么样的人 when the sentence would otherwise sound rigid; reserve 真实人物 for literal real people and avoid 真实人性 unless the source really means universal human nature. In non-psychology books, do not treat 潜意识 as an acceptable default for subconscious/unconscious; in craft prose, it usually creates a false psychology register. Prefer context-specific forms such as 意识之下, 未被意识到, 深层, 心智深处, 暗流, 深层需求, or 意识之下的欲望. Only a book that is explicitly discussing psychology as psychology should justify any clinical/technical psychology term in a local note. For sphere/realm/level/inside, prefer 层面, 内在, 层层深处, 这一层, 暗自流转, or another readable relation rather than a literal object or space. 修复方式：重建中文关系，不保留源语名词壳。true character 作技法术语时以“人物本性”为锚点；若句子读起来发硬，可按语境改成“本性”“本性的揭示”“真实是谁”“到底是什么样的人”。源文确指现实真人时才用“真实人物”；除非原文确指普遍人性，否则不要用“真实人性”。非心理学书中，不要把“潜意识”当作 subconscious/unconscious 的可接受默认译法；在写作技法散文里，它通常会制造错误的心理学腔。优先按上下文译成“意识之下”“未被意识到”“深层”“心智深处”“暗流”“深层需求”“意识之下的欲望”等。只有一本书明确是在把心理学当作心理学讨论时，才应在本书说明中为临床/技术心理学术语作个案记录。sphere/realm/level/inside 宜译成“层面”“内在”“层层深处”“这一层”“暗自流转”等可读关系，不要硬变成物体或空间。
• Low-token audit: Start with candidate scans rather than full rereading: rg -n "一个潜意识|潜意识心灵|潜意识自我|潜意识冲动|潜意识里|潜意识中的|潜意识中|潜意识的|源自潜意识|来自潜意识|潜伏在潜意识|潜意识欲望|潜意识需求|巨大领域|这些领域|未说出口的领域|球体|同心层|同心球|最内层|真实人物|已说" chapters/final chapters/translated glossary/terms.csv qa/terminology qa/chapter_controls output/epub_work/EPUB, then compare only candidate passages with nearby source. Add a broader secondary scan for suspicious abstract interfaces: rg -n "潜意识|无意识|领域|内里|内部|球体|同心层|真实人物" chapters/final chapters/translated. 低 token 审计：不要先盲读全书，先跑候选扫描：rg -n "一个潜意识|潜意识心灵|潜意识自我|潜意识冲动|潜意识里|潜意识中的|潜意识中|潜意识的|源自潜意识|来自潜意识|潜伏在潜意识|潜意识欲望|潜意识需求|巨大领域|这些领域|未说出口的领域|球体|同心层|同心球|最内层|真实人物|已说" chapters/final chapters/translated glossary/terms.csv qa/terminology qa/chapter_controls output/epub_work/EPUB，再只对候选与邻近原文。再追加一轮宽松抽象接口扫描：rg -n "潜意识|无意识|领域|内里|内部|球体|同心层|真实人物" chapters/final chapters/translated。
• Exception guard: Do not blanket-replace. 真实人物 is correct when the source literally means real people, 领域 may be fine for social/private/public domains, and 无意识 may be acceptable when the source explicitly discusses unconscious thought in a readable contrast. Historical QA files may quote bad phrases as evidence; do not erase evidence, but ensure current PASS terminology and final text do not endorse the defect. 例外防护：不要机械替换。源文确指现实真人时“真实人物”是正确译法；社会/私人/公共等抽象域里“领域”可能成立；源文明确讨论无意识思想且中文自然时，“无意识”可以保留。历史 QA 文件可以引用旧坏例作为证据；不要抹掉证据，但当前 PASS 术语和终稿正文不能继续认可该问题。
• Recheck: Scan final Markdown and generated EPUB XHTML after rebuilding. The book-local retrospective should record the family names, detection regexes, confirmed fixes, and justified exceptions so later reviewers do not rediscover the same issue as isolated word choices. 复查：重建后同时扫描终稿 Markdown 和生成 EPUB XHTML。书内复盘应记录问题族名、检测正则、确认修复和合理例外，避免后续审校把同类问题重新当成零散词语问题发现。

Enumerative Punctuation Drag / 排比标点拖拽

• Symptom: Source commas, semicolons, dashes, or list rhythm are mechanically mapped into target-language punctuation, producing stiff rows of clauses. 症状：把原文逗号、分号、破折号或列举节奏机械搬进目标语，形成僵硬的并列从句队列。
• Find by: Scanning for repeated semicolons, dash chains, parallel clauses with no rhetorical payoff, or chapter titles that preserve printed table-of-contents punctuation. 发现方式：扫描重复分号、破折号链、没有修辞收益的排比从句，以及保留纸书目录标点的章节标题。
• Fix by: Convert lists into natural target-language hierarchy: sentence groups, short clauses, run-in explanations, subtitles, or true tables when needed. 修复方式：按目标语习惯重组层级：句群、短分句、段内说明、副标题，必要时改为真实表格。
• Recheck: Confirm the punctuation now reflects target-language reading logic, not source punctuation inertia. 复查：确认标点服务目标语阅读逻辑，而不是原文标点惯性。
• zh-Hans hard-gate note: In Simplified Chinese publication text, semicolons are rare. Audit them with rg "；" across chapters/final/, frontmatter/, and metadata/; most hits should be rewritten as sentence breaks, commas, or re-layered clauses unless there is a true high-level parallel structure. 简体中文硬门禁补记：出版级简体中文里，分号应极少出现。先用 rg "；" 扫描 chapters/final/、frontmatter/ 和 metadata/；除非确有高层并列关系，否则大多数命中都应改写为断句、逗号或重组层级。
• Paper-like ordinal guard: In reader-facing prose, paragraph-opening 第一，第二，第三 often reads like a paper outline rather than a book. Keep explicit numbering only for true step lists, tables, notes, legal/scientific hierarchy, or places where the source argument depends on numbered order. In ordinary explanatory prose, recast as natural transitions (先看..., 再看..., 最后...) or topic leads (形容词。, 背景故事。). 论文式序号防护：面向读者的正文里，段首 第一，第二，第三 很容易读成论文提纲，而不是一本书。只有真正的步骤清单、表格、注释、法律/科学层级，或原文论证依赖编号顺序时才保留显性编号。普通解释性正文应改成自然过渡（如“先看……”“再看……”“最后……”）或主题句（如“形容词。”“背景故事。”）。
• Low-token audit: For zh-Hans prose, scan paragraph starts with rg -n "^(第一|第二|第三|第四|第五|第六|第七|第八|第九|第十)[，、]" chapters/final chapters/translated output/epub_work/EPUB, then classify notes/true step lists as exceptions and rewrite confirmed body-prose hits. 低 token 审计：简体中文正文可先扫段首序号：rg -n "^(第一|第二|第三|第四|第五|第六|第七|第八|第九|第十)[，、]" chapters/final chapters/translated output/epub_work/EPUB，再把注释/真实步骤清单列为例外，确认正文提纲腔后改写。

Standalone Quote Lost Source Interface / 独立引句丢失原文接口

• Symptom: In books that analyze dialogue, rhetoric, or famous literary/cinematic lines, the translation keeps only the Chinese rendering of a standalone quote block, so readers cannot inspect the source wording that is itself under discussion. 症状：在讨论对白、修辞或经典台词的书里，独立引句块只保留中文译文，不给原文接口，导致读者无法直接核对正被分析的原句声调与措辞。
• Find by: Scan for standalone quotation blocks, attribution lines, repeated em-dash quote attributions, or isolated dialogue examples first with rg, then verify against nearby source context. Focus on canonical lines, famous citations, and analyzed dialogue snippets, not ordinary prose paragraphs. 发现方式：先用 rg 扫描独立引句块、署名行、反复出现的破折号式引句归属或孤立台词示例，再对照邻近原文复核。重点看经典名句、著名引文和被分析的台词片段，不要把普通正文都拉进来。
• Fix by: Pair the original source line or short passage with the translation in reader-visible order, usually original first and translation second. Keep the bilingual pairing limited to places where the exact wording matters to the argument; do not bloat ordinary narrative prose into parallel text.

Source-Language Quote Without Target Interface / 源语长引文缺目标语接口

• Symptom: A reader-facing note or quoted evidence preserves a long Latin, French, Greek-transliterated, or other source-language passage without a usable target-language rendering, or inserts a Chinese citation title/page reference into the middle of the source sentence. The target reader sees evidence but cannot read what the evidence says. 症状：面向读者的注释或证据引文保留较长拉丁文、法文、希腊转写或其他源语段落，却没有可用的目标语译文/大意，或把中文书名、页码、出处插入源语句子中间。目标语读者看见证据，却读不懂证据究竟说了什么。
• Find by: Start from random-sample source-language interface hits, then scan final chapters for long Latin-script runs, diacritics, Greek transliteration, and mixed Chinese/source-language citation breakage such as 中文?Latin or Latin?中文. Review candidates by note role: short names, titles, and bibliographic abbreviations may be acceptable; long quoted evidence needs a target-language interface. 发现方式：先从随机抽样发现的源语接口问题出发，再扫描终稿中较长拉丁字母串、变音字母、希腊转写，以及 中文?Latin、Latin?中文 这类中外文出处破损。按注释功能分类：短人名、书名、文献缩写可以保留；承担证据作用的长引文需要目标语接口。
• Risk: The note becomes a source-language wall inside a translated book. Readers cannot evaluate the quotation, and broken citation insertion can make the foreign sentence itself unreadable. 风险：注释在译本中变成源语墙。读者无法判断引文如何支持正文；出处插入破损时，源语句子本身也会不可读。
• Fix by: Preserve the source quote when exact wording matters, but add a concise target-language rendering or 引文大意 immediately next to it. If a citation has been inserted mid-sentence, move the citation outside the quote and restore the quote's continuity before adding the target-language interface. 修复方式：确需保留原文措辞时可以保留源语引文，但必须在近旁补出简洁目标语译文或“引文大意”。若出处被插入句中，应先把出处移出引文、恢复引文连续性，再补目标语接口。
• Low-token audit: Use a machine scan for long Latin-script passages and mixed punctuation breakage, then repair confirmed evidence quotes rather than ordinary bibliographic fragments. Keep an audit report with candidate count, repaired count, and justified exceptions. 低 token 审计：先用机器扫描长拉丁字母段和中外文标点/出处破损，再修复确认承担证据功能的引文，不把普通书目片段全量改写。审计报告需记录候选数、修复数和合理例外。
• Recheck: Rebuild EPUB and ask fresh random-review agents to inspect note/caption strata; stale failed rounds cannot count after the repair. 复查：重建 EPUB，并让新 seed 随机评审继续检查注释/说明层；修复前失败轮次不得计入通过证据。 修复方式：把原句或短原段与译文成对呈现，通常原文在前、译文在后。双语配对只用于“原句措辞本身就是论证对象”的位置，不要把普通叙述正文膨胀成对照文本。
• Retention boundary: Do not keep source text merely because it is dialogue. Ordinary scene transcripts, greetings, filler, and plot-functional lines should normally be translated into the target language only. Keep the source line only when it is a famous/canonical quotation, a named literary/cinematic line whose wording is being analyzed, or the local argument depends on source-language form, sound, word length, pun, or syntax. 保留边界：不要因为它是对白就自动保留源文。普通场景转写、寒暄、填充语和只推动情节的台词，通常应只给目标语译文。只有名言/经典引文、明确被分析措辞的文学/影视台词，或局部论证依赖源语词形、声韵、词长、双关、句法时，才保留源句。
• Ordinary-dialogue block guard: Even when a craft paragraph briefly needs the source wording to discuss word length, repetition, or cadence, do not promote ordinary fragmentary dialogue into standalone bilingual quote blocks. Prefer target-language prose, or a minimal inline source cue such as a single word being analyzed; reserve full source block plus target block for canonical quotes or passages whose exact wording is the reader-facing evidence. 普通对白块防护：即使技法段落需要短暂提到源语措辞来说明词长、重复或节奏，也不要把普通碎对白提升成独立双语引文块。优先使用目标语正文，或只保留正在分析的最小内联源语提示；完整“源语块在前、目标语块在后”只用于经典引文，或精确措辞本身就是读者可见证据的段落。
• Recheck: Confirm that every added source line serves a reader-facing analytical purpose, that the translation remains primary for continuous reading, and that the chapter has not drifted into indiscriminate bilingual duplication. 复查：确认新增原文都服务于读者可见的分析目的，译文仍然是连续阅读的主文本，章节没有滑向无差别双语堆叠。
• False-positive guard: In quote-analysis books, do not classify a visible English (or other source-language) line as untranslated residue if it is immediately paired with a translation and the pair is clearly functioning as a reader-facing quote interface. Audit only lines that remain source-language-only with no translated partner, or quote pairs whose presentation prevents normal reading. 误判防护：在引句分析型书中，如果一条可见英文（或其他源语）句子紧跟译文成对出现，并且明显承担读者可见的“原句接口”功能，就不要把它误判为 untranslated residue。真正要审计的是只有源语、没有译文配对的残留，或虽成对出现但呈现方式已经妨碍正常阅读的引句块。
• Rendering guard: When a source quote is authored with Markdown blockquote syntax, inspect the generated EPUB XHTML as well as the source Markdown. Literal > / > markers, collapsed verse lines, or source quote lines joined into one paragraph are reader-facing quote-interface defects even if the translation itself is correct. 渲染防护：如果原文引句用 Markdown blockquote 写法保留，必须同时检查生成后的 EPUB XHTML。若 > / > 原样进入读者正文、诗行被压成一段，或原句行序被错误拼接，即使译文内容正确，也属于读者可见的引句接口缺陷。
• Quote-punctuation guard: When a colon introduces a standalone or long quoted passage, scan the reader-facing unit for orphan opening/closing Chinese quote marks. A closing ” without a corresponding visible opening “ is a quote-interface defect, unless it is a documented multi-line quote whose opening mark appears in the same continuous quoted block.
• Transliterated-quote guard: Classical Greek, Latin, or similar transliterated lines can expose doubled source opening marks such as 原句为 ''... or corrupted citation tails such as 载 ai 第二卷 after OCR/repair passes. Scan exact sampled fragments plus patterns like 原句为 '', ''[A-Za-z], and 载 ai|载ai; repair only confirmed quote/citation interfaces against the source. 引号标点防护：冒号引出独立引文或长引文时，要扫描读者可见单元里是否有孤立开/闭中文引号。只有闭引号 ” 而没有对应开引号 “，属于引文接口缺陷；除非它是有记录的多行引文，开引号位于同一连续引文块的前文。
• Audit-unit guard: If a source quote or dialogue line is intentionally retained, keep its target-language translation immediately adjacent in normal reader-facing book layout. Do not solve sampling convenience by adding QA labels such as 原文： / 译文：, by joining languages with /, or by forcing bilingual material into one paragraph. The publishable layout is source paragraph/line first, target paragraph/line second, with the same paragraph structure on both sides when practical. 审计单元防护：如果有意保留原文引句或对白，必须让目标语译文在正常读者版式中紧邻出现。不能为了抽样方便，在正文加入 原文： / 译文： 这类 QA 标签，不能用 / 把两种语言串接，也不能强行把双语内容压进同一段。可出版版式应是源语段/行在前，目标语段/行在后；可行时两边保持相同的段落结构。
• Verse/dialogue layout guard: For verse, famous lines, and short analyzed dialogue, preserve the line/paragraph rhythm. If an English line is its own paragraph, the corresponding Chinese line should also be its own paragraph. Do not collapse verse into comma-separated prose, do not use Markdown > as the only structure, and do not leave literal > / > markers in the generated EPUB. 诗行/台词版式防护：诗行、名句和短被分析台词，要保留行/段节奏。英文若独立成段，对应中文也应独立成段。不要把诗行压成逗号串接的散文，不要只靠 Markdown > 表示结构，更不能让 > / > 裸标记进入生成后的 EPUB。
• Complete-block guard: If several source lines together form one complete quoted unit, especially verse or a multi-line literary speech, present the complete source block first, then the complete target block. Do not interleave English line, Chinese line, English line, Chinese line unless the source itself is designed as alternating bilingual text. Preserve the source line breaks inside the source block and mirror them in the target block when practical. 完整引文块防护：如果多行源语合起来才构成一个完整引文单元，尤其是诗段或多行文学台词，应先呈现完整源语块，再呈现完整目标语块。不要排成英文一行、中文一行、英文一行、中文一行的交错格式，除非原文自身就是交替双语文本。源语块内部保留原行分隔；可行时目标语块也对应保留行分隔。
• Consistency guard: If the same famous or analyzed line appears more than once, the target rendering must remain consistent unless the local context explicitly justifies a variant. A later occurrence of a line such as Of all the gin joints... must not become a syntactically incomplete variant like 走进了我的 when an earlier reader-facing quote already has a complete rendering. 一致性防护：同一句名句或被分析台词多次出现时，译法应保持一致；除非局部上下文明示需要变体。像 Of all the gin joints... 这样的后续引文，不应在前文已有完整译法时变成“走进了我的”这类句法残缺变体。
• Source-form argument guard: If the source argument depends on spelling, initials, rhyme, word shape, pun, portmanteau formation, or an alphabet-constrained list, keep the source-language form in the same reader-facing unit with target-language glosses. A translated-only list can make the explanation incoherent, for example saying "phrases that begin with ba" while showing only Chinese idioms that do not begin with ba; a portmanteau like narcisstick should not become an opaque target-language coinage without the source form and morphology being visible after the target gloss. 源语形式论证防护：如果原文论证依赖拼写、首字母、押韵、词形、双关、混成构词或按字母筛选的列表，必须在同一读者可见单元保留源语形式并配目标语释义。只给中文意译会让解释断裂，例如正文说“以 ba 开头的短语”，但列表里只有并不以 ba 开头的中文惯用语；像 narcisstick 这样的混成词，也不能只造一个中文生词，必须在目标语释义后保留源词和构词说明。
• Source-prosody argument guard: If the argument depends on source-language syllable count, word length, repeated source words, cadence, rhythm, alliteration, assonance, consonance, meter, or word counts, do not leave only the target-language line as evidence. Either keep the analyzed source line with its translation in the same reader-facing unit, or explicitly say the count/claim is "in the source/English line" when the evidence spans a long scene. A sentence such as "squeezes vast content into the fewest, simplest monosyllables" must not point only to Chinese dialogue that is not monosyllabic. If the source repeats one word such as okay five times, do not present several target-language variants as if they were the repeated evidence; name the source word and explain the target variants separately. 源语声韵论证防护：如果论证依赖源语音节数、词长、同一源词反复、语势、节奏、头韵、谐元音、辅音重复、格律或词数统计，不要只给目标语台词作证据。应在同一读者可见单元保留被分析的源语台词并配译文；若证据跨越长场景，至少明确写出“按英文原文/源文统计”。例如正文说“把巨大内容压缩进最少、最简单的单音节词”时，不能只列并非单音节的中文译句。若原文反复的是 okay 这类同一个源词，不要把几个中文变体写成被重复的证据；应点明源词，再另行解释中文按语境如何处理。
• Philological quantity guard: In historical or classical notes, source quantity usually means vowel length, syllable weight, or prosodic quantity, not audio volume. Scan source and target for quantity|accent|metre|meter|syllable|vowel|prosody plus target residues such as 音量; repair confirmed hits to 音长, 音值, 长短, or another context-supported philological term. 训诂音值防护：历史或古典语文学注释里的 quantity 通常指元音长短、音节轻重或韵律量，不是现代声音大小。应扫描 quantity|accent|metre|meter|syllable|vowel|prosody 及目标语残留如 音量；确认后改为 音长、音值、长短 或其他有上下文支撑的训诂术语。
• Source-prosody quote-block guard: If a literary quote is itself the evidence for tongue position, phoneme, stress, or alliteration, a meta note saying the analysis is "about the English" is not enough. Show the complete source quote block first, then the complete target block. Do not make the translated line carry source phonetic evidence it cannot carry. 源语声韵引文块防护：如果文学引文正是舌位、音素、重音或头韵分析的证据，只补一句“分析针对英文”并不够。必须先给完整源语引文块，再给完整目标语译文块；不要让中文译句承担它无法承担的源语发音证据。
• Low-token audit: Collect English-only reader units with a paragraph scanner or rg, then inspect only those candidates and nearby translations. A practical rule is: any paragraph with Latin letters above a small threshold and zero CJK should either be metadata/code, a justified source-only exception, or be merged with its translation in the same paragraph/audit unit. 低 token 审计：用段落扫描脚本或 rg 收集英文-only 读者单元，再只检查这些候选及邻近译文。实用规则是：任何拉丁字母超过少量阈值且没有 CJK 的段落，要么是 metadata/code，要么是有记录的 source-only 合理例外，要么必须与译文合并进同一段落/审计单元。
• Parenthetical proper-name initial guard: When target prose keeps a Latin/English original in parentheses, dropped initials are factual defects, not harmless typography. Examples include rand Signor for Grand Signor, ennadius for Gennadius, regory II for Gregory II, or uratori for Muratori. Audit with bounded regexes for lowercase Latin parentheticals and known proper-name residues, then verify against the same source paragraph before changing ordinary Latin terms. For long history projects, a low-token whole-book pass can pair each lowercase parenthetical with the same chapter source text and ask whether exactly one one-letter-capitalized counterpart exists (term -> Aterm...Zterm). Unique source-backed matches are strong repair candidates; ambiguous matches such as ing -> Ming/Zing require context review. After any batch repair, scan for replacement side effects such as doubled initials (MM..., GG...) or overlapping repairs (MedMina) before closing the family. 括注专名首字母防护：目标语正文在括号中保留拉丁/英文原名时，首字母丢失属于事实/专名错误，不是普通排版问题。例如 rand Signor 应为 Grand Signor，ennadius 应为 Gennadius，regory II 应为 Gregory II，uratori 应为 Muratori。先用括号内小写拉丁词和已知专名残留的有界正则收集候选，再逐条对照同段源文；普通拉丁术语和有意小写形式应记录为合理例外。长篇历史书可做低 token 全书审计：把每个小写拉丁括注同本章源文配对，检查是否存在唯一的一字母大写前缀对应项（term -> Aterm...Zterm）。唯一且有源文支撑者是强修复候选；ing -> Ming/Zing 这类歧义项必须看上下文。批量修复后还要扫描双首字母副作用（MM...、GG...）和重叠替换副作用（如 MedMina），才能关闭问题族。
• Whole-book long-run guard: When one random sample exposes a long foreign quotation after a Chinese gist, audit the whole book with a bounded Latin-script run such as [A-Za-z][A-Za-z\s,.;:()\[\]'"-]{95,}[A-Za-z.] over chapters/final, chapters/translated, and generated XHTML. Keep short names and bibliographic abbreviations as exceptions, but repair long reader-facing evidence strings. If the Chinese gist is already complete and the source wording is not itself under analysis, 外文原句略 is acceptable; otherwise translate the meaning before shortening the source string. 全书长串防护：若随机样本发现“已有中文大意后仍暴露长外文原句”，必须用有界拉丁字母长串正则做全书审计，例如在 chapters/final、chapters/translated 和生成 XHTML 中扫描 [A-Za-z][A-Za-z\s,.;:()\[\]'"-]{95,}[A-Za-z.]。短人名、书目缩写可记为例外，但读者可见的长证据原串要修复。若中文大意已完整且源语措辞本身不是论证对象，可用 外文原句略；否则应先译出含义，再压缩源语原串。
• Low-token audit: To catch over-retained ordinary dialogue, scan English-only lines and scene-transcript triggers before agent review: rg -n "^[A-Za-z].*$|^… [A-Za-z].*$|^—[A-Za-z].*$|Hello\\.|Hi\\.|Anthony\\.|A friend of mine|Tickets|Bermuda|Elbow Beach|I turn down|Say, don.t you feel|I don.t see the point" chapters/final chapters/translated output/epub_work/EPUB, then classify each hit as famous/analyzed source interface, ordinary dialogue to translate-only, metadata/name clutter, or code/source-form analysis. Greetings, filler, and ordinary plot-functional script fragments are confirmed defects if they remain bilingual merely because they are dialogue; source retention is justified only by fame/canonicity or a local argument about source wording, sound, length, pun, syntax, or form. Do not ask an agent to reread the whole book before this candidate pass. 低 token 审计：追查普通对白源文过度保留时，先扫英文-only 行和场景转写触发式：rg -n "^[A-Za-z].*$|^… [A-Za-z].*$|^—[A-Za-z].*$|Hello\\.|Hi\\.|Anthony\\.|A friend of mine|Tickets|Bermuda|Elbow Beach|I turn down|Say, don.t you feel|I don.t see the point" chapters/final chapters/translated output/epub_work/EPUB，再把命中分为名句/被分析源语接口、应只译中文的普通对白、metadata/人名杂音、代码/源语形式分析。寒暄、填充语和普通推动情节的剧本碎片，若只是因为“它是对白”而保留双语，就是确认缺陷；只有名句/经典性，或局部论证依赖源语措辞、声韵、词长、双关、句法、形式时才保留源文。不要在候选收集前让 agent 盲读全书。
• Low-token audit: Add a quote-layout scan over final Markdown and generated XHTML for reader-visible QA labels and joiners: rg -n "原文：|译文：| / 译文：|> >|> >|^> " chapters/final chapters/translated output/epub_work/EPUB. Then inspect only the candidate quote units against nearby source and target text, distinguishing normal adjacent bilingual paragraphs from label leakage. 低 token 审计：对终稿 Markdown 和生成 XHTML 追加引文版式扫描，查读者可见 QA 标签和拼接符：rg -n "原文：|译文：| / 译文：|> >|> >|^> " chapters/final chapters/translated output/epub_work/EPUB。随后只把候选引文单元与邻近源文/译文对照，区分正常相邻双语段落和标签泄漏。
• Low-token audit: For complete-block layout, scan repeated famous-line triggers first instead of rereading the whole book: rg -n "Tomorrow and tomorrow|Creeps in|To the last syllable|And all our yesterdays|The way to dusty death|Towards thee|to the last I grapple|from hell|for hate|Of all the gin joints|Not that there|It was the best|I am an invisible man|Why, man|Well, I’m sure|taking a break|he’s working" chapters/final chapters/translated. Then inspect nearby paragraphs and classify each as complete source block, complete target block, inline explanatory gloss, or reasonable exception. 低 token 审计：完整引文块版式不要盲读全书，先扫重复名句触发式：rg -n "Tomorrow and tomorrow|Creeps in|To the last syllable|And all our yesterdays|The way to dusty death|Towards thee|to the last I grapple|from hell|for hate|Of all the gin joints|Not that there|It was the best|I am an invisible man|Why, man|Well, I’m sure|taking a break|he’s working" chapters/final chapters/translated。随后只检查邻近段落，把候选分为完整源语块、完整目标语块、正文内短释义或合理例外。
• Low-token audit: Add repeated famous-line scans such as rg -n "gin joints|她偏偏走进|走进了我的|走进了我这一间|Tomorrow and tomorrow|尘土之死|Towards thee|Not that there" chapters/final chapters/translated output/epub_work/EPUB, then compare only repeated quote units and documented variants. 低 token 审计：追加同一名句重复扫描，例如 rg -n "gin joints|她偏偏走进|走进了我的|走进了我这一间|Tomorrow and tomorrow|尘土之死|Towards thee|Not that there" chapters/final chapters/translated output/epub_work/EPUB，再只复核重复引文单元和有记录的变体。
• Low-token audit: Add source-prosody scans such as rg -n "单音节|多音节|双音节|三音节|四音节|词数|有声词语|英文原文统计|英文台词|英文短语|同一个词|重复.*(okay|行吗|好吗|没关系)|押韵|头韵|谐元音|节奏|monosyll|polysyll|syllable|word count|alliteration|assonance|consonance|cadence|rhythm|meter" chapters/src chapters/final chapters/translated output/epub_work/EPUB; then inspect only candidates where the reader-visible evidence is translated-only, where source/translation are crammed into inline parentheses, or where target variants replace a repeated source token. 低 token 审计：追加源语声韵扫描，例如 rg -n "单音节|多音节|双音节|三音节|四音节|词数|有声词语|英文原文统计|英文台词|英文短语|同一个词|重复.*(okay|行吗|好吗|没关系)|押韵|头韵|谐元音|节奏|monosyll|polysyll|syllable|word count|alliteration|assonance|consonance|cadence|rhythm|meter" chapters/src chapters/final chapters/translated output/epub_work/EPUB；随后只复核读者可见证据是译文-only、源文/译文被塞进括注混排、或目标语变体替代同一源词反复证据的候选。

Reader-Facing Note Marker Leak / 读者可见注号裸露

• Symptom: Chapter body text exposes raw trailing digits or source-authoring note markers instead of a readable note interface, so random-sample review catches naked 3 / 9 endings or similar residue in reader-facing text. 症状：章节正文把尾随数字或源稿注号原样暴露出来，没有形成可读的注释接口，于是随机抽检会抓到裸露的 3 / 9 之类尾注残留。
• Find by: Run rg or equivalent scans over chapters/final/ for sentence-final digits, then compare with note inventory and built XHTML. Check whether the source markdown itself is readable, not only whether the EPUB builder later repairs it. 发现方式：先对 chapters/final/ 用 rg 或等效扫描找句末数字，再对照注释清单和生成 XHTML。不要只看 EPUB 构建后有没有自动补好，还要检查源 markdown 本身是否可读。
• Low-token audit: Start with rg -n "。\\d+\\s|，\\d+\\s|？\\d+\\s|！\\d+\\s|[\\u4e00-\\u9fff]\\d+\\s" chapters/final chapters/translated, then classify hits as note markers, dates, years, measurements, list numbering, or notes-section entries before editing. 低 token 审计：先用 rg -n "。\\d+\\s|，\\d+\\s|？\\d+\\s|！\\d+\\s|[\\u4e00-\\u9fff]\\d+\\s" chapters/final chapters/translated 收集候选，再把命中区分为注号、日期、年份、计量、列表编号或 notes 章节条目，确认后再编辑。
• Fix by: Replace raw note digits with a controlled note-marker convention recognized by the build pipeline. If random samples or chapter controls treat source Markdown as reader-facing audit units, do not store production noteref markers in chapters/final/; put the injection locations in a non-reader-facing metadata map or equivalent build layer, let the EPUB builder inject noterefs in memory, and keep source Markdown clean. Use a visible note marker only when the review template explicitly accepts it. 修复方式：把裸数字改成构建流水线识别的受控注号约定。如果随机抽检或章节控制把源 Markdown 当作读者可见审计单元，不要把制作性 noteref 标记写进 chapters/final/；应把注号注入位置放入非读者可见的 metadata 映射或等效构建层，由 EPUB 构建器在内存中注入 noteref，并保持源 Markdown 干净。只有评审模板明确接受时，才使用读者可见注号。
• Low-token audit: Search source chapters and built XHTML separately, for example rg -n "<!--noteref:|〔注|见注|&lt;!--noteref" chapters/final chapters/translated output/epub_work/EPUB. Source chapters should have zero production markers; generated XHTML should have resolved <sup epub:type="noteref"> links and no escaped comments. 低 token 审计：分别扫描源章节和生成 XHTML，例如 rg -n "<!--noteref:|〔注|见注|&lt;!--noteref" chapters/final chapters/translated output/epub_work/EPUB。源章节应没有制作标记；生成 XHTML 应呈现已解析的 <sup epub:type="noteref"> 链接，不能残留转义注释。
• Metadata-map guard: After any body sentence carrying a note is rewritten, run a metadata-map stale-entry diagnostic before rebuilding. Compare the number of intended metadata note markers with generated XHTML class="noteref" count; a mismatch means a rewritten sentence no longer matches the map or a source raw digit still lacks a metadata injection point. metadata 映射防护：任何带注释的正文句被改写后，重建前必须运行 metadata 映射失效诊断。把 metadata 中预期注号 marker 数与生成 XHTML 的 class="noteref" 数量对账；若不一致，说明改写后的句子已无法匹配映射，或仍有源层裸数字缺少注入位置。
• Do not rely on builder fallback: If generated XHTML converts 。1 to a valid noteref but chapters/final still shows a naked 1, the family is still open because random samples and source review are reader-facing audit units. 不要依赖构建兜底：如果生成 XHTML 已把 。1 转成有效 noteref，但 chapters/final 仍显示裸 1，问题族仍未关闭，因为随机样本和源层复核同样属于读者可见审计单元。
• Recheck: Confirm chapters/final/ random samples no longer expose raw or visible production note markers, metadata/build-layer injection count matches the intended note inventory, built XHTML renders clickable noterefs, and each note link resolves to the intended note target. 复查：确认 chapters/final/ 抽样文本不再暴露裸数字或可见制作注号，metadata/构建层注入数量与预期注释清单一致，生成 XHTML 呈现可点击 noteref，且每个注释链接都落到目标注释。

Unclear Pronoun Reference / 代词指代不清

• Symptom: A pronoun, demonstrative, or ellipsis that was clear in the source becomes ambiguous after translation, especially in dialogue, multi-character scenes, legal argument, or technical explanation. 症状：原文清楚的代词、指示词或省略，在译文中变得含混；常见于对话、多人物场景、法律论证和技术说明。
• Find by: Reading only the translated paragraph and asking who/what each pronoun points to; then compare with the source chain. 发现方式：只读译文段落，逐个确认“他/她/它/这/那”指向谁或什么，再对照原文链条。
• Fix by: Replace selected pronouns with names, roles, objects, or repeated terms when Chinese clarity requires it; avoid over-repetition that damages style. 修复方式：中文清晰度需要时，把部分代词换成人名、身份、物体或术语；同时避免重复过密损伤文风。
• Recheck: Confirm that a reader unfamiliar with the source can follow the relation without guessing. 复查：确认没读过原文的读者也能不猜测地跟上关系。

Source-Syntax Residue / 源语句法残留

• Symptom: The target text is understandable but still follows source clause order, modifier stacking, passive structure, or abstract noun scaffolding. 症状：译文能懂，但仍保留源语从句顺序、修饰堆叠、被动结构或抽象名词支架。
• Find by: Chinese-only reading, long-sentence breath checks, and source comparison for clauses that could be reordered without changing meaning. 发现方式：中文独立阅读、长句换气检查，并对照原文找出可重排而不改义的从句。
• Special check: Audit dangling contrastive or connective fragments such as isolated "相比之下。", "然而。", or similar source-led hinges. Do not stop at punctuation repair; make sure the comparison or turn has a concrete landing point in the target-language clause. 特别检查：额外审计被截成残句的转折/连接语，例如孤立的“相比之下。”、“然而。”等源语驱动铰链。不要只修标点，要确认比较或转折关系已经在目标语主句里真正落地。
• Special check: Audit source inserted examples such as "for instance" or "for example" that land mid-clause as awkward target-language fragments, for example "但她不能比如捂住肚子". In Chinese, move the example into a natural means phrase such as "不能通过……之类的动作", "不能以……这种方式", or a full sentence with "比如说". 特别检查：审计源语插入式例子，例如 "for instance"、"for example" 被硬塞进目标语分句中部，形成“但她不能比如捂住肚子”这类怪句。中文应把例子改成自然手段表达，例如“不能通过……之类的动作”“不能以……这种方式”，或另起完整的“比如说”句式。
• Special check: Audit English balance idioms in craft-analysis prose, especially throw(s) ... out of balance, life is thrown out of balance, and put life back in balance. Do not preserve the source physical verb as 把早晨/生活抛出平衡; write the target relation directly, such as 打乱了这个早晨的平衡, 让生活失衡, 打乱生活平衡, or 恢复生活平衡. In the same pass, audit emotional mood labels such as sour mood, tetchy, and sullen; avoid face/body pseudo-literals like 一脸酸气, and use context-supported Chinese such as 满心不快, 阴沉烦躁, 心情不佳, or a more specific hidden pain/fear explanation. 特别检查：审计写作分析中的英语平衡习语，尤其是 throw(s) ... out of balance、life is thrown out of balance、put life back in balance。不要把源语物理动词保留成“把早晨/生活抛出平衡”；应直接写目标语关系，例如“打乱了这个早晨的平衡”“让生活失衡”“打乱生活平衡”或“恢复生活平衡”。同轮审计 sour mood、tetchy、sullen 等情绪标签，避免“一脸酸气”这类脸部/身体伪直译，应按上下文写成“满心不快”“阴沉烦躁”“心情不佳”，或更具体的隐秘痛苦/恐惧说明。
• Special check: Audit source category examples where places, cultures, groups, and people are mechanically connected, such as characters in low-context cultures, such as North America. In Chinese, make the category relation explicit (处在北美这类低语境文化中的人物) and avoid turning nationalistic backgrounds into ideology (民族主义背景) when the source means national or national-origin backgrounds (国族背景). 特别检查：审计源语分类例子中地点、文化、群体、人物被机械接上的句子，例如 characters in low-context cultures, such as North America。中文应显化范畴关系（如“处在北美这类低语境文化中的人物”），且不要把 nationalistic backgrounds 误成意识形态“民族主义背景”；若原文指 national / national-origin background，应译为“国族背景”等。
• Special check: Audit English capacity and deictic-purpose frames such as sensitive enough to see... and dedicate their lives to doing exactly that. In Chinese, do not preserve 足以看见 or 恰恰做这件事 when the clause points to awareness or a prior paradox. Prefer target-language relations such as 足够敏锐，意识到... or 献给这件逆势而行的事, after confirming the antecedent in nearby source context. 特别检查：审计 sensitive enough to see...、dedicate their lives to doing exactly that 这类英语能力/指代目的结构。中文不要硬保留“足以看见”“恰恰做这件事”；若原文表达意识到后果或回指前文悖论，应按邻近原文确认后改成“足够敏锐，意识到……”或“献给这件逆势而行的事”等目标语关系。
• Special check: Audit abstract moral/legal words such as misdeeds, wrongdoing, fault, blame, guilt, sin, and offense for near-shape or wrong-register Chinese choices. A rendering like 错行 for misdeeds is a reader-facing word-choice defect; use context-appropriate words such as 过错, 错事, 恶行, 责任, 罪责, or 罪 after source-context comparison. 特别检查：审计 misdeeds、wrongdoing、fault、blame、guilt、sin、offense 等抽象道德/法律评价词，防止近形误选或语域误选。把 misdeeds 译成“错行”是读者可见错词；应对照上下文改为“过错”“错事”“恶行”“责任”“罪责”“罪”等合适表达。
• Special check: Audit literal body, motion, and position metaphors copied from source-language idioms, such as "glides into sentences", "in the guts", "pushed over the top", or "after interest is lost". These often look vivid in isolation but become stiff or physically odd in Chinese. 特别检查：审计从源语习语机械搬来的身体、动作和位置隐喻，例如 "glides into sentences"、"in the guts"、"pushed over the top"、"after interest is lost" 一类表达。它们单看似乎有画面，进入中文后常变成生硬或物理感别扭的句子。
• Special check: Audit abstract-force injury metaphors copied as literal bodily harm, such as "suffered the scald of irony" becoming “被反讽烫伤”. In Chinese, preserve the cognitive/emotional effect with expressions like “尝过……的刺痛”, “感到……刺痛”, or another context-supported idiom rather than making irony, desire, truth, or memory physically burn the reader. 特别检查：审计抽象力量造成身体伤害的硬译，例如把 "suffered the scald of irony" 译成“被反讽烫伤”。中文应保留认知/情绪效果，例如“尝过……的刺痛”“感到……刺痛”或其他有上下文支撑的习惯表达，不要让反讽、欲望、真相、记忆等抽象物在字面上烫伤读者。
• Special check: Audit throat-grab, drag-to-crawl, and concealed-hurt idioms when the source uses body language for rhythm or emotion, such as "the fear of embarrassment grabs them by the throat", "their stride shuffle to a crawl", or "concealing the hurt". Translate the effect as being stuck, slowed, stung, embarrassed, or emotionally hurt instead of importing a literal body-action chain. 特别检查：审计源语用身体动作表达节奏或情绪的 throat-grab、drag-to-crawl、concealed-hurt 等习语，例如 "the fear of embarrassment grabs them by the throat"、"their stride shuffle to a crawl" 或 "concealing the hurt"。应译出卡住、变慢、刺痛、尴尬或受伤感等效果，不要把字面身体动作链搬进中文。
• Special check: Audit source collocations where an abstract noun, text, camera, language, or relationship is made to pour, deliver, resonate, swallow, impose a tongue, cast someone off, or call/generate danger. Phrases like "characters pour out exposition", "indirect dialogue delivers exposition", "lovers pour out their history", "poured out three pages", "resonance of a brick", "camera ... swallow", "imposing their Germanic tongue", "cast off from their marriages", and "calling on capacities / generating jeopardy" should be converted into natural target-language effects: explain too much at once, provide necessary information, recount a shared past, write at length, sound dull, take into frame, impose a language, feel excluded, force out capacity, or enter danger. 特别检查：审计源语搭配把抽象名词、文本、摄影机、语言或关系写成倾倒、输送、共鸣、吞食、强加舌头、抛离、召唤/生成危险的句子。遇到 "characters pour out exposition"、"indirect dialogue delivers exposition"、"lovers pour out their history"、"poured out three pages"、"resonance of a brick"、"camera ... swallow"、"imposing their Germanic tongue"、"cast off from their marriages"、"calling on capacities / generating jeopardy" 等，应转成目标语自然效果：一股脑交代信息、交代必要信息、讲述共同过往、写上很长篇幅、听起来钝硬、收入画面、强加语言、被关系排除、逼出能力、进入险境。
• Special check: Audit craft-analysis idioms that look harmless in English but become literal body, camera, or space mechanics in Chinese: "conversation lubricated with alcohol descends", "value has run its course", "wraps the conversation around her little finger", "staring into the back of his head", "genius working up a sweat", "shape the scene to and around its turning point", "shape language in all the contours", or "the camera sees / moves through global storytelling space". Convert them into the target-language relation: alcohol loosens talk, a value has reached its endpoint, a character controls the conversation, someone is trapped in self-absorption, genius comes from hard labor, the scene forms around a turning point, language is shaped at multiple levels, and the camera presents or moves through narrative space. 特别检查：审计英语里看似普通、进入中文后变成身体/摄影机/空间机械动作的写作分析习语，例如 "conversation lubricated with alcohol descends"、"value has run its course"、"wraps the conversation around her little finger"、"staring into the back of his head"、"genius working up a sweat"、"shape the scene to and around its turning point"、"shape language in all the contours"、"the camera sees / moves through global storytelling space"。应转成目标语关系：酒精让谈话松动、价值走到尽头、人物支配谈话、自我凝视、天才来自艰苦劳动、场景围绕转折点成形、从多个层面塑造语言、镜头呈现或移动于叙事空间。
• Special check: Audit parenthetical source clauses such as influences, if not controls when they are mechanically inserted into Chinese as 影响、即使不控制，也会影响. Rewrite as a natural concessive relation, for example "即便没有完全控制，也影响着...", and remove duplicated verbs caused by source clause order. 特别检查：审计 influences, if not controls 这类英语插入式让步结构，避免机械译成“影响、即使不控制，也会影响”。应改成中文自然让步关系，例如“即便没有完全控制，也影响着……”，并删除由源语语序造成的重复动词。
• Special check: Audit violent state contradictions where the source says someone was harmed while still alive, for example burned and crucified while they were still alive. Do not translate the result as "被烧死、钉死...那时还活着"; preserve the simultaneity as "还活着时，就被..." or an equivalent target-language structure. 特别检查：审计暴力状态逻辑自撞，例如原文说 burned and crucified while they were still alive。不要译成“被烧死、钉死……那时还活着”；应保留“还活着时，就被……”这类同时关系。
• Special check: Audit English stage/camera/instruction wording that becomes awkward physical mechanics in Chinese, for example a scene at a shipyard, looked over his shoulder, Forcefully utilize a big nail, sit back, or secret place from the world / teeter through life. Translate the teaching point or emotional relation directly: a scene set in a shipyard, someone standing behind and watching, hammering a spike in, stepping back, or a private hidden corner and an unsteady life. 特别检查：审计英语场景/镜头/说明句进入中文后变成别扭物理机械的表达，例如 a scene at a shipyard、looked over his shoulder、Forcefully utilize a big nail、sit back、secret place from the world / teeter through life。应直接译出教学点或情绪关系：一场设在造船厂的戏、某人站在身后看着、把长钉锤进去、退后一步、躲开世界的隐秘角落与一路踉跄的人生。
• Special check: Audit compact source grammar/history explanations before translating them as Chinese grammar labels or false language-family facts. If -ing is being discussed as an English form in a progressive-like phrase, do not call every instance 动名词; prefer -ing 形式 or 进行式结构 unless the source clearly means the grammatical gerund. Likewise, do not simplify language-history relations into false claims such as 古德语的一种方言 or 古法语，即拉丁语的一种方言; write safer source-supported relations such as 古日耳曼语根基 or 拉丁语渊源. 特别检查：审计紧凑的源语语法/语言史说明，不要把它们译成误导性的中文语法标签或错误语系事实。若原文讨论的是进行式一类短语中的英语 -ing 形式，不要统称为“动名词”；除非原文明确指 grammatical gerund，否则优先写 -ing 形式 或 进行式结构。同样，不要把语言史关系简化成 古德语的一种方言、古法语，即拉丁语的一种方言 这类错误事实；可按源文支撑写成 古日耳曼语根基、拉丁语渊源 等更稳妥关系。
• Special check: Audit craft-analysis verbs that make scenes, values, beats, forces, or abilities perform English analytical motions, such as "arcing the scene around its value changes", "a turning point pivots", "the charge arcs toward the light", "value charge turns dark / shines", "subtextual actions echo each other", "bulldoze through beats", "moral power controls pace", "assemble abilities", or "drop her head on her arms". Prefer direct target-language relations: let values change, turn, brighten, progress along an arc, echo/correspond as actions, press through beats, pull pace back, integrate abilities, or rest/伏 the head on arms. A source term such as "value charge" may be stable in glossary/source-interface layers, but the target body should not mechanically preserve a physics term like 电荷; use value state, direction, or change unless the local terminology policy explicitly locks a different target term. 特别检查：审计写作分析段里让场景、价值、节拍、力量或能力执行英语分析动作的动词，例如 "arcing the scene around its value changes"、"a turning point pivots"、"the charge arcs toward the light"、"value charge turns dark / shines"、"subtextual actions echo each other"、"bulldoze through beats"、"moral power controls pace"、"assemble abilities" 或 "drop her head on her arms"。优先改为目标语直接关系：价值发生变化、转向、转亮、沿弧线推进、行动彼此呼应、压过节拍、把节奏拉回、整合能力、把头伏到手臂上。value charge 这类源语术语可在术语表/源语接口层稳定，但目标语正文不要机械保留 电荷 这类物理词；除非书内术语 policy 明确锁定，否则用价值状态、走向或变化。
• Special check: When a random sample finds craft-analysis physicalization, generalize it immediately to the whole book. High-yield English triggers include "nail the meaning", "nail characterization", "nail each action/reaction", "nails him with the truth/fact", "hammered home", "bend him farther away", "funnel their struggle", "energy explodes into", "charge erodes toward", "fear runs through", "nose is rubbed in it", "fuse burning of danger", "double-fisted punches", "flinches and withers", "rapid-fire exposition", "deliver exposition", "stubs its toe", "takes hold", "turns sharply into irony", "strikes the eyes and ears", "digging deeper into capacities", "dams emotion momentum", "flow toward", "compresses its power", "moral power controls the pace", "lurching out in defeat", "hooks/holds/pays off", "crafting words as an inner action", "confusion stifles creativity", "run its course", "working up a sweat", and "the width of all things Roman". Chinese candidates often contain suspicious pairings such as 钉住意义, 钉住.*事实, 事实.*锤, 锤进耳朵, 钉住人物塑造, 精准钉住, 弯折得更远, 输送斗争, 爆发进, 朝负面侵蚀, 恐惧在身上奔跑, 鼻子被按在, 危险引线燃烧, 握着真相和道德的双拳, 退缩并萎缩, 快速发射信息, 输送信息交代, 脚趾撞, 需求抓住, 进入.*反讽, 打进眼睛和耳朵, 打到观众.*耳朵.*眼睛, 挖掘.*能力, 筑坝拦住, 流向关键变化, 压缩其力量, 分量.*拉回, 失败地冲出去, 钩住、抓住并兑现, 勾住兴趣, 建立兴趣, 回报兴趣, 封住讲话, 打磨成内在行动, 落实为外在活动, 混乱压住创造力, 跑完自己的路程, 天才.*出汗, or 宽到一切...的宽度. Repair by expressing the target relation directly: let a meaning land, force someone into silence, repeat facts until the audience goes numb, establish characterization, precisely express each action/reaction, move farther from a desire object, focus conflict through a third thing, release energy in the climactic beat, weaken/turn negative, fear surges through someone, have the issue shoved in someone's face, danger approaches, truth/morality strike rhetorically, a person loses nerve or withers in force, information is quickly delivered, provide necessary information, story pauses briefly, a need takes over, irony appears/turns, perception receives something, capacities are mobilized, momentum is held, attention concentrates, power intensifies, moral force controls rhythm, someone is defeated, suspense attracts and rewards attention, words carry inner action into outer activity, confusion blocks creativity, a value reaches its endpoint, genius comes from hard labor, or the world extends as far as an empire's reach. For strikes the eyes and ears, translate the sensory channel relation as "以声音和图像直接抵达观众"; for hooks, builds, and pays off interest, translate the suspense relation as "吸引兴趣、逐步累积、最后给出回报"; for caps a speech with a metaphor, use "以隐喻收束讲话". 特别检查：随机抽检一旦发现写作分析里的抽象概念物理化，要立即归纳为全书问题族追杀。高收益英文触发式包括 "nail the meaning"、"nail characterization"、"nail each action/reaction"、"nails him with the truth/fact"、"hammered home"、"bend him farther away"、"funnel their struggle"、"energy explodes into"、"charge erodes toward"、"fear runs through"、"nose is rubbed in it"、"fuse burning of danger"、"double-fisted punches"、"flinches and withers"、"rapid-fire exposition"、"deliver exposition"、"stubs its toe"、"takes hold"、"turns sharply into irony"、"strikes the eyes and ears"、"digging deeper into capacities"、"dams emotion momentum"、"flow toward"、"compresses its power"、"moral power controls the pace"、"lurching out in defeat"、"hooks/holds/pays off"、"crafting words as an inner action"、"confusion stifles creativity"、"run its course"、"working up a sweat"、"the width of all things Roman"。中文候选常见为 钉住意义、钉住.*事实、事实.*锤、锤进耳朵、钉住人物塑造、精准钉住、弯折得更远、输送斗争、爆发进、朝负面侵蚀、恐惧在身上奔跑、鼻子被按在、危险引线燃烧、握着真相和道德的双拳、退缩并萎缩、快速发射信息、输送信息交代、脚趾撞、需求抓住、进入.*反讽、打进眼睛和耳朵、打到观众.*耳朵.*眼睛、挖掘.*能力、筑坝拦住、流向关键变化、压缩其力量、分量.*拉回、失败地冲出去、钩住、抓住并兑现、勾住兴趣、建立兴趣、回报兴趣、封住讲话、打磨成内在行动、落实为外在活动、混乱压住创造力、跑完自己的路程、天才.*出汗、宽到一切...的宽度。修复时直接写出目标语关系：让意义落定、把人逼到无话可说、反复强调事实直到观众麻木、确立人物塑造、准确落实每个行动/反应、离欲望对象更远、借第三物集中冲突、在高潮节拍中释放、削弱并转向负面、恐惧涌过人物、事情天天怼到脸上、危险迫近、真相/道德形成修辞出击、人物畏缩或气势枯萎、快速交代信息、交代必要信息、故事短暂停顿、需求占据主导、反讽出现/转入、感官接收事物、调动能力、动能被扣住、注意力集中、力量增强、道德力量主导节奏、人物败下阵来、悬念吸引并回报注意力、词语承载内在行动并显于外在活动、混乱妨碍创造力、价值走到尽头、天才源自艰苦劳动、世界延伸到帝国势力所及之处。strikes the eyes and ears 应译为“以声音和图像直接抵达观众”这类感官通道关系；hooks, builds, and pays off interest 应译为“吸引兴趣、逐步累积、最后给出回报”这类悬念关系；caps a speech with a metaphor 应译为“以隐喻收束讲话”。
• Round-derived addendum: Also watch for ordinary English analysis verbs that become unnatural Chinese mechanical actions in samples, including comes into her head, seals meaning, let the scene go to work on you, voice dialogue as music, look into the heart of things, mete out exposition, draws an analogy ... cuisine, speed-building rhythm, and character-specific volume of culture. Suspicious Chinese candidates include 冒进.*脑海, 封住意义, 自行落到.*身上, 把对白发声, 看进事物, 未来事件望去, 分配信息交代, 一道饮食, 饮食经验, 断奏式.*建构速度, 文化量, 精准确立, and 弹成.*跳句. Prefer target-readable relations such as "闯进/钻进脑海", "收束意义", "让场景对读者发生作用", "用音乐唱出对白", "理解核心并期待后续", "交代信息", "同艺术作品或一道具体菜肴建立类比", "速度逐步加快的节奏", "只属于这个人物的文化积累", "准确勾勒人物形象", and "不合逻辑的跳跃句". 轮次沉淀补充：还要警惕普通英语分析动词在样本中变成中文机械动作，例如 comes into her head、seals meaning、let the scene go to work on you、voice dialogue as music、look into the heart of things、mete out exposition、draws an analogy ... cuisine、speed-building rhythm、character-specific volume of culture。中文候选包括 冒进.*脑海、封住意义、自行落到.*身上、把对白发声、看进事物、未来事件望去、分配信息交代、一道饮食、饮食经验、断奏式.*建构速度、文化量、精准确立、弹成.*跳句。优先改成“闯进/钻进脑海”“收束意义”“让场景对读者发生作用”“用音乐唱出对白”“理解核心并期待后续”“交代信息”“同艺术作品或一道具体菜肴建立类比”“速度逐步加快的节奏”“只属于这个人物的文化积累”“准确勾勒人物形象”“不合逻辑的跳跃句”等目标语关系。
• Round-derived addendum: For craft-analysis books, do not leave analytic relations as literal physics or false technical gauges. Audit value charge, charge, exact/specific moment, exact words, look into, eyes and ears, grips interest, topping/surpassing beats, guilt-tripping, Docent/Romantic [name], third-person intelligence/voice, line design, character-specific, payoff, McMansion, fly-by-night, glimpse of meaning, column of talk, and line-level sound analysis such as alliteration or would. Chinese candidates include 价值电荷, 正电荷, 负电荷, 电荷, 充电, 精确时刻, 精确节拍, 精确词语, 看进, 抓住.*眼睛.*耳朵, 抓住兴趣, 超越前一个节拍, 内疚操控, 导览者.*凯末尔, 浪漫主义者.*凯末尔, 浪漫自我, 第三人称智能, 第三人称声音, 人物特异性, 人物专属性, 缩短句行, 拉长句行, 爆成笑声, 麦克豪宅, 一夜就跑, 被一瞥意义触发, 她那一栏话语, and 助动词进入. Repair as target-readable teaching language: value state/change/direction, accurate timing/wording, study/read the subtext, attract visual/auditory attention, maintain interest, move one beat further, use guilt as a tactic, choose natural role labels, use narrative consciousness/third-person manner, line/utterance length, turn payoff into laughter, localize cultural idioms, and preserve short source-form interfaces for sound/grammar analysis. 轮次沉淀补充：写作技法书不要把分析关系留成字面物理动作或伪技术仪表。审计 value charge、charge、exact/specific moment、exact words、look into、eyes and ears、grips interest、topping/surpassing beats、guilt-tripping、Docent/Romantic [name]、third-person intelligence/voice、line design、character-specific、payoff、McMansion、fly-by-night、glimpse of meaning、column of talk，以及头韵、would 等逐字音形分析。中文候选包括 价值电荷、正电荷、负电荷、电荷、充电、精确时刻、精确节拍、精确词语、看进、抓住.*眼睛.*耳朵、抓住兴趣、超越前一个节拍、内疚操控、导览者.*凯末尔、浪漫主义者.*凯末尔、浪漫自我、第三人称智能、第三人称声音、人物特异性、人物专属性、缩短句行、拉长句行、爆成笑声、麦克豪宅、一夜就跑、被一瞥意义触发、她那一栏话语、助动词进入。修复为目标语教学语言：价值状态/变化/走向、准确时机/词语、研读潜台词、吸引视听注意、维持兴趣、节拍更进一步、把内疚当策略、选择自然角色标签、叙述意识/第三人称口吻、台词长短、把包袱转化为笑声、本地化文化习语，并在音形/语法分析处保留短源语接口。
• Round-derived addendum: When samples reveal source idioms that look ordinary but become false Chinese facts or non-words, immediately audit the whole book for the exact Chinese residue and the nearby source trigger. High-yield triggers include connects you more to life, build toward a loss of identity, conscious concern for a secure doctor/patient relationship, put him at the scene by name, taking a break from my wife, forgetting my son's birthday, endorse a whiskey, racing thoughts, mad, ticking clock, squeezes vast content, value at stake, and fixed desires for stability. Chinese candidates include 更连接生命, 连接生命, 建向, 意识关切, 安全医患关系, 按名字把.*放到现场, 请个假, 帮我忘掉, 奔跑念头, 疯狂滴答作响的钟, 巨大内容, 所押注, 押注在, 价值电荷, and 对稳定的固定欲望. Repair by writing the target relation directly: 更贴近生命, 逐步走向身份丧失, 有意识地关心关系是否安全, 目击者可以指认他在场, 离开妻子喘口气, 忘了儿子的生日, 给威士忌做代言, 停不下来的念头, 像一座疯狂滴答作响的钟, 丰富含义, 利害攸关的价值, 价值状态, and 维持稳定的固定欲望. 轮次沉淀补充：抽检一旦发现普通英语习语进入中文后变成虚假事实或不成词，要立刻用精确中文残留和邻近源语触发式全书审计。高收益触发式包括 connects you more to life、build toward a loss of identity、conscious concern for a secure doctor/patient relationship、put him at the scene by name、taking a break from my wife、forgetting my son's birthday、endorse a whiskey、racing thoughts、mad, ticking clock、squeezes vast content、value at stake、fixed desires for stability。中文候选包括 更连接生命、连接生命、建向、意识关切、安全医患关系、按名字把.*放到现场、请个假、帮我忘掉、奔跑念头、疯狂滴答作响的钟、巨大内容、所押注、押注在、价值电荷、对稳定的固定欲望。修复时直接写目标语关系：更贴近生命、逐步走向身份丧失、有意识地关心关系是否安全、目击者可以指认他在场、离开妻子喘口气、忘了儿子的生日、给威士忌做代言、停不下来的念头、像一座疯狂滴答作响的钟、丰富含义、利害攸关的价值、价值状态、维持稳定的固定欲望。
• Low-token audit: Search target text for suspicious injury verbs near abstract nouns, for example rg -n "反讽.{0,8}(烫|灼|烧|刺伤)|被.{0,8}(反讽|欲望|真相|记忆).{0,8}(烫|灼|烧|刺)" chapters/final output/epub_work/EPUB, then compare only the candidate sentences with nearby source. 低 token 审计：用抽象名词附近的伤害动词收集候选，例如 rg -n "反讽.{0,8}(烫|灼|烧|刺伤)|被.{0,8}(反讽|欲望|真相|记忆).{0,8}(烫|灼|烧|刺)" chapters/final output/epub_work/EPUB，再只把候选句与邻近原文对照。
• Low-token audit: Search for malformed inserted-example residue with rg -n "不能比如|可以比如|会比如|应比如|要比如|但.{0,8}比如|比如捂住|for instance|for example" chapters/final chapters/translated output/epub_work/EPUB, then classify hits as ordinary examples, source-only evidence, or confirmed syntax residue. 低 token 审计：用 rg -n "不能比如|可以比如|会比如|应比如|要比如|但.{0,8}比如|比如捂住|for instance|for example" chapters/final chapters/translated output/epub_work/EPUB 搜索插入式例子残留，再区分普通举例、源语证据和确认句法残留。
• Low-token audit: Search for balance-idiom and mood-label residue with rg -n "抛出平衡|丢出平衡|扔出平衡|甩出平衡|掷出平衡|一脸酸气|酸气|酸溜溜|out of balance|back in balance|sour mood|tetchy|sullen" chapters/src chapters/final chapters/translated output/epub_work/EPUB, then compare only paired source/final candidates. Treat natural target forms such as 让生活失衡, 打乱生活平衡, 恢复生活平衡, or context-specific hidden pain/fear as reasonable exceptions. 低 token 审计：用 rg -n "抛出平衡|丢出平衡|扔出平衡|甩出平衡|掷出平衡|一脸酸气|酸气|酸溜溜|out of balance|back in balance|sour mood|tetchy|sullen" chapters/src chapters/final chapters/translated output/epub_work/EPUB 搜索平衡习语和情绪标签残留，再只复核源文/终稿成对候选。让生活失衡、打乱生活平衡、恢复生活平衡 或上下文具体化的隐秘痛苦/恐惧可记为合理例外。
• Low-token audit: Search for ordinary idiom residue and craft-term hard literals with rg -n "更连接生命|连接生命|巨大内容|建向|意识关切|安全医患关系|请个假|帮我忘掉|奔跑念头|疯狂滴答作响的钟|按名字把.*放到现场|所押注|押注在|价值电荷|对稳定的固定欲望" chapters/final chapters/translated output/epub_work/EPUB, then compare only the candidate paragraph with nearby source. Source-side expansion, if needed: rg -n "connects you more to life|build toward|conscious concern|put him at the scene by name|taking a break from my wife|forgetting my son|endorse a whiskey|racing thoughts|ticking clock|squeezes vast content|value at stake|fixed desires for stability" chapters/src. 低 token 审计：用 rg -n "更连接生命|连接生命|巨大内容|建向|意识关切|安全医患关系|请个假|帮我忘掉|奔跑念头|疯狂滴答作响的钟|按名字把.*放到现场|所押注|押注在|价值电荷|对稳定的固定欲望" chapters/final chapters/translated output/epub_work/EPUB 搜索普通习语残留和技法术语硬译，再只复核候选段及邻近原文。必要时用源语扩展式：rg -n "connects you more to life|build toward|conscious concern|put him at the scene by name|taking a break from my wife|forgetting my son|endorse a whiskey|racing thoughts|ticking clock|squeezes vast content|value at stake|fixed desires for stability" chapters/src。
• Low-token audit: Add classification and moral-word scans: rg -n "低语境文化中的人物，例如|民族主义背景|例如北美|文化中的人物|错行|misdeed|wrongdoing|fault|blame|guilt|sin|offense|过错|恶行|错事|罪行|失误|责任" chapters/src chapters/final chapters/translated output/epub_work/EPUB; then inspect only source/final pairs, because many 责任 or 错误 hits are reasonable exceptions. 低 token 审计：追加分类句和道德评价词扫描：rg -n "低语境文化中的人物，例如|民族主义背景|例如北美|文化中的人物|错行|misdeed|wrongdoing|fault|blame|guilt|sin|offense|过错|恶行|错事|罪行|失误|责任" chapters/src chapters/final chapters/translated output/epub_work/EPUB；随后只复核源文/终稿成对候选，因为大量“责任”或“错误”命中属于合理例外。
• Special check: Audit abstract-force drag/push/pull metaphors, such as "turns sharply into the powerful irony" becoming “被一记更强的反讽拉了回去”. When the source describes a rhetorical turn, value shift, or conceptual relation, translate the turn directly instead of inventing a physical force acting on the sentence or reader. 特别检查：审计抽象力量拖拽/推动/拉扯类硬译，例如把 "turns sharply into the powerful irony" 译成“被一记更强的反讽拉了回去”。当原文说的是修辞转折、价值转向或概念关系时，应直接译出“转向/急转/进入/形成”，不要让抽象物变成对句子或读者施力的外物。
• Low-token audit: Search target text for abstract nouns near physical force verbs, for example rg -n "(反讽|真相|答案|兴趣|欲望|痛苦|恐惧|羞辱|冲突|价值|关系).{0,12}(拉|拖|推|拽|扯)|(被|由).{0,12}(反讽|真相|答案|兴趣|欲望|痛苦|恐惧|羞辱|冲突|价值|关系).{0,12}(拉|拖|推|拽|扯)" chapters/final chapters/translated. 低 token 审计：扫描抽象名词附近的物理施力动词，例如 rg -n "(反讽|真相|答案|兴趣|欲望|痛苦|恐惧|羞辱|冲突|价值|关系).{0,12}(拉|拖|推|拽|扯)|(被|由).{0,12}(反讽|真相|答案|兴趣|欲望|痛苦|恐惧|羞辱|冲突|价值|关系).{0,12}(拉|拖|推|拽|扯)" chapters/final chapters/translated。
• Low-token audit: For craft books, also scan concrete suspect collocations before asking an agent to reread: rg -n "钉住|精准钉住|弯折|输送.*(斗争|信息|交代)|推动.*穿越|撞上.*真实|宽到|正电荷.*侵蚀|恐惧在.*奔跑|鼻子.*按在|爆发进|潜意识的水流|危险引线|握着.*双拳|退缩并萎缩|快速发射|脚趾撞|需求抓住|进入.*反讽|打进眼睛|挖掘.*能力|筑坝拦住|流向关键变化|压缩其力量|分量.*拉回|失败地冲出去|钩住、抓住并兑现|打磨成内在行动|落实为外在活动|混乱压住创造力|跑完自己的路程|天才.*出汗|酒精润滑|滑向.*真相|绕回自己的小指|后脑勺里面|摄影机看见|穿过整体叙事空间|所有轮廓中塑造|抓住观众的耳朵|冠军取悦|只列出以字母" chapters/final chapters/translated, then classify each hit as confirmed fix or reasonable exception. 低 token 审计：写作技法书还应先用具体可疑搭配收集候选，再交给 agent 小上下文复核：rg -n "钉住|精准钉住|弯折|输送.*(斗争|信息|交代)|推动.*穿越|撞上.*真实|宽到|正电荷.*侵蚀|恐惧在.*奔跑|鼻子.*按在|爆发进|潜意识的水流|危险引线|握着.*双拳|退缩并萎缩|快速发射|脚趾撞|需求抓住|进入.*反讽|打进眼睛|挖掘.*能力|筑坝拦住|流向关键变化|压缩其力量|分量.*拉回|失败地冲出去|钩住、抓住并兑现|打磨成内在行动|落实为外在活动|混乱压住创造力|跑完自己的路程|天才.*出汗|酒精润滑|滑向.*真相|绕回自己的小指|后脑勺里面|摄影机看见|穿过整体叙事空间|所有轮廓中塑造|抓住观众的耳朵|冠军取悦|只列出以字母" chapters/final chapters/translated，随后把每个命中归类为确认修复或合理例外。
• Low-token audit: Include round-derived source-syntax and grammar/fact residues rg -n "敏感到足以|足以看见|恰恰做这件事|只是很不同|那是什么，好吗|称呼规范|思索.*之间的性质|从他肩后看过去|有力地使用一枚大钉子|躲避世界的秘密之处|摇摇晃晃地穿过生活|向后坐下|古德语的一种方言|古法语，即拉丁语的一种方言|动名词短语|动名词只给|使用动名词短语" chapters/final chapters/translated output/epub_work/EPUB, then compare each candidate against nearby source before editing. 低 token 审计：追加轮次沉淀的源语句法和语法/事实残留扫描：rg -n "敏感到足以|足以看见|恰恰做这件事|只是很不同|那是什么，好吗|称呼规范|思索.*之间的性质|从他肩后看过去|有力地使用一枚大钉子|躲避世界的秘密之处|摇摇晃晃地穿过生活|向后坐下|古德语的一种方言|古法语，即拉丁语的一种方言|动名词短语|动名词只给|使用动名词短语" chapters/final chapters/translated output/epub_work/EPUB，每个候选都先对照邻近原文再编辑。
• Low-token audit: Add round-derived hinge/deliver and abstract-collocation residues rg -n "三部电影都系在|交付这条信息交代|信息交代事实|事实知识充足|狡诈策略|穿越私人生活|母亲所说的话|亚洲式形状|调谐到|现在的刀锋|同故事对抗力量挣扎|被讲出和问出|整体性的感觉|数万款项" chapters/final chapters/translated output/epub_work/EPUB, then source-expand with rg -n "hinge on|deliver exposition|story fact|factually informed|devious tactics|navigate his personal life|mother tongue|Asian shape|tuned to|cutting edge of the present|antagonistic forces of story|spoken and asked|feel for the whole|stolen tens of thousands" chapters/src only for unclear candidates. 低 token 审计：追加轮次沉淀的 hinge/deliver 和抽象搭配残留扫描：rg -n "三部电影都系在|交付这条信息交代|信息交代事实|事实知识充足|狡诈策略|穿越私人生活|母亲所说的话|亚洲式形状|调谐到|现在的刀锋|同故事对抗力量挣扎|被讲出和问出|整体性的感觉|数万款项" chapters/final chapters/translated output/epub_work/EPUB，只对不明确候选再用源语扩展式 rg -n "hinge on|deliver exposition|story fact|factually informed|devious tactics|navigate his personal life|mother tongue|Asian shape|tuned to|cutting edge of the present|antagonistic forces of story|spoken and asked|feel for the whole|stolen tens of thousands" chapters/src。
• Low-token audit: After the round-derived addendum above, include rg -n "弹成|不合逻辑的跳句|断奏式|建构速度|一道饮食|饮食经验|文化量|精准确立|确立.*人物塑造|把对白发声|发声为音乐|看进事物|未来事件望去|分配信息交代|冒进.*脑海|封住意义|自行落到.*身上|落到你身上|终身狂热影响|即使不控制|影响、即使|被烧死、钉死|还活着.*被烧死" chapters/final chapters/translated output/epub_work/EPUB. 低 token 审计：对上述轮次沉淀项，还要追加 rg -n "弹成|不合逻辑的跳句|断奏式|建构速度|一道饮食|饮食经验|文化量|精准确立|确立.*人物塑造|把对白发声|发声为音乐|看进事物|未来事件望去|分配信息交代|冒进.*脑海|封住意义|自行落到.*身上|落到你身上|终身狂热影响|即使不控制|影响、即使|被烧死、钉死|还活着.*被烧死" chapters/final chapters/translated output/epub_work/EPUB。
• Low-token audit: For round-derived craft analysis residue, add rg -n "价值电荷|正电荷|负电荷|电荷|充电|精确时刻|精确节拍|精确词语|看进|抓住.*眼睛.*耳朵|抓住兴趣|超越前一个节拍|内疚操控|导览者.*凯末尔|浪漫主义者.*凯末尔|浪漫自我|第三人称智能|第三人称声音|人物特异性|人物专属性|信息交代事实|缩短句行|拉长句行|爆成笑声|麦克豪宅|一夜就跑|被一瞥意义触发|她那一栏话语|助动词进入|一张又一张画布地|方法必须反向移动|烟屁股|备用胎|微波炉节目|不可说之物这一层球体|平庸炸弹|人格不可能" chapters/final chapters/translated output/epub_work/EPUB, then classify false positives such as legitimate idioms, verse punctuation, stable target terms like 行动主线, or source-interface quotations before editing. 低 token 审计：对轮次沉淀的写作分析残留，追加 rg -n "价值电荷|正电荷|负电荷|电荷|充电|精确时刻|精确节拍|精确词语|看进|抓住.*眼睛.*耳朵|抓住兴趣|超越前一个节拍|内疚操控|导览者.*凯末尔|浪漫主义者.*凯末尔|浪漫自我|第三人称智能|第三人称声音|人物特异性|人物专属性|信息交代事实|缩短句行|拉长句行|爆成笑声|麦克豪宅|一夜就跑|被一瞥意义触发|她那一栏话语|助动词进入|一张又一张画布地|方法必须反向移动|烟屁股|备用胎|微波炉节目|不可说之物这一层球体|平庸炸弹|人格不可能" chapters/final chapters/translated output/epub_work/EPUB，编辑前先区分成语、诗歌引文标点、行动主线 这类稳定目标语术语、必要源语接口等合理例外。
• Low-token audit: Add the round-derived abstract/voice residue scan rg -n "丑闻缠身地|声名狼藉地富有|富有而出名|回看合乎逻辑|合乎逻辑的感知|叙述智能|人物专属声音|人物声音这一侧|角色身份|人物身份|以角色身份|词语文本|可见的尖端|巨大内容.*沉入|孤独区域|临时冬眠|令人烦扰的迟钝" chapters/final chapters/translated output/epub_work/EPUB, then compare only candidates with nearby source before editing. 低 token 审计：追加本轮沉淀的抽象/声音接口残留扫描：rg -n "丑闻缠身地|声名狼藉地富有|富有而出名|回看合乎逻辑|合乎逻辑的感知|叙述智能|人物专属声音|人物声音这一侧|角色身份|人物身份|以角色身份|词语文本|可见的尖端|巨大内容.*沉入|孤独区域|临时冬眠|令人烦扰的迟钝" chapters/final chapters/translated output/epub_work/EPUB，只把候选与邻近原文对照后再编辑。
• Evaluative-adverb guard: Do not turn degree or social-color adverbs into unsupported facts. For example, scandalously rich and famous may mean rich and famous to an almost shocking or ostentatious degree; do not translate it as 丑闻缠身 unless the source explicitly asserts actual scandals. 评价副词防护：不要把程度或社会色彩副词改成无依据事实。例如 scandalously rich and famous 可表示富有和出名到近乎惊人/招摇的程度；除非原文明说丑闻事实，不要译成“丑闻缠身”。
• Fix by: Reorder information for target-language comprehension while preserving facts, emphasis, causality, and narrative viewpoint. 修复方式：按目标语理解顺序重排信息，同时保留事实、重心、因果和叙述视角。
• Recheck: Verify the revised sentence no longer sounds translated, and that no source relation was lost. 复查：确认修订句不再像翻译腔，同时没有丢失原文关系。

Reader-Facing Source-Name Clutter / 读者界面原名杂音

• Symptom: A target-language paragraph interrupts ordinary reading with multiple source-language names, role names, or parenthetical originals in one sentence, even though the exact source spelling is not being analyzed. 症状：目标语正文一句里塞入多个源语人名、角色名或括注原名，打断普通阅读；而该处并不是在分析这些源语拼写本身。
• Find by: Scan sampled units and final chapters for dense Latin-script names, repeated parenthetical originals, and mixed source/target references to the same person in nearby sentences. Exempt quote-analysis lines where the source wording itself is the object of analysis. 发现方式：扫描抽样单元和终稿章节中的密集拉丁字母人名、重复括注原名，以及邻近句里同一人物源语/目标语混用的情况。若该处正在分析原句措辞本身，应作为合理例外。
• Fix by: Keep a single useful source interface at first natural mention, then use the target-language name or role consistently. Move work titles into title parentheses if needed, but do not parenthesize every person name in explanatory prose. When the source uses role labels for contrasted selves or narrators, choose readable target-language role names and keep them stable, for example Docent Kemal / Romantic Kemal should become consistent target roles instead of awkward literal labels. 修复方式：在第一次自然出现处保留一个有用的原名接口，之后稳定使用目标语姓名或身份。作品标题必要时可括注原名，但不要在解释性正文里给每个人名都加括注。若原文用角色标签区分叙述者、自我或人物面向，应选择目标语可读角色名并保持稳定，例如 Docent Kemal / Romantic Kemal 应处理成稳定的目标语角色称谓，而不是生硬直译标签。
• Common-name guard: For books that discuss many films, plays, novels, people, and places, do not invent a Chinese title from the source meaning when a mainland/common Chinese rendering exists. Also do not substitute the actor/creator name for the work title. Confirmed defect patterns include The Larry Sanders Show becoming 《加里·山德林秀》, Curb Your Enthusiasm becoming 《抑制热情》, Pi becoming a different film such as 《源代码》, and Shakespeare full titles being carried into every later mention as heavy literal forms such as 《理查三世悲剧》 or 《奥赛罗悲剧》. Where a title such as Every Man out of His Humour has competing or sparse evidence, use the user/editor-approved title map and lock it in glossary/terms.csv. 通行专名防护：一本书密集讨论电影、戏剧、小说、人名和地名时，不要在已有大陆/通行译名的地方按源语字面硬造中文题名，也不要把演员/创作者名误作作品名。已确认的问题模式包括把 The Larry Sanders Show 误成 《加里·山德林秀》，把 Curb Your Enthusiasm 直译成 《抑制热情》，把 Pi 误作另一部电影 《源代码》，以及把莎士比亚全题名在后文反复写成 《理查三世悲剧》、《奥赛罗悲剧》 这类沉重直译。若 Every Man out of His Humour 这类题名证据稀疏或存在分歧，应采用用户/编辑确认的题名映射，并写入 glossary/terms.csv 锁定。
• First-mention interface guard: A first natural body mention may need the source title, author, or place original, but headings should not carry this burden. Use the Chinese title or translated place name for reading, then a compact parenthetical source interface only once when it helps identification, e.g. 《人人扫兴》（Every Man out of His Humour）, 《消消气》（Curb Your Enthusiasm）, 底特律滨水区（Detroit waterfront）, or 东京柏悦酒店（Park Hyatt Tokyo）. Later mentions should drop the parenthetical unless the source form itself is being analyzed. 首次接口防护：正文第一次自然出现处可补原题、作者或地名原文，但标题不承担这个负担。阅读面使用中文题名或地名译名，再在确有识别价值时用一次紧凑括注，如 《人人扫兴》（Every Man out of His Humour）、《消消气》（Curb Your Enthusiasm）、底特律滨水区（Detroit waterfront）、东京柏悦酒店（Park Hyatt Tokyo）。后文除非正在分析原文形式，否则不再括注。
• Low-token audit: Search for dense parenthetical source names and unstable role labels, for example rg -n "（[A-Z][A-Za-z .'-]+）|Donald Hall|Robert Frost|Erwin Piscator|Thornton Wilder|讲解员凯末尔|浪漫凯末尔" chapters/final chapters/translated, then exempt only bibliography, title interfaces, and source-form analysis. 低 token 审计：扫描密集括注原名和不稳定角色标签，例如 rg -n "（[A-Z][A-Za-z .'-]+）|Donald Hall|Robert Frost|Erwin Piscator|Thornton Wilder|讲解员凯末尔|浪漫凯末尔" chapters/final chapters/translated，只豁免书目信息、作品标题接口和源语形式分析。
• Recheck: Read the paragraph without the source text and confirm the names no longer feel like production notes or terminology scaffolding. 复查：不看原文只读该段，确认这些名字不再像制作说明或术语脚手架。

Verse Alignment and Route Arc Drift / 诗节对齐与路线弧线漂移

• Symptom: A verse-aligned poem, messenger poem, pilgrimage route, itinerary chapter, or formally numbered source is translated into target-language units that keep roughly the same count but no longer correspond to the same source verses. The target may move later-scene content into an earlier route chapter, invent a final verse, skip a variant verse marker, or collapse geography into generic travel prose. 症状：诗节对齐的诗、信使诗、巡游路线、朝圣路线或带正式编号的源文，被译成数量大体相同但不再逐节对应的目标语单元。译文可能把后文场景搬进前面的路线章，自创末节，漏掉异文节号，或把地理路径压成泛泛旅行文字。
• Risk: Readers lose the poem's structural argument: route order, scene-to-scene movement, mythological geography, and message timing. A smooth target-language chapter can therefore be a severe fidelity failure even when headings and counts appear to pass. 风险：读者会失去诗歌的结构逻辑：路线先后、场景转换、神话地理和传信时机。即使标题和数量看似通过，顺滑译文也可能构成严重忠实度失败。
• Find by: Pair source headings with target headings and scan for numbering gaps or substitutions such as 1.67, 2.56, missing 2.45a, or a run of modern summary sentences where the source has concrete place names. Compare the beginning and end of each chapter arc before reading every line. 发现方式：先把源文标题与译文标题配对，扫描 1.67、2.56、缺失 2.45a 等编号空洞或替换；再检查源文有具体地名处，译文是否变成现代概括句。逐句细读前，先对照每章开端与结尾的叙事弧线。
• Low-token audit: Use heading counts and exact scans before asking an agent to reread prose, for example rg -n "^## " chapters/src chapters/final, rg -n "1\\.67|2\\.56|2\\.45a|归期|现代|告诉她|愿你" chapters/final chapters/translated, plus a small-context source comparison around the first confirmed drift. 低 token 审计：先用标题计数和精确扫描，不要让 agent 盲读整章，例如 rg -n "^## " chapters/src chapters/final、rg -n "1\\.67|2\\.56|2\\.45a|归期|现代|告诉她|愿你" chapters/final chapters/translated，再只对首个确认漂移点附近做小上下文源文对照。
• Fix by: Restore one target unit per source verse or documented variant, keep the source chapter boundary, and rebuild the target prose from the local verse and its neighboring arc. If a public-domain reference translation is used, treat it as a witness, not as wording to copy. 修复方式：恢复每个源诗节或已登记异文对应一个目标语单元，保住源文章界，并从本节源文与邻近弧线重建中文。若使用公版参考译本，只作 witness，不复制措辞。
• Recheck: Rerun heading-count comparison, scan for invented verse ids, then perform a full-chapter source-to-target pass. The fixing round must be FIXED_RECHECK_REQUIRED; only the following full-chapter zero-issue round can PASS. 复查：重跑标题计数，扫描自创节号，再做整章源文对译文复查。修复轮必须记为 FIXED_RECHECK_REQUIRED；只有随后新增的整章零问题轮才能 PASS。

Over-Explanation and Invented Motive / 过度解释或加戏

• Symptom: The translation explains what the source leaves implicit, supplies motives, modernizes attitudes, or adds emotional judgment not present in the source. 症状：译文解释了原文留白，补动机，现代化人物态度，或添加原文没有的情绪判断。
• Find by: Comparing explanatory phrases, psychological verbs, moral labels, and added causal connectors against the source. 发现方式：把解释性短语、心理动词、道德标签和新增因果连接逐项对照原文。
• Messenger-poem guard: In poems where the source gives a formal message, consolation, or itinerary, do not replace the source's concrete images with modern therapeutic paraphrase such as repeated 告诉她, 不要, 愿你, 归期, or generalized hope language unless those functions are present in the local verse. Treat a run of smooth motivational sentences after a concrete source passage as a candidate invented-addition family. 信使诗防护：信使诗中，源文若给出正式传语、安慰或路线，不要把本节具体意象改成现代心理安慰式概括，例如反复出现的“告诉她”“不要”“愿你”“归期”或泛化希望语；除非本节源文确有这些功能。具体源文后出现一串顺滑励志句，应视作可疑加戏问题族。
• Low-token audit: Start with phrase scans such as rg -n "告诉她|若她|愿你|归期|心意|不要|希望|痛苦没有白白" chapters/final chapters/translated, then inspect only candidate units with the neighboring source verse and glossary rows before widening the audit. 低 token 审计：先扫候选短语，例如 rg -n "告诉她|若她|愿你|归期|心意|不要|希望|痛苦没有白白" chapters/final chapters/translated，再只把候选单元连同邻近源诗节和术语表交给复核，确认后再扩大审计。
• Fix by: Move necessary reader help into a short note when there is real misunderstanding risk; otherwise remove the addition and let the source ambiguity stand. 修复方式：确有误读风险时，把必要帮助移入克制短注；否则删除新增解释，保留原文暧昧或留白。
• Recheck: Ask whether the passage now lets readers infer at the same distance intended by the source. 复查：确认读者推断距离与原文一致。

Terminology Drift Under Readability Pressure / 顺读压力下的术语漂移

• Symptom: A translator varies terms for smoother prose, but the variation changes a concept, institution, case name, identity, ship, place, or fictional science item. 症状：为了顺读变换术语，却改变了概念、制度、案例名、身份、船名、地名或虚构科学设定。
• Find by: glossary/terms.csv, display_policy, forbidden_body_renderings, source-term scans, and chapter control terminology tables. 发现方式：检查 glossary/terms.csv、display_policy、forbidden_body_renderings、源语词扫描和章节控制术语表。
• Fix by: Mark terms as locked, preferred, or context-variable; keep locked terms stable and document allowed variants. 修复方式：区分 locked、preferred 或可随上下文变化的术语；锁定术语必须稳定，允许变体必须记录。
• Concept-triad guard: When a craft book defines a recurring triad, lock the whole set rather than polishing each occurrence independently. For the said / the unsaid / the unsayable, prefer 所言 / 未言 / 不可言说: 所言 avoids the misleading completed-time feel of 已言/已说, 未言 names conscious withheld content, and 不可言说 captures subconscious material beyond words. Audit and forbid drift such as 已说之物, 未说之物, 不可说之物, and 说出的/未说出的/无法说出的 when they are used as the defined terms. 概念三分组防护：技法书定义一组三分概念时，要锁定整组术语，而不是逐处自由润色。the said / the unsaid / the unsayable 宜译为“所言 / 未言 / 不可言说”：所言 避免 已言/已说 的时间完成态误导，未言 指有意识但被扣留的内容，不可言说 对应意识之下、语言不能及的材料。作为定义术语出现时，应审计并禁用“已说之物”“未说之物”“不可说之物”“说出的/未说出的/无法说出的”等漂移译法。
• Defined-concept quotation guard: When a recurring translated concept has a book-specific technical meaning, mark the concept name with target-language main quotation marks in body prose, but not in headings where the heading role already marks the term. In zh-Hans, use Chinese double quotation marks such as “所言”, “未言”, and “不可言说”; reserve single quotation marks for nested quotes. Do not let later polishing strip these marks from explanatory body sentences. 定义概念引号防护：反复出现的译名若在本书中具有专门定义，正文论述中应用目标语主引号标识概念名；小标题处可不加，因为标题功能本身已经标示术语。简体中文用中文双引号，如“所言”“未言”“不可言说”；单引号只用于引号内嵌套。后续润色不得为了“顺眼”把解释性正文中的这些引号删掉。
• Poetic natural-name guard: In classical poetry, bird, animal, and plant names often carry seasonal, erotic, or mythic functions. Do not smooth an uncertain name into a familiar but wrong species, and do not leave a typo-like transliteration in the body. For Sanskrit poetry, exact source triggers such as चक्रवाकी, केतक, केसर, चातक, बलाका, हंस, and सारिका should be audited against glossary/terms.csv, forbidden renderings, and the local image function before release. Confirmed target residues such as 斑鸠, 酮迦, 俱舍罗, 白鹭, or gender drift like 年轻人 for a young woman should trigger a whole-book candidate scan. 诗歌自然物名防护：古典诗歌中的鸟、兽、植物名常承担季节、情欲或神话功能。不要把未判清的名称顺手改成熟悉但错误的物种，也不要在正文留下像错字的音译。梵语诗歌中，चक्रवाकी、केतक、केसर、चातक、बलाका、हंस、सारिका 等源语触发式，应在 release 前对照 glossary/terms.csv、禁用写法和局部意象功能审计。若确认出现 斑鸠、酮迦、俱舍罗、白鹭，或把年轻女子误成“年轻人”等性别漂移，应立即做全书候选扫描。
• Low-token audit for poetic natural names: Scan source and target separately before broad reread, for example rg -n "चक्रवाकी|केतक|केसर|चातक|बलाका|हंस|सारिक" chapters/src and rg -n "斑鸠|酮迦|俱舍罗|白鹭|年轻人|发间穿着" chapters/final chapters/translated output/epub_work/EPUB glossary/terms.csv; then inspect only the matched verse units with nearby source context. 诗歌自然物名低 token 审计：先分开扫描源文和译文，不要直接要求大范围重读，例如 rg -n "चक्रवाकी|केतक|केसर|चातक|बलाका|हंस|सारिक" chapters/src 与 rg -n "斑鸠|酮迦|俱舍罗|白鹭|年轻人|发间穿着" chapters/final chapters/translated output/epub_work/EPUB glossary/terms.csv；随后只复核命中的诗节及邻近源文。
• Recheck: Scan final chapters and generated XHTML for forbidden body renderings and unauthorized source-term parentheses. 复查：扫描终稿章节和生成 XHTML，查禁用写法和未授权原词括注。
• Russian rank and Caucasus place-name guard: For Russian-to-Chinese literary books, close-looking transliterations for military ranks and Caucasus place names must be locked from glossary/terms.csv before release. If a random sample finds drift such as 叶萨乌尔 for locked есаул -> 叶绍尔 or 别什套山 for locked Бешту -> 别什图山, audit final chapters, translated drafts, generated XHTML, glossary, and chapter-control watchlists for both the confirmed bad rendering and nearby phonetic variants. 俄语军衔与高加索地名防护：俄译中文学项目中，军衔和高加索地名常有形近音译，发布前必须以 glossary/terms.csv 锁定。若随机抽检发现 есаул -> 叶绍尔 漂成“叶萨乌尔”，或 Бешту -> 别什图山 漂成“别什套山”，必须同时审计终稿、译稿、生成 XHTML、术语表和章节 control 观察表，覆盖已确认坏写法及邻近音译变体。
• Russian weapon terminology guard: When a Russian glossary locks culturally specific weapons, do not smooth them into a generic familiar weapon name. If шашка is locked as 长刀 and 马刀 is listed as forbidden or over-generalized, scan source triggers (шашка|шашкой|шашку|шашки) and target residues (马刀|长刀) across final chapters, translated drafts, generated XHTML, source text, and the glossary; repair every confirmed target residue to the locked term and record any exceptions. 俄语武器术语防护：俄语术语表锁定具有文化语境的武器名时，不要为了顺口改成泛化常用武器名。若 шашка 已锁定为“长刀”，且“马刀”在术语表中列为禁用或泛化过强，必须同时扫描源语触发词（шашка|шашкой|шашку|шашки）和目标语残留（马刀|长刀），覆盖终稿、译稿、生成 XHTML、源文和术语表；确认命中一律改回锁定译名，并记录合理例外。

Context-Disambiguation Drift / 上下文选义漂移

• Symptom: A polysemous or context-dependent source word is translated with a plausible early meaning, but later context shows the wrong sense. The paragraph may read smoothly, so ordinary readability review misses it. 症状：多义词或依赖上下文的源语词在前文被译成一个看似合理的义项，但后文证明选义错误。句子可能很顺，所以普通顺读审校容易漏掉。
• Risk: Readers receive a false fact, wrong conceptual category, wrong emotional relation, or unstable term family while the prose still looks polished. This is especially dangerous in literary symbolism, historical titles, legal or institutional words, craft-analysis terms, scientific terms, and dialogue idioms. 风险：读者会得到错误事实、错误概念类别、错误情感关系或不稳定术语族，而译文表面仍像成稿。文学象征、历史称谓、法律/制度词、写作技法术语、科学术语和台词习语尤其高危。
• Find by: During translation and chapter control, keep an internal watchlist of source words whose sense depends on later clauses, later paragraphs, speaker identity, plot outcome, or technical definitions. During review, revisit earlier watchlist items after downstream context has been translated. Reader feedback that says "this word choice feels wrong later" should trigger this family immediately. 发现方式：翻译和章节 control 中，对依赖后文从句、后文段落、说话人身份、情节结果或术语定义才能判义的源语词建立内部观察清单。审校时，在后文已译出后回到前文观察项复查。读者反馈“后文看起来这个词选错了”时，立即归入本问题族。
• Fix by: Use skills/expert-translation-quality/SKILL.md. If the sense is unresolved, choose a target phrasing that preserves the source ambiguity instead of forcing a guessed sense. Once later context resolves it, revise the earlier occurrence. If the wrong choice may recur, audit same source trigger, same target residue, glossary rows, and nearby source-to-target pairs. 修复方式：使用 skills/expert-translation-quality/SKILL.md。义项未决时，先用不错误收窄的目标语表达保留原文暧昧，不要强行猜义。后文判清后，回到前文修订。若可能复现，审计同一源语触发式、同一目标语残留、术语表行和邻近源文-译文对。
• Low-token audit: Start with the human or sample-cited source phrase, then search paired source and target text. Prefer exact source triggers and confirmed target residues; avoid asking an agent to reread whole chapters before a candidate pass. 低 token 审计：从人工反馈或样本点名的源语短语开始，再搜索成对源文和译文。优先使用精确源语触发式和已确认目标语残留；不要在候选收集前让 agent 重读整章。
• Spanish ironic truth guard: Early modern Spanish comparisons such as mejor les ayude Dios que ellos dicen la verdad may deny or undercut a rumor instead of affirming that the speakers tell the truth. When verdad appears near mejor ... que, aunque, rumor words, or narrator self-defense, audit the local logic before translating it as 说的是真话. 西语反讽真话防护：近代西语中类似 mejor les ayude Dios que ellos dicen la verdad 的比较式可能是在否认或消解流言，而不是确认说话者“说真话”。当 verdad 与 mejor ... que、aunque、流言词或叙述者自辩同现时，先审计局部逻辑，再决定是否能译作“说的是真话”。
• Sanskrit purpose/negation guard: Purpose compounds and postpositive cause markers can be reversed when a nearby target sentence contains a negation or a speed adverb. For example, śīghra-saṃpāta-hetoḥ means "for the sake of descending quickly"; it should not become a negated instruction such as 不作迅疾降临. Audit target phrases such as 不作, 不要, 勿, 迅疾, 立刻, 迅速, and 先停 against the nearby Sanskrit purpose relation before PASS. 梵语目的/否定防护：梵语目的复合词和原因标记遇到中文否定或速度副词时，容易把方向译反。例如 śīghra-saṃpāta-hetoḥ 是“为了迅速降下”，不能译成“不作迅疾降临”这类否定指令。PASS 前应扫描 不作、不要、勿、迅疾、立刻、迅速、先停 等目标语触发词，并对照邻近梵语目的关系。
• Sanskrit parallelism guard: In Sanskrit lyric passages, repeated case endings and paired compounds can carry the poem's emotional structure. Do not compress a sequence such as body/body, heat/heat, tear/tear, hot-sigh/hot-sigh, longing/longing into a generic 无尽思念; omission may look smooth but breaks the parallel rhetoric and fidelity. Audit sampled failures by searching both source triggers such as अङ्गेन, तप्तं, सास्रेण, उष्णोच्छ्वासं, उत्कण्ठ and target residues such as 无尽思念, 贴近你的思念, 热苦, 泪眼, then restore each parallel member in readable target prose. 梵语平行修辞防护：梵语抒情诗中，反复的格尾和成对复合词常承载情绪结构。身体/身体、热苦/热苦、泪/泪、热叹/热叹、相思/相思这类序列，不能压成“无尽思念”之类总括；译文即使顺，也会破坏平行修辞和忠实度。抽检发现后，先扫源语触发词 अङ्गेन、तप्तं、सास्रेण、उष्णोच्छ्वासं、उत्कण्ठ 与目标语残留 无尽思念、贴近你的思念、热苦、泪眼，再以自然中文恢复每一项平行关系。
• Sanskrit agency/relation guard: Compact Sanskrit verse often implies who benefits from an event through case endings, not through explicit target-language subjects. Do not turn an incidental reaction into a false offering or confuse the beneficiary. For example, when Siddhas honor the cloud because their companions tremble and embrace them at the cloud's thunder, do not translate it as 她们替爱侣献给你的迎礼; rebuild the relation in Chinese before polishing. 梵语施受/关系防护：梵语诗常用格尾暗示事件受益者和施受关系，而不是像中文一样明说主语。不要把附带反应误译成虚假的献礼或弄错受益者。例如悉陀因云雷使爱侣惊拥自己而敬云，不能写成“她们替爱侣献给你的迎礼”；应先把中文关系重建清楚，再追求诗性。
• Sanskrit aphorism hard-shell guard: Short Sanskrit aphorisms often use broad words like sukha, duḥkha, or wheel imagery. A literal target sentence may sound logical in pieces but fail as a Chinese saying, such as 谁的幸福会永远来到. First express the invariant relation (谁的幸福会永远不变？谁的痛苦又会绝对不变？), then keep the source image if it remains readable. 梵语格言硬壳防护：梵语短格言常用 sukha、duḥkha、车轮等宽泛词。逐词译文局部看似可通，合起来却不像中文格言，例如“谁的幸福会永远来到”。应先译出关系不变性，如“谁的幸福会永远不变？谁的痛苦又会绝对不变？”，再保留可读的源文意象。
• Sanskrit fine-sense and sensory-link guard: When a compact adjective or compound is uncertain, do not attach a plausible sensory detail to the nearest target noun just because the Chinese sentence becomes vivid. For sagandha, avoid turning the bird's call into 草木气息的鸣声; for malinavasana, do not weaken grief-marked dirty clothing into merely 穿旧的衣裙. Audit exact source triggers such as सगन्ध and मलिनवसन, plus target residues such as 草木气息, 穿旧, and other adjective-noun pairs where smell, sound, stain, or grief has been reassigned. 梵语细义与感官关系防护：紧凑形容词或复合词未判稳时，不要为了中文生动，把一个看似合理的感官细节顺手挂到最近的名词上。sagandha 不应把鸟鸣写成“草木气息的鸣声”；malinavasana 不应把带有哀伤污痕的衣裳弱化成单纯“穿旧的衣裙”。审计时扫描 सगन्ध、मलिनवसन 等源语触发式，以及“草木气息”“穿旧”等目标语残留，特别检查气味、声音、污痕、哀伤状态是否被错接或削弱。
• Sanskrit bird-compound guard: Do not map a semantic bird compound to the nearest attractive bird species when the source names the bird by behavior or diet. For gṛha-bali-bhujām, the house-offering-eaters are crows, not white cranes; scan the exact compound trigger plus residues such as 白鹤筑巢, 白鹤, 白鸟, and other bird-name substitutions when reviewing Sanskrit nature imagery. 梵语鸟类复合词防护：源语若用行为或饮食义来命名鸟类，不要顺手把语义复合词映射成最顺眼的鸟种。gṛha-bali-bhujām 这种“食家祭者”指的是乌鸦，不是白鹤；审校梵语自然物象时，要扫描这一类精确复合触发式以及 白鹤筑巢、白鹤、白鸟 等鸟名替换残留。
• Sanskrit mythic-name interface guard: Mythic weapons, legendary acts, and eponymous places must keep a reader-facing proper-name interface instead of collapsing into generic description or modern explanation. चन्द्रहास in a Ravana-war context should not become an unmarked 月刃剑; सुरभितनयालम्भजां ... रन्तिदेवस्य कीर्तिम् should not flatten into 以牛祭成就功业. Add or update glossary rows for the name/allusion, scan forbidden generic residues in final Markdown and generated XHTML, and decide whether a short note or a target-readable named rendering is enough. 梵语神话专名接口防护：神话武器、传说行为和因名得义的地名，必须保留读者能识别的专名接口，不能压成普通描写或现代说明。罗波那战场语境里的 चन्द्रहास 不宜无标记地译成“月刃剑”；सुरभितनयालम्भजां ... रन्तिदेवस्य कीर्तिम् 也不应扁平化为“以牛祭成就功业”。修复时补或更新术语表行，扫描终稿 Markdown 与生成 XHTML 中的泛化残留，再决定用目标语可读专名还是短注接口。
• Recheck: The chapter control or closure check must record polysemy_context_review: "PASS" and polysemy_unresolved_count: 0. If a deliberate source ambiguity remains, record why the target wording preserves rather than resolves it. 复查：章节 control 或闭环检查必须记录 polysemy_context_review: "PASS" 和 polysemy_unresolved_count: 0。若原文故意保留暧昧，必须记录目标语为何是在保留暧昧，而不是未判清。
• Borrowed-word mood guard: In Russian literary dialogue, embedded French or European loanwords may name a social mood rather than a literal modern body or temper category. For example, spleen in salon speech should be checked against melancholy, ennui, or depression-like context before rendering it as 脾气. If the confirmed residue is a wrong but smooth target word such as 驱散您的脾气, scan exact residues and nearby synonyms in final/translated chapters before PASS. 外来词情绪义防护：俄语文学对白中的法语或欧洲外来词，常指社交语境中的心绪，而不是现代汉语里的身体或脾性类别。例如沙龙话语里的 spleen 应先按忧郁、烦闷、厌倦等上下文判义，不宜顺手译成“脾气”。若确认残留为“驱散您的脾气”这类顺但错的目标词，PASS 前必须扫描终稿/译稿中的精确残留和邻近同义候选。

Technique-Term and Idiom Interface / 技法术语与成语接口

• Symptom: A craft term, famous line, or English idiom is translated literally enough to become opaque in the target language, even though the exact source form matters for teaching or reader recognition. 症状：写作技法术语、经典台词或英语成语被直译得不透明；而这些地方的源语形式又关系到教学说明或读者识别。
• Find by: Scan sampled low-score units plus glossary/title maps for terms such as on-the-nose, action spine, movement, in-character/non-character, narrating intelligence, fits the ear, fix blame, offer he couldn't refuse, fly-by-night, flattering, run its course, okay?, just very different, codes of address, and idiom-like phrases that appear in quoted or analytical contexts. 发现方式：结合低分样本、术语表和标题映射，扫描 on-the-nose、action spine、movement、in-character/non-character、narrating intelligence、fits the ear、fix blame、offer he couldn't refuse、fly-by-night、flattering、run its course、okay?、just very different、codes of address 等术语/成语，以及出现在引文或分析语境中的惯用表达。
• Fix by: Use a target-readable term first, then keep the source form in parentheses only when the book-local glossary/display policy allows it or when the source phrase itself is a required quote interface. If a glossary forbids the source form in body text, put the source interface in the approved note/glossary layer instead. Render a famous line such as offer he couldn't refuse with the established target phrasing and, in quote-analysis contexts, pair it with the original source line. For craft terms such as action spine / spine of action, do not preserve a literal body term like 行动脊柱; use a target-readable concept such as 行动主线, and recast compounds like "dueling action spines" into a clear target-language relation. For dramatic-structure movement, avoid 运动/运动段落; use 推进段, 动作段, or a context-specific equivalent. For prose-mode in-character/non-character, prefer readable labels such as 人物发声/非人物叙述 or 借人物声音写作, not literal "身份声音" compounds. For narrating intelligence, use 叙述意识 or a context-specific narrator-awareness phrase, not 叙述智能. For idioms and dialogue such as fly-by-night, flattering, and fits the ear, translate the function in the scene: 不靠谱的临时买卖, 太抬举我了, 听起来顺耳, etc.; do not add fraud, excessive emotion, or anatomical literalism. 修复方式：先给目标语可读术语；只有书内术语表/display policy 允许，或源语短语本身是必要引句接口时，才在正文括号中保留源语。若术语表禁止源词进入正文，应把源语接口放到批准的注释/术语表层。offer he couldn't refuse 这类经典台词应采用读者能识别的惯用译法；在引句分析语境里，还要与原句成对呈现。action spine / spine of action 这类技法术语不得硬译成正文里的 行动脊柱，应改为 行动主线 等中文可读概念；"dueling action spines" 这类组合要改成清楚的目标语关系。戏剧结构里的 movement 不要译成 运动/运动段落，应按上下文改为 推进段、动作段 等；散文模式里的 in-character/non-character 宜译成“人物发声/非人物叙述”或“借人物声音写作”这类可读标签，不要硬造“身份声音”。narrating intelligence 宜译为“叙述意识”或按上下文改写为叙述者意识，不要写“叙述智能”。fly-by-night、flattering、fits the ear 这类成语/台词应按场景功能处理为“不靠谱的临时买卖”“太抬举我了”“听起来顺耳”等，不要加出诈骗、过激情绪或器官直译。
• Dialogue idiom guard: In analyzed dialogue, keep the English quote interface when the source wording matters, but translate the function rather than the filler word. okay? inside a frantic challenge may become 行吗？ or another pressure marker, not a polite 好吗？; just very different should usually become a natural understatement such as 只是不太一样, not 只是很不同; codes of address should not be narrowed to address forms only when it means broader interaction or speech norms. 台词成语防护：被分析台词中，若源语措辞本身重要，应保留英文引句接口，但译文要译功能，不要译填充词表面。急促逼问里的 okay? 可译为“行吗？”等施压标记，不宜译成客气的“好吗？”；just very different 通常应处理成“只是不太一样”这类自然低调说法，而不是“只是很不同”；codes of address 若指更广的交往/话语规范，不应收窄为只关乎称谓的“称呼规范”。
• Medical idiom guard: Do not translate a colloquial source illness phrase into a target-language folk diagnosis that sounds medically wrong or unintentionally comic. In a quoted line such as I'm just one stomach flu away from my goal weight, use a functional and recognizable phrase such as 一次肠胃炎, not 肠胃感冒 or a less natural collocation such as 一场肠胃炎. 医学俗语防护：不要把源语里的俗称病名译成目标语里听起来医学上不稳、还会意外搞笑的民间诊断。例如 I'm just one stomach flu away from my goal weight 这类台词，可按功能译为“一次肠胃炎”，不要写成“肠胃感冒”，也不要用“一场肠胃炎”这类搭配不够自然的表达。
• Blame/relationship-loss guard: In craft analysis, do not let compact English abstractions become false target-language actions. fix blame means determining or assigning where responsibility lies, not necessarily pursuing punishment; public transportation portrait may be a bus-side ad image rather than an abstract portrait; mutual loss in relationship analysis is a shared loss borne by both people, not 相互失落. 责任/关系损失防护：写作技法分析中，不要把紧凑英语抽象词译成错误的目标语动作。fix blame 指确定或归置责任，不一定是“追究责任”；public transportation portrait 可能是公交车身广告头像，不是抽象的“公共交通画像”；关系分析里的 mutual loss 是双方共同承受的损失，不是“相互失落”。
• Crowd/collective noun guard: In political or theatrical dialogue, do not translate mob mechanically as 暴民 when the scene means the audience, populace, crowd, or Roman people whose favor grants power. Use 民众, 大众, 观众, or a context-specific collective unless the source clearly stresses riotous violence or contempt. 群体名词防护：政治或戏剧台词中的 mob 不要机械译成 暴民。若语境指掌声、拥戴和政治权力来源中的观众/大众/罗马民众，应按上下文译为 民众、大众、观众 等；只有原文明示暴乱、暴力或强烈蔑称时才用 暴民。
• Foreign-language social phrase guard: When the source intentionally leaves French, Latin, Italian, or another prestige/social language in a Russian or multilingual literary scene, the EPUB body still needs a reader-facing interface. A full foreign sentence should not stand alone without a target-language rendering or an immediate note marker, and snippets such as pour mazure should be rendered in target-language action terms such as 跳玛祖卡 with the approved note interface if the foreignness matters. 外语社交短语接口防护：俄语或多语文学场景故意保留法语、拉丁语、意大利语等社交/身份语言时，EPUB 正文仍必须给读者可用接口。整句外语不能无中文译文或即时脚注标记地单独出现，pour mazure 这类片段应转成“跳玛祖卡”等目标语动作表达；若外语身份功能重要，再按批准的注释接口保留说明。
• Object-term self-collision guard: When a source term names a material object with modifiers, do not stack incompatible target terms in one noun phrase. For Russian двойной лорнет, 双筒单筒望远镜 is self-contradictory; choose a coherent reader-facing object term such as 双筒长柄望远镜, then scan exact residues and nearby object names (单筒望远镜, 双筒望远镜, лорнет) across final chapters, translated drafts, and generated XHTML. 器物术语自撞防护：源语器物名带修饰语时，不要把互相冲突的目标语器物词硬叠成一个名词短语。俄语 двойной лорнет 不能写成“双筒单筒望远镜”这类自相矛盾的名称，应改为“双筒长柄望远镜”等读者能理解的器物名；随后扫描“单筒望远镜”“双筒望远镜”“лорнет”等残留，覆盖终稿、译稿和生成 XHTML。
• Russian clothing/object metaphor guard: In Russian literary caricature, clothing parts may be described with concrete texture words such as щетина around подгалстушник; do not carry the animal/object surface literally into Chinese if it makes the image self-collide. A sampled phrase like 硬领的鬃毛 should trigger scans for 硬领的鬃毛|鬃毛|硬领衬|щетина|подгалстушник across final chapters, translated drafts, generated XHTML, source text, and sample manifests, then rebuild the confirmed body text into a coherent garment image such as 硬挺的领边托着他的下巴. 俄语衣饰/器物比喻防护：俄语文学讽刺描写中，衣饰部件可能带 щетина、подгалстушник 这类质感词；若直搬成中文会让衣饰变成动物意象或器物自撞，不得保留词面硬壳。抽样命中“硬领的鬃毛”时，应扫描 硬领的鬃毛|鬃毛|硬领衬|щетина|подгалстушник，覆盖终稿、译稿、生成 XHTML、源文和抽样 manifest；确认正文命中后改成“硬挺的领边托着他的下巴”等连贯衣饰图像。
• Russian Caucasus culture-term guard: Culture-specific terms in Russian frontier or Caucasus scenes must follow the project glossary instead of drifting to generic modern social words. If калым is locked as 聘金, scan source triggers (калым|калыма|калыму|калымом) and target candidates (聘礼|聘金|彩礼) across final chapters, translated drafts, generated XHTML, source text, glossary, and random samples. Repair confirmed reader-facing калым renderings to the locked term and record any documented exceptions. 俄语高加索文化术语防护：俄语边疆/高加索场景中的文化专词必须按项目术语表统一，不要漂成泛化或现代口语词。若 калым 已锁定为“聘金”，应同时扫描源语触发词（калым|калыма|калыму|калымом）和目标语候选（聘礼|聘金|彩礼），覆盖终稿、译稿、生成 XHTML、源文、术语表和随机样本；确认读者正文中对应 калым 的命中后改为锁定译名，并记录合理例外。
• Low-token audit: Use rg -n "顶在鼻尖|on-the-nose|行动脊柱|故事脊柱|行动主线|相互决斗|固定责任|追究责任|确定责任|无法拒绝的条件|无法拒绝的提议|受宠若惊|奉承|一个运动|两个运动|运动段落|运动段|第二个运动|解决运动|人物身份声音|非人物身份声音|角色身份声音|人物内声音|叙述智能|这个智能|人物专属声音|人物声音这一侧|以角色身份|贴合耳朵|骗一把就跑|一夜就跑|不靠谱的.*买卖|跑完自己的路程|公共交通画像|拳击家|相互失落|关系的失落|弧贯|讲解员凯末尔|浪漫凯末尔|暴民|逗暴民|mob is Rome|肠胃感冒" chapters/final chapters/translated output/epub_work/EPUB, then classify hits as confirmed defects, established terms, famous-line interfaces, or reasonable exceptions. If source expansion is needed, search rg -n "fix blame|share blame|place blame|public transportation portrait|pugilists|mutual loss|narrating intelligence|this intelligence|stomach flu" chapters/src. 低 token 审计：可用 rg -n "顶在鼻尖|on-the-nose|行动脊柱|故事脊柱|行动主线|相互决斗|固定责任|追究责任|确定责任|无法拒绝的条件|无法拒绝的提议|受宠若惊|奉承|一个运动|两个运动|运动段落|运动段|第二个运动|解决运动|人物身份声音|非人物身份声音|角色身份声音|人物内声音|叙述智能|这个智能|人物专属声音|人物声音这一侧|以角色身份|贴合耳朵|骗一把就跑|一夜就跑|不靠谱的.*买卖|跑完自己的路程|公共交通画像|拳击家|相互失落|关系的失落|弧贯|讲解员凯末尔|浪漫凯末尔|暴民|逗暴民|mob is Rome|肠胃感冒" chapters/final chapters/translated output/epub_work/EPUB 收集候选，再区分为确认问题、稳定术语、经典台词接口或合理例外。需要源语扩展时，搜索 rg -n "fix blame|share blame|place blame|public transportation portrait|pugilists|mutual loss|narrating intelligence|this intelligence|stomach flu" chapters/src。
• Rebuild guard: After repairing a literal technique term such as 行动脊柱, scan both source Markdown and generated EPUB XHTML. If Markdown is clean but an old EPUB still contains the bad term or old quote layout, rebuild before validation; stale staging is a reader-facing regression, not a reasonable exception. 重建防护：修复 行动脊柱 这类技法术语硬译后，必须同时扫描源 Markdown 和生成 EPUB XHTML。若 Markdown 已清但旧 EPUB 仍含坏术语或旧引文排版，必须先重建再验证；旧 staging 污染是读者可见回归，不是合理例外。
• Recheck: Read the revised paragraph as a target-language teaching paragraph without the source. The term should be intelligible before the source form helps it. 复查：不看原文，只把修订段当目标语教学段阅读；术语应先能读懂，源语形式只起辅助作用。

Title and Subtitle Overload / 标题与小标题超载

• Symptom: The translation preserves source title chains, original names, parenthetical source terms, or explanatory subtitles in headings where target-language title design should stand alone. 症状：标题保留原文长链、原名、括注原词或解释性副题，而目标语标题本应独立成立。
• Find by: Reviewing nav.xhtml, title maps, chapter headings, subtitles, and mobile wrapping risk. 发现方式：检查 nav.xhtml、标题映射、章题、副题和手机换行风险。
• Fix by: Separate nav_title, display_title, subtitle, and optional title_note; put source names or first-mention notes in the first natural body occurrence, not the heading. 修复方式：拆分 nav_title、display_title、subtitle 和可选 title_note；原名或首次出现注释放正文第一次自然出现处，不塞进标题。
• Recheck: Confirm headings carry structure and tone, not terminology explanations. 复查：确认标题承载结构和声调，而不是术语解释。
• Heading/body boundary guard: A title can look acceptable in wording while still swallowing the first body sentence or paragraph, especially after OCR cleanup, chapter splitting, or mechanical title normalization. Treat any # heading that is unusually long, contains a full narrative sentence after a part marker, or appears in nav.xhtml with body prose as a reader-facing defect family, not a cosmetic Markdown issue. Audit with a cheap heading-length scan such as ^# .{80,} over chapters/final, chapters/translated, and generated XHTML, then split the body prose back below the heading and rerun coverage plus EPUB navigation checks. 标题/正文边界防护：标题文字本身可能看似可接受，却把正文首句或首段吞进标题行；这常发生在 OCR 清理、章节切分或机械规范章题之后。凡 # 标题异常过长、在“第几部分/第几节”后直接接叙事句，或在 nav.xhtml 中把正文带入目录，都应归为读者可见问题族，不要当作普通 Markdown 外观问题。先用低成本标题长度扫描，如在 chapters/final、chapters/translated 和生成 XHTML 中查 ^# .{80,}；确认后把正文拆回标题下方，并重跑覆盖率和 EPUB 导航检查。

Retrospective Hook / 复盘要求

Added Analytical Recap and Coverage-Aware Repair / 加戏性分析总结与覆盖率修复

• Symptom: A translated paragraph or standalone reader-facing unit adds a synthesis sentence that is not present in the source, often introduced by cues such as 换言之, 这说明, 由此, 这也解释, 作者/吉本在这里, or a broad institutional/political recap. The sentence may sound plausible, but it changes the unit from translation into commentary. 症状：译文段落或独立读者可见单元加入源文没有的综合判断，常见提示包括“换言之”“这说明”“由此”“这也解释”“作者/吉本在这里”，或把事实叙述升格为制度、政治、作者意图总结。句子可能看似合理，但已经从翻译变成评论。
• Find by: Start from random-sample blockers, then scan final chapters, translated drafts, generated XHTML, and sample manifests for the cue phrases. Treat broad cues as candidates, not automatic defects; verify each hit against the local source context before removal. 发现方式：先从随机样本 blocker 出发，再在终稿、译稿、生成 XHTML 和抽样 manifest 中扫描这些提示语。宽泛提示词只能作为候选，不能自动判错；每个命中必须对照局部源文后再决定。
• Risk: Unsupported recap alters fidelity and can make the book read like study notes. If the chapter is close to the coverage floor, deleting the added sentence can expose under-translation and fail coverage gates. 风险：无源总结会损害忠实度，使成书读起来像读书笔记。若章节本来贴近覆盖率下限，直接删除加戏句还会暴露欠译，导致覆盖率门禁失败。
• Fix by: Remove or replace the unsupported interpretation. When coverage drops, do not restore the commentary merely to satisfy the gate; add a short source-backed factual bridge or translate the omitted source material at the same location. Keep the new text factual: who acted, what happened, what consequence the source states. 修复方式：删除或替换无源解释。若覆盖率下降，不要为了过门禁恢复评论句；应在同一位置补入有源文依据的事实连接，或翻译被遗漏的源文内容。新增文字只写事实：谁行动、发生了什么、源文说明了什么后果。
• Low-token audit: Use exact-string scans for cue phrases and sampled sentences, then inspect only those paragraphs plus nearby source. Record candidates kept as justified exceptions when they are terminology explanations, unit conversions, or editor-note paraphrases supported by source. 低 token 审计：用精确字符串扫描提示语和样本句，再只检查命中段落及邻近源文。若候选属于术语解释、计量换算或有源文支撑的编者注释改写，应记录为合理例外。
• Classical-history trigger expansion: In long historical translations, sampled invented recaps may be fluent one-sentence or whole-paragraph bridges around warfare, legitimacy, religion, or institutions. Add exact candidates such as 戏剧性宣言, 政治宣言, 横跨南北, 战略视野, 这种胜利迅速耗尽人心, 这个细节说明, and 宗教裂痕又使 to the first-pass scan, then delete only source-unsupported hits and preserve source-backed historical judgments as exceptions. 古代史触发词扩展：长篇历史译本中的加戏总结，可能是围绕战争、正统性、宗教或制度的流畅单句，也可能是整段桥接说明。首轮候选扫描可加入 戏剧性宣言、政治宣言、横跨南北、战略视野、这种胜利迅速耗尽人心、这个细节说明、宗教裂痕又使 等精确短语；随后只删除无源文支撑的命中，把源文本身支持的历史判断记录为例外。
• Logistics-word exception: A word such as 补给线 is not itself proof of added commentary in military history. Treat it as a candidate only when it appears inside a standalone recap or modern analytical bridge; record source-supported logistics descriptions as exceptions after checking the nearby source passage. 后勤词例外：补给线 这类词本身并不能证明加戏。只有当它出现在独立总结段或现代分析桥接段中时才作为候选；若邻近源文确实在叙述通道、粮食、交通、围城消耗或撤退路线，应记录为合理例外。
• Recheck: Rerun translation coverage after repairs, then rebuild EPUB and rescan generated XHTML. A repaired random round remains non-PASS and must be followed by a new-seed spot-check. 复查：修复后重跑 translation coverage，再重建 EPUB 并扫描生成 XHTML。发生修复的随机抽检轮仍为非 PASS，必须追加新 seed 复抽。

Note-Marker Interface Leak / 注码接口泄漏

• Symptom: Source footnote numbers remain as bare body digits or bare note-definition prefixes, or inline note markers keep source-like spaces (text [12] text) that make the EPUB read like ordinary numbers instead of note links. 症状：源文脚注编号残留为正文裸数字或注释定义裸编号，或者正文注码保留源语式空格（如 正文 [12] 正文），使 EPUB 中的注码读起来像普通正文数字，而不是清晰的注释入口。
• Caret-source guard: Also scan for source-caret note residue such as inline ^3611 or line-start ^3611 note text; convert inline references to [3611] and definitions to [^3611]: note text, then recheck that no non-definition ^number residue remains. Caret 源注码防护：还要扫描 ^3611 这类源格式残留；正文引用改为 [3611]，注释定义行改为 [^3611]: note text，并复查不再存在非脚注定义形式的 ^数字 残留。
• Bare dotted definition guard: Also scan for line-start dotted note definitions such as 129. note text; convert confirmed note definitions to [^129]: note text, then recheck ^\s*\d{1,5}[.]\s+ across final Markdown, translated drafts, and generated XHTML. 裸点号定义防护：还要扫描 129. note text 这类行首点号注释定义；确认属于注释定义后改为 [^129]: note text，并在终稿 Markdown、译稿和生成 XHTML 中复查 ^\s*\d{1,5}[.]\s+ 不再命中。
• Bare inline reference guard: Build the note-definition number set for each chapter, then scan non-definition body text for those numbers when they appear immediately after sentence or list punctuation, such as text. 42 in source-like spacing, 正文。42, or 正文；211. Convert confirmed note references to [42] or [211], and record exceptions for years, page numbers, chapter numbers, quantities, dates, regnal numbers, and quoted source numerals. 正文裸数字引用防护：逐章建立注释定义编号集合，再扫描非定义正文中紧跟句末或列举标点的同编号，例如源语式空格里的 text. 42、中文正文里的 正文。42 或 正文；211。确认属于注码后改为 [42]、[211]；年份、页码、章节号、数量、日期、纪年号和引文内源文数字应记录为例外。
• CJK-adjacent reference guard: After fixing punctuation-adjacent references, also scan the same note-definition number set when a candidate follows a CJK word and is followed by whitespace, punctuation, or line end, such as historian29, 法规471、[48], or 服从129 普世会议 in target text. Treat these as the same body-reference family and convert only confirmed note numbers to bracketed markers. 中文词后引用防护：修复标点后裸注码以后，还要用同一章注释定义编号集合扫描“中文词后 + 空白/标点/行尾前”的候选，例如目标语正文中的 古代史家29、法规471、[48]、服从129 普世会议。这仍属同一正文注码问题族；只把确认属于注释编号的候选改成方括号注码。
• Find by: Compare the chapter's note-definition numbers with inline body occurrences. Use structural scans over final Markdown, translated drafts, generated XHTML, and chapter-control/gate reports; do not rely only on EPUBCheck, because a visually bad bare number can still be technically valid text. 发现方式：先取得本章注释定义编号，再扫描正文中同编号的裸露出现。审计范围应覆盖终稿 Markdown、译稿、生成 XHTML 以及章节控制/门禁记录；不要只依赖 EPUBCheck，因为视觉上错误的裸数字仍可能是技术上合法的正文。
• Fix by: Convert confirmed body note references to the approved marker form ([12]) and convert note-definition prefixes to the same stable form before the note label. Remove spaces around inline markers in CJK body text, but keep readable spacing after definition markers when the book's note style requires it. 修复方式：把确认属于注码的正文引用统一为批准的标记形式（如 [12]），把注释定义行的裸编号也改为同一稳定形式。中文正文中的行内注码前后不留空格；若书籍注释定义样式需要，可保留定义标记后的可读空格。
• Paragraph guard: Also scan whether note-definition lines are attached to the preceding body paragraph. If a definition marker starts immediately after body text without a blank line, insert a blank line before it so the note remains a separate reader-facing block and paragraph-coverage gates do not hide the problem. 分段防护：还要扫描注释定义行是否贴在前一正文段后。若定义标记紧接正文、前面没有空行，应在定义行前补空行，使注释成为独立的读者可见块，也避免段落覆盖门禁掩盖问题。
• Multiple-note paragraph guard: If one reader-facing paragraph begins with a note marker and later contains another marker such as （注159） or （注167a）, split each confirmed note into its own paragraph. Otherwise random sampling may merge unrelated notes or attach editor commentary to a following body unit. 多注释合并防护：若一个读者可见段落以注释标记开头，后面又含 （注159）、（注167a） 等另一个注释标记，应把每条确认注释拆成独立段落。否则随机抽样会把互不相干的注释合并，或把编者补注误并到后续正文单元。
• Low-token audit: Build the candidate number set from actual note definitions in each chapter, then scan only those numbers as standalone candidates. Record false positives such as years, page numbers, chapter numbers, and quoted source numerals as exceptions. 低 token 审计：先从每章实际注释定义建立候选编号集合，再只扫描这些编号的独立出现。年份、页码、章节号、引文中的源文数字等误报，应作为例外记录。
• Recheck: Rerun the bare-definition scan, bare-body-reference scan, and inline-spacing scan after fixes; if any chapter was changed, append a non-PASS fix round and then a fresh full-chapter zero-issue recheck before release. 复查：修复后重跑裸定义、正文裸引用和行内空格扫描；凡章节被修改，都必须先记录不能 PASS 的修复轮，再追加最新一轮全章零问题复查，方可 release。

Foreign Quote Interface and Source Residue / 外文引文接口与源格式残留

• Symptom: A long Latin, French, Greek-transliterated, Italian, or other source-language quotation remains reader-facing without a target-language explanation, or the repair leaves production wording such as 中文接口：... in the EPUB body. Related OCR/source residues include glued question marks (?Lala, 页?Et), orphan ASCII brackets (。], M.], ]), or old [ note text ] wrappers around note definitions. 症状：拉丁文、法文、希腊转写、意大利文等长外文引文直接留在读者正文中，没有目标语说明；或修复时把 中文接口：... 这类制作标签留进 EPUB 正文。相邻问题还包括 OCR/源格式残留，如粘连问号（?Lala、页?Et）、孤立 ASCII 方括号（。]、M.]、 ]），以及脚注定义里的旧式 [ 注文 ] 外壳。
• Find by: Scan final Markdown, translated drafts, and generated XHTML at paragraph level, not only line level, because Markdown verse lines can merge into one long XHTML paragraph. Use long Latin-script runs plus missing target cues (引文大意, an actual target-language translation, or a real note marker), and separately scan source-residue or pseudo-interface patterns such as 原文续作, 续文为：, 中文接口：, English interface, Original citation, 外文引句补足此处所述, 其大意为：这段外文引文用以印证, 按：外文原句保留作版本, \?[A-Za-z], [A-Za-z]\?[A-Za-z], \?{2,}, terminal ], and note-definition wrappers. Record page numbers, years, book/letter numbers, and intentional source uncertainties as exceptions. 发现方式：必须按段落扫描终稿 Markdown、译稿和生成 XHTML，不能只按行扫；因为 Markdown 诗行会在 XHTML 中合成一个长段。用长 Latin-script 片段加缺少目标语提示（引文大意、真实目标语译意或真实脚注标记）定位候选；另行扫描源格式残留或伪接口模式，如 原文续作、续文为：、中文接口：、English interface、Original citation、外文引句补足此处所述、其大意为：这段外文引文用以印证、按：外文原句保留作版本、\?[A-Za-z]、[A-Za-z]\?[A-Za-z]、\?{2,}、行尾 ] 和脚注定义外壳等。页码、年份、书信编号和有意保留的源文疑读要记录为例外。
• Bibliographic citation-interface guard: In history notes, raw page/pages, Latin source abbreviations followed by target-language page labels, and machine bridge phrases such as 此注指向...出处：, 目的在于说明...编注, or 吉本在此... are reader-facing source residue when they replace a finished translated note. Scan page|pages|此注指向|出处：|目的在于说明|吉本在此|吉本在这里|编注 across final chapters, translated drafts, generated XHTML, and random-sample manifests; then convert confirmed hits to target-language citation prose or remove the unsupported bridge. 书目引文接口防护：历史注释中，裸露的 page/pages、拉丁书目缩写后接中文页码，以及 此注指向...出处：、目的在于说明...编注、吉本在此... 这类机器桥接语，如果替代了完成态译注，就是读者可见源格式残留。应在终稿、译稿、生成 XHTML 和抽样 manifest 中扫描 page|pages|此注指向|出处：|目的在于说明|吉本在此|吉本在这里|编注；确认命中后改成目标语引文句式，或删除无原文支撑的桥接评论。
• Citation-tail corruption guard: Bibliographic tails can be damaged by OCR or quote-boundary repair even when the note is otherwise translated: examples include empty quote parentheses, smart quotes after page numbers (165“, 475”), comma-only chapter markers (Evagrius (,40)), broken title words (Annal. for Animal.), dropped initials in titles or authorities (Menagiana -> enagiana, Greg. Acropol. -> reg. Acropol., Mascou -> ascou), missing initials in quoted proper names (M. Anicius Gallus -> . Anicius allus), damaged Latin quotes (Multis incivilibus gestis Gallus Cæsar -> ultis incivilibus gestis allus Cæsar), broken foreign terms (Mediomatrici -> ediomatrici), page joins such as Olearius,”553, and lost volume/chapter ranges such as Hist. de C. P. l. vi. c. 22-46 collapsing to Hist. de C. P. -26. Scan exact damaged fragments from failed samples plus generic patterns for quote marks adjacent to page numbers and citation ranges reduced to a bare hyphen-number; normalize only the confirmed citation tail, preserving the source authority and edition data. 引文尾部损坏防护：即使注释主体已经译好，书目尾部也可能因 OCR 或引号边界修复而损坏，例如空引号括号、页码后残留智能引号（165“、475”）、只剩逗号的章节标记（Evagrius (,40)）、书名词损坏（Annal. 应为 Animal.），以及 Olearius,”553 这类页码粘连。应先扫描失败样本中的精确坏片段，再扫页码旁异常引号等泛化模式；只修确认命中的书目尾部，同时保留原有作者、版本和页码证据。
• Truncated quote and dropped-initial guard: A random sample that ends with an unfinished foreign fragment, or shows one-letter loss in a Latin-script name/term (Moab -> oab, Germans -> ermans, Maria -> aria, Murad -> urad, Morad -> orad), should trigger exact bad-fragment scans across chapters/final, chapters/translated, and rebuilt XHTML. Also scan for nested parenthetical names such as 目标名（目标名（Source））. Verify each candidate against the same source note before restoring the missing initial or simplifying the name interface; do not batch-repair ordinary substrings where the bad fragment is part of a correct word. 截断引文与首字母脱落防护：抽样若发现外文片段在句末截断，或拉丁字母姓名/术语丢失一个首字母（如 Moab 变 oab、Germans 变 ermans、Maria 变 aria），应把精确坏片段扫过 chapters/final、chapters/translated 和重建后的 XHTML。还要扫描 目标名（目标名（Source）） 这类嵌套括注。每个候选都先对照同一源注再补首字母或简化姓名接口；坏片段若只是正确单词的普通子串，不得批量替换。
• Round-derived scan extension: when one dropped-initial source name or raw quote interface is confirmed, include exact scans for nearby prose glue and short damaged terms such as 写道 意思, 贺拉斯描述为当地所有, ons Seleuc, anor 是, adchar, and adschar in chapters/final, chapters/translated, rebuilt XHTML, and the current random-sample manifests. Restore only candidates confirmed against the matching source note. 轮次补充：确认一个首字母脱落或外文接口残留后，应把抽样 manifest 中的相邻坏短语也纳入精确扫描，例如 写道 意思、贺拉斯描述为当地所有、ons Seleuc、anor 是、adchar、adschar；扫描范围包括 chapters/final、chapters/translated、重建 XHTML 和当前抽样 manifest。只修复能与对应源注核实的候选。
• Fix by: Do not expose production labels, and do not replace a missing translation with a generic pseudo-gist such as 外文引句补足此处所述... or 这段外文引文用以印证.... If the quotation matters to the argument, provide a concrete target-language gist of the quoted words; if nearby target prose already explains it, remove the redundant pseudo-interface rather than adding another formula. For source residue, restore the quote boundary and citation placement instead of leaving 原文续作 or 续文为： in reader-facing text, remove old ASCII note wrappers while preserving note numbers, and keep deliberate source uncertainties only when they are visually intelligible to readers. 修复方式：不要把制作标签暴露给读者，也不要用 外文引句补足此处所述...、这段外文引文用以印证... 这类泛化伪大意替代真实翻译。若引文本身承担论证功能，必须补出该引文具体说了什么；若附近中文已经说明其作用，应删除多余伪接口，而不是再套公式。源格式残留应恢复引文边界和出处位置，不得把 原文续作 或 续文为： 留在读者文本里；旧 ASCII 注文外壳要去掉但保留脚注编号；只有读者能理解的源文疑读才可保留。
• Copy-neighbor guard: When batch-repairing many notes, never reuse or shift a gist from the previous or next note just because the Latin-script pattern looks similar. Each added 引文大意 must be checked against the exact quote on that line or verse block; copied-neighbor gists are fidelity defects, not harmless interface text. 相邻复制防护：批量修复大量注释时，不得因为拉丁字母形态相似就复用或错挪前后注释的大意。每条新增 引文大意 都必须对照本行或本诗段的确切引文；复制相邻注释的大意属于忠实度缺陷，不是无害的接口文字。
• Orphan-fragment and OCR guard: When random review finds a lone Latin-script word or a verse line with obvious OCR residue, treat it as the same foreign-quote family, not as a one-line typo. Scan for orphan line fragments such as a standalone Nomine, duplicated adjacent 引文大意, malformed citation bridges such as 出处：...能够辨出 ..., double target punctuation after quote repair, and likely OCR variants inside preserved Latin/French/Italian lines (Nune for Nunc, marmorels for marmoreis). Repair the quote boundary, add a concrete target-language gist when the quotation remains, and record justified manuscript-corruption exceptions separately. 孤立残片与 OCR 防护：随机抽检若发现单独一词的拉丁字母残片，或诗行里有明显 OCR 残留，应归入同一外文引文接口族，而不是当成单行错字。需扫描孤立行残片（如单独的 Nomine）、相邻重复 引文大意、畸形出处桥（如 出处：...能够辨出 ...）、引文修复后的双重中文标点，以及保留拉丁文、法文、意大利文诗句中的疑似 OCR 变体（如 Nune 应为 Nunc、marmorels 应为 marmoreis）。修复时恢复引文边界；引文保留时补具体中文大意；传本残缺或疑读必须单独记录为例外。
• Audit-unit guard for source-only verse: If a Latin or other source-language verse is preceded by a Chinese gist but separated into its own paragraph, the random sampler may correctly expose the verse as a source-only reader unit. Merge the source line and its concrete target-language meaning into the same reader-facing paragraph, or use adjacent source/target quote blocks where both blocks are complete and intelligible. 源语诗行审计单元防护：如果拉丁文或其他源语诗句前一段已有中文大意，但诗句本身被分成独立段落，随机抽检仍会正确地把它暴露为 source-only 读者单元。应把源句与具体中文含义合并到同一个读者可见段落，或采用完整且可理解的源语块/译文块相邻版式。
• Legal/historical phrase guard: Short Latin legal or antiquarian phrases can still be blocking even when surrounded by Chinese prose, because they may carry the evidence itself. If a sample keeps phrases such as ad jus corporis ... ecclesiarum non hominum singulorum pertinentia, ubi sacrificia mactantur, or Coliseum ... ob stultitiam Romanorum ... ad calcem deletum, scan exact Latin triggers across final chapters, translated drafts, generated XHTML, and sample manifests. Keep the Latin only when useful as evidence, but add a concrete Chinese meaning immediately nearby, such as "属于教会团体而非个人", "在那里献祭牺牲", or "大半被烧成石灰".
• Short quote interface extension: Even a two-word Latin evidence phrase such as intacti tormentis can fail a reader-facing audit if no Chinese meaning follows. When the retained source phrase carries the note's proof, add an immediate compact gloss (即...) and record source-backed exceptions separately from untranslated residue. 法律/古物短语防护：短拉丁法律语或古物学短语即使夹在中文段落中，也可能是阻塞问题，因为它们承担的正是证据内容。若样本保留 ad jus corporis ... ecclesiarum non hominum singulorum pertinentia、ubi sacrificia mactantur、Coliseum ... ob stultitiam Romanorum ... ad calcem deletum 等短语，应扫描终稿、译稿、生成 XHTML 和抽样 manifest 中的精确拉丁触发词。拉丁原文可作为证据保留，但附近必须补具体中文含义，如“属于教会团体而非个人”“在那里献祭牺牲”“大半被烧成石灰”。
• Recheck: Rebuild the EPUB before claiming closure, then scan generated XHTML again for both source-residue patterns and generic pseudo-gist phrases. A Markdown-only PASS is insufficient when the defect appears only after paragraph merging. If a random spot-check finds one sample, classify the full family, audit the whole book, write the fix round as non-PASS, and open the next seed only after a zero-hit recheck. 复查：声称闭环前必须重建 EPUB，并再次扫描生成 XHTML，覆盖源格式残留和泛化伪大意短语。若缺陷只在段落合并后出现，Markdown 层 PASS 不足以放行。随机抽检只要命中一个样本，就归纳为全族问题并全书审计；该修复轮不能记 PASS，必须零命中复查后再开启新 seed。

Round-Derived Spotcheck Guards / 抽检轮次派生防护

• Unsupported analytical-bridge triggers: When samples expose unsupported historical synthesis, add exact target phrases from the failed paragraphs to the whole-book scan, including 整体胜势, 正规军力, 这种辩护的力量, and 这里“有教益”显然. Treat these as candidates, not automatic defects: remove only sentences whose claim, evaluation, or causal bridge is absent from the source or note, and preserve source-backed summaries as exceptions. 无支撑分析桥触发词：抽样若发现无源文支撑的历史综合句，应把失败段落中的精确目标语短语纳入全书扫描，例如 整体胜势、正规军力、这种辩护的力量、这里“有教益”显然。这些只作为候选，不是自动缺陷；只有当句子的判断、评价或因果桥接在源文/源注中没有依据时才删除，源文支持的概括应记录为例外。
• Split-continuation guard: If the source footnote marker or OCR text interrupts a single note, do not invent a new editor note such as 注137a merely because the continuation follows a number. Compare the source sentence across the marker, merge confirmed continuation text into the original note, and retain only genuinely separate source/editor notes. 续文误拆防护：若源文脚注标记或 OCR 文本把同一条注释切开，不得因为续文跟在数字之后就虚构 注137a 这类新编注。必须跨标记核对源句，把确认属于同一注释的续文并回原注；只有真实独立的源注/编注才保留单独编号。
• Bibliography-tail OCR extension: Notes can fail through a damaged authority name or citation tail even when the surrounding Chinese is fluent. Add exact scans for sampled fragments such as Wesseling, in Praefat. and Not. ad, Pyrenæan ountains, and broken currency/citation joins like 罗马 穆拉托里第二十八论文（solidi）; repair the bibliographic interface in target-language prose while preserving source authority, page, and edition evidence.
• Bibliographic authority hybrid extension: Mixed forms such as 德 Thou or raw Shaw, Travels, 第...页 are citation-interface defects when the note otherwise uses target-language bibliography. Scan exact sampled authority fragments plus nearby source forms (Thuanus, Shaw, Marmol), normalize confirmed hits to target-language authority prose, and keep Latin-script names only in parentheses or where the edition identity requires them.
• Bibliography-title initial-loss extension: Damaged German or Latin title tails such as eschichte des..., chichte des..., Röm. Rechts, or raw Savigny, eschichte... can be reader-facing OCR residue. Scan exact failed title fragments plus nearby authority names (Savigny, Hullman), normalize confirmed hits to target-language bibliographic prose, and record legitimate retained source titles as exceptions. 书目尾部 OCR 扩展：即使周围中文流畅，注释也可能因为权威名或引文尾部损坏而失败。应把样本中出现的精确坏片段纳入扫描，例如 Wesseling, in Praefat. and Not. ad、Pyrenæan ountains，以及 罗马 穆拉托里第二十八论文（solidi） 这类货币/引文粘连；修复时用目标语书目句式整理接口，同时保留作者、页码和版本证据。
• Proper-name drift guard: If one authority name has multiple target renderings, lock the chosen form in the active glossary and scan old variants across final Markdown, translated drafts, generated XHTML, and random manifests. For example, after choosing 唐维尔 for D'Anville, scan and reject stale variants such as 当维尔, 丹维尔, 德昂维尔, 昂维尔, and nested replacements such as 唐维尔（唐维尔）. 专名漂移防护：同一权威姓名若出现多个目标语译名，应把选定形式写入当前术语表，并在终稿、译稿、生成 XHTML 和抽样 manifest 中扫描旧译名。例如确定 D'Anville 译作 唐维尔 后，应扫描并拒收 当维尔、丹维尔、德昂维尔、昂维尔，以及 唐维尔（唐维尔） 这类机械替换嵌套。
• Quote-meaning bridge guard: A source quote may keep a concise original-language fragment, but the target-language meaning must read as finished note prose. Scan production-like bridges such as ：意思是：, 意思是：, 说：意思是：, 表达 意思是, and failed-replacement residue such as ??? across final chapters, translated drafts, generated XHTML, and sample manifests. Convert confirmed hits to finished phrasing such as 其意为：, direct quotation, or a normal citation sentence; retain ordinary phrases like 这句话的意思是 only when the whole unit is already fluent and intelligible.
• Analytical-bridge trigger addendum: Phrases such as 意义，在于, 意义在于, and 本注的意义在于 are high-risk candidates for target-only explanatory synthesis in history notes or recap paragraphs. Audit each hit against the nearby source before changing it; delete unsupported inserted commentary, but preserve source-backed uses when the source itself states the evaluation or consequence.
• Duplicate transition guard: A short standalone paragraph that repeats a transition already translated in the previous source-aligned paragraph is a target-only duplicate, even if it is fluent. Scan exact sampled transition phrases and run a nearby-duplicate pass for short paragraphs whose opening also appears in the previous few paragraphs; delete only confirmed misplaced duplicates after source comparison.
• taught false-friend guard: In historical prose, English taught to tremble and obey means "made/conditioned to fear and obey", not 教会 as an institution. Scan exact Chinese residues such as 被教会颤抖, 教会颤抖, and 颤抖和服从; preserve source-backed expressions like "born to tremble and obey" when no taught false friend is present.
• Note/body boundary guard: If a source note ends and body prose resumes, do not let the note paragraph swallow the following body paragraph. Audit samples where a note contains a long body-like argument after a citation tail, especially around exact triggers such as 君士坦丁对基督徒..., and restore the source note/body paragraph boundary before rebuilding.
• Source-quote compression guard: If a sampled note reduces a verse or source quote to analytical paraphrase (用...名...铺陈, 并列举...诸族, 蛮族洪流突然冲破), compare with the source quote. Either translate the concrete images in a marked 诗句大意/quote-gist sentence or preserve the complete source quote with adjacent target-language meaning; do not replace poetic evidence with a modern analytical summary.
• Round 359 source-interface addendum: Treat sampled residues such as 空气性本质, 朴素装备, 任性残酷, 若干色调, 不按任何规则方法被评论, 正统的筛子, 阿卜德拉梅, 阿卜德拉赫曼或, and 今天你把我作为题目 as whole-book scan candidates. Classify confirmed hits as source-syntax residue, metaphor false literalization, proper-name interface drift, or foreign-quote interface drift; fix them with idiomatic target-language prose while preserving the source authority or original-language form in parentheses when needed.
• Round 360 note/prose interface addendum: When a sample exposes mixed raw authority abbreviations, dangling quote bridges, or modern analytical recap prose, scan exact triggers such as Willerm. Tyr., fronte recedant. Imperii, （回注），其意为, qui ab imperatore fuerat, 请听欺诈而非欺诈, 这段法律文句说明, 这里最尖锐之处在于, 这里最可怕的不是, 表面是信任，实则, 风暴无效消散, and 烦扰选择. Repair confirmed hits by either giving an immediate Chinese gloss with source form in parentheses, normalizing the bibliography to target-language author/title/volume prose, deleting unsupported target-only analytical bridges, or rebuilding source-syntax residue into idiomatic target-language narration.
• Round 365 note/source-interface addendum: Treat damaged continuation markers, note/body boundary drift, source-only Latin legal or verse phrases, raw bibliography tails, invented explanatory notes, and broken legal-status phrases as one reusable random-spotcheck family cluster. Start with exact scans for the sampled residues, for example 注3a, stylum in pyxide。, 原句为 Marmaricus, Sapherinas pelles。, Suethans（, Heyne, Excurs, Nardini, 第, （注70） 在记述这场战争, （注67）然而, 骑士财产等级或秩序者, and target-only synthesis such as 维系记忆、礼仪和穷人救济; then compare only matched small contexts with the source. Repair by making Chinese reader interfaces immediate, restoring source note numbers and body boundaries, deleting unsupported target-only synthesis, and normalizing legal-social classifications. Rebuild EPUB and scan generated XHTML before closing the round; repaired issue rounds still require a new seed.
• Round 366 stale-sample and micro-residue addendum: If a random sample exposes an exact target string that no longer exists in current chapters/final, chapters/translated, or generated XHTML, record it as a stale sample residue and verify with exact scans before editing. For confirmed current hits, scan and repair sampled strings such as 启示碎片, 按论证推进, 没有区分所有责任, 注100a）a, 让纳比见加尼耶, Dicanice，, and awkward label words such as 责名. Delete target-only synthesis, remove duplicated/misaligned paragraphs, normalize "apud"/transmission relations as 转引自, delete stray OCR letters after note markers, and add target-language readings for retained foreign terms. Rebuild and scan final/translated/generated outputs; repaired rounds remain non-PASS and need a fresh seed.
• Round 367 source-form and explanatory-bridge addendum: Treat sampled forms such as Coenum allicanum, 经过修饰的勉强, and standalone geography/history bridges such as 黎巴嫩山地的自然条件解释了马龙派的政治命运 as whole-book audit triggers. Compare retained foreign proper names against source/control evidence so dropped initials or damaged title tails are corrected, rewrite ritual or idiomatic source phrases into natural target-language prose, and delete target-only explanatory synthesis unless the source itself supplies the claim. Rebuild, scan final/translated/generated outputs, and keep the repaired round non-PASS before opening a fresh seed.
• Round 368 note-tail, theology, and boundary addendum: Treat sampled strings such as 三者分别提供编年叙事, 不差别, 他传教士的匕首, Hist. de C. P. , 29, and paragraph-start fragments such as 喘息； as whole-book triggers. Delete source-adjacent analytical note tails that merely summarize already translated authority evidence, replace opaque theological calques with readable target-language doctrinal interfaces, distinguish Assassin/Ismaelian agents from ordinary gospel missionaries, preserve bibliographic book/chapter ranges such as l. v. c. 28, 29, and merge paragraph splits that start a reader-facing unit mid-sentence. Rebuild and scan final/translated/generated outputs before a new seed; repaired rounds remain non-PASS.
• Round 369 proper-name role gloss, schism-object calque, and analytical-bridge addendum: Treat sampled residues such as Peter Bulgarian chief, literal abstractions such as 分裂首先争夺对象, and source-adjacent explanatory bridges such as 军人特权之间作出选择 or 这一退场正是惩罚的实质 as whole-book audit triggers. Remove raw role labels from retained foreign names when the Chinese sentence already supplies the role, recast compressed ecclesiastical-political phrases into readable target-language relationships, and delete unsupported body-level interpretation that turns a note gloss or narrative consequence into an added argument. Rebuild, scan final/translated/generated outputs, and keep the repaired round out of the PASS streak.
• Round 370 OCR-title, name-variant, source-syntax, and boundary-compression addendum: Treat sampled strings such as Occulus Pastoralis, Benaud、Reginald 或 Arnold de Chatillon, 自然自发的慷慨, 可耻必要：同一个反叛臣民交战并惩罚他, and target-only paragraph joins such as 标枪贯穿了他的身体。 叙利亚之所以值得争夺 as whole-book triggers. Verify apparent source OCR spellings against local context and established names, normalize title/name variants into a reader-facing Chinese interface, rewrite abstract source noun chains into idiomatic target prose, and delete or split target-side synthesis that fuses separate source contexts. Rebuild and scan final/translated/generated outputs before opening the next seed; repaired rounds remain non-PASS.
• Round 371 source-term residue, orphan editor marker, and recap-synthesis addendum: Treat sampled strings such as bare pagans in Chinese note prose, orphan 按。 tails, target-only explanatory note sentences like 这些例证反复出现, and duplicate recap paragraphs such as 波斯驻军、地方顺从者和宗教压力 or 盖利默在第十里碑败退后 as whole-book triggers. Translate operative source terms first and retain the foreign word only as parenthetical evidence, normalize editor attributions to ——编者按。, delete unsupported note-tail synthesis, and remove duplicate recap paragraphs when the following source-aligned sequence already translates the material. Rebuild and scan final/translated/generated outputs before a fresh seed; repaired rounds remain non-PASS.
• Round 372 quote/OCR, polysemy, and note-boundary addendum: Treat sampled residues such as 人，以及永远！, 这一句把戴克里先, 贞德 for chastity, damaged authority names like oguet, damaged quote starts like urgite, and source-sample compression such as 读者也许会喜欢原始语调的一个样本 without any original-language sample as whole-book audit triggers. Repair by giving the retained foreign quote an immediate concrete target-language meaning, restoring source names and initials against chapters/src, deleting unsupported analytical note bridges, correcting context-bound false friends, and replacing confirmed orphan note-tail semicolons with full stops. Keep ordinary explanatory phrases or semicolons only after source/context review, record them as exceptions, rebuild, and scan final/translated/generated outputs before opening a fresh seed; repaired rounds remain non-PASS.
• Round 373 source-interface and quote-contamination addendum: Treat sampled residues such as raw chronology labels (Olymp. xcv. 4), dropped initials in retained source terms (irtha for Girtha, e miserum for Me miserum), verse contamination across adjacent notes (defluxit Orontes inserted into a different Juvenal quote), bare hostile labels (Ariomanites), figure placeholders without labels, and Latin or French quotation strings without immediate target-language meaning as one whole-book audit family. Scan exact triggers first, then related retained-source forms such as Transversâ, quam sanæ, synodaliter, commendams, and Mecreant. Repair by restoring the source form against chapters/src, adding a concise Chinese meaning or role gloss in the same reader-facing unit, replacing generic figure placeholders with descriptive labels, and deleting unsupported target-only interpretation rather than converting it into commentary. Rebuild and rescan final/translated/generated outputs before opening a fresh seed; repaired rounds remain non-PASS.
• Round 374 hard-literal history, law, theology, and note-boundary addendum: Treat sampled residues such as 用刀剑和弓弦处死法, 把眼睛和思想导向, 把法律的真实变成了我们的真实, 生于旁非利亚的西德, 匈人凭箭预取胜机, 可售或自愿的美色, 禁止永久法人占地诸法, 法国统马官, 圣灵单独发出或双重发出, 原文为： followed by a blank paragraph, and raw citation tails like vol. ii. or damaged Latin agrun modum as whole-book audit triggers. First scan exact target strings in final Markdown, translated drafts, rebuilt XHTML, glossary/forbidden-rendering files, and the random manifests; then inspect only the matched paragraph plus nearby source before deciding whether it is a defect or an exception. Repair by converting metaphorical or institutional source phrasing into idiomatic target-language prose, keeping precise legal/theological terms with a reader-safe Chinese interface, restoring source names or Latin forms against chapters/src, and merging quote openings with their retained original-language quote and target gist so the reader-facing unit is complete. Rebuild and rescan final/translated/generated outputs before opening a fresh seed; repaired rounds remain non-PASS.
• Round 375 reader-interface, title-chain, and citation-form addendum: Treat sampled residues such as 经过私人保证的巧妙缓和, 赌博债传给, 喜悦骚动一平息为注意, lowercase or damaged source terms like standalone iaour, visible spacing around target-language quotation marks (国家的 “显赫” 尊位), bibliographic raw tails such as Smith's Inquiry into the Wealth of Nations, vol. i., citation authority leakage such as raw Evagrius, dangling note wrappers like （回注）［, and old printed synopsis chains using repeated 。—— as whole-book audit triggers. Scan exact sampled strings first, then related interface patterns in chapter-opening synopsis lines, notes, manifests, final Markdown, translated drafts, and rebuilt XHTML. Repair only confirmed hits by turning hard source syntax into idiomatic target prose, normalizing title/authority references to target-language citation form, removing visible quote-spacing and orphan wrappers, restoring source terms against chapters/src, and rewriting old title chains as readable target-language synopsis sentences. Record substring false positives such as iaour inside correct Giaour as exceptions, rebuild, and keep repaired rounds out of the clean PASS streak.
• Round 376 name-apposition, empty-quote, and place-name addendum: Treat sampled residues such as malformed divine or boundary appositions (神圣界神特尔米努斯 for the god Terminus, the sacred boundary), empty quote-gist bridges (“” 意为), corrupted historical place names (推阿提拉 for Thyatira), visible note-marker spacing after inline note numbers, and target-language quote spacing such as 从 “他的” 无知 or “他们” 暴露 as reader-interface audit triggers. Scan exact bad strings first across final Markdown, translated drafts, rebuilt XHTML, and current random manifests; then compare only matched small contexts against chapters/src and existing in-book glossary/name usage. Repair by making appositions explicit in Chinese, restoring omitted source quotations before their gist, normalizing place names to established in-book forms with the source name in parentheses when helpful, and removing confirmed reader-visible interface spaces. Broad note-marker scans can be noisy in legacy or mojibake draft material; record them as candidates and repair confirmed sampled/current hits rather than batch-rewriting unrelated lines. Rebuild and open a fresh seed after repair.
• Round 377 note-boundary, duplicate-fragment, and unsupported-note-tail addendum: Treat sampled residues such as sentence splits around a footnote marker (骄傲铭文 separated from 向后世宣告), orphan note labels that stand alone without note text, duplicated note or quote fragments (机器使它们仿佛自发运动 repeated after the full note), raw source-language title interfaces (decrees of Lyons), malformed learned appositions (地理学家王者威廉), unsupported explanatory tails (正文所据的要点, 目力或消息仍可迅速抵达), and reader-facing notes or sentences ending in a dangling comma as defect-family triggers. Scan exact sampled strings first in final Markdown, translated drafts, rebuilt XHTML, and the round manifest; then audit nearby note-number neighborhoods and same-chapter note blocks before broad regex work. Repair by restoring the source-backed note/body boundary, translating or removing raw title interfaces, converting learned epithets into idiomatic apposition, deleting unsupported note-tail commentary, and closing dangling punctuation. Preserve real note content when an orphan label points to a source footnote; do not merely delete the marker unless the source has no corresponding note text. Rebuild, validate the repaired round, and require a fresh seed because the repaired round cannot count as clean PASS.
• Round 378 explanatory-bridge and truncated-example addendum: Treat polished but source-unsupported bridge sentences as P2 when they add analytical framing, logistics, moral evaluation, ethnogenesis, or reader-guiding note interpretation, even if the Chinese prose is fluent. Start with exact sampled triggers such as 道路不再只受, 赎金、通行费和战利品, 税制问题因此不再只是, 源文同时保存, 政治与语言共同体, 非战斗人口, 制度的慈悲固然悦目, and 正对应正文; then scan related high-risk markers like 源文, 对应正文, 可贵之处, 换言之, 也就是说, and 这说明 only as candidates. Compare each matched small context with chapters/src; delete confirmed target-only commentary, but preserve legitimate author/editor notes and source-backed explanations. If a unit ends in a setup phrase such as 以下例子可以说明, repair by restoring the omitted source example, not by merely deleting the setup. Rebuild and open a new seed because a round with these repairs is FAIL_REPAIRED.
• Round 379 bibliography, false-friend, orphan-continuation, and recap-tail addendum: Treat damaged bibliography titles (iddle Ages for Middle Ages), note endings like 米尔曼注。这个注释, false-friend renderings of taught as 教会, orphan source-quote continuations such as a standalone fines Imperii Italici or 见恩诺狄乌斯, and recap tails such as 共同功德, 这些经历解释了, 速度造成震慑, 同一个外在动作, 令人不安，又确实有效, or 战法预演 as current-run audit triggers. Scan exact strings in final Markdown, translated drafts, and generated XHTML after rebuild; then scan related title fragments, dangling editor labels, and explanatory verbs only as candidates. Repair by normalizing bibliographic titles into target-language citation prose, restoring missing citation endings, translating source verbs by function (taught -> 教导), joining orphan quote/citation continuations back under their note label, and deleting source-unsupported recap tails. Preserve source-backed explanatory conclusions when the source itself states the judgment.
• Round 047 literary-readability, index-interface, and false-friend addendum: Treat sampled hard-literal opening prose, source-only mottoes or Latin verse split away from their Chinese meaning, bare bibliography titles, index headwords such as IMPORTS or BUCKRAM'D, context-bound false friends such as loose stones rendered as 松石, and legacy source note symbols such as bare [*] as whole-book audit triggers. Start with exact target strings from the failed samples, then scan nearby family terms in final Markdown, translated drafts, generated XHTML, glossary/terms, forbidden-rendering files, chapter controls, and the random manifests. Repair only confirmed current hits by rewriting source-syntax chains into natural target-language prose, putting retained foreign/source titles and their concrete target-language meaning in the same reader-facing unit, translating index headwords by their local function while preserving the source word in parentheses, and deleting obsolete source note symbols when numeric or structured note references already carry the citation. Record stale samples, source-supported foreign titles, and intentional bibliographic source evidence as exceptions; rebuild, rescan, and open a fresh seed because the repaired round remains non-PASS.
• Round 048 truncated-boundary, statistic-format, and source-title addendum: Treat sampled reader units cut off at the end of a chapter, paragraphs beginning with a sentence fragment after an erroneous blank line, malformed source numbers such as 14. 994, citation-title truncations such as Etymological Diet., and hard-literal aphorism or index quotations as whole-book audit triggers. Scan exact bad strings in final Markdown, translated drafts, generated XHTML, and adjacent source split files; if the continuation lives in the next source unit, restore only the missing tail needed to complete the current reader sentence and record the cross-unit source evidence. Repair by joining broken sentence paragraphs, normalizing quoted statistics against source (14,994), correcting citation abbreviations conservatively, and turning source wit into intelligible target-language prose while preserving the cited headword. Rebuild and open a fresh seed; repaired rounds remain outside the clean PASS streak.
• Round 049 audit-unit boundary, table, and publishing-term addendum: Treat sampled source-only motto paragraphs, outline notes that keep a dangling opening quote across multiple reader units, paragraphs ending with a colon/semicolon before the actual answer or verse, Markdown table separators made from em-dash-like glyphs, and publishing terms such as cancel rendered as 取消页 as whole-book audit triggers. Start with exact failed strings, then scan final Markdown and translated drafts for source-only Latin paragraphs, ^| —— table separators, paragraphs ending in ： or ； before a following quote/answer, and glossary-backed publishing terms (cancel, cancels, cancel leaf). Repair by putting retained source-language mottoes and their concrete Chinese meaning in the same reader-facing unit, making outline/list fragments complete without dangling quote punctuation, joining or rephrasing setup paragraphs with their answer when the random unit would otherwise be incomplete, converting table separators to standard Markdown pipes, and using reader-safe book-production terms such as 替换页 or 替换一页. Rebuild, rescan exact bad strings, and open a fresh seed because the repaired round is not clean PASS.
• Round 050 quote/note boundary, source-blank, and account-table addendum: Treat sampled quoted-letter paragraphs that open with “ but do not close, source blank placeholders such as 收入约为——, financial account tables where source DR./CR. or separator rows are flattened into false totals, and note markers glued to the previous note while the note body starts in the next paragraph as whole-book audit triggers. Start with exact sampled strings, then scan final Markdown, translated drafts, generated XHTML, and manifests for 收入约为——$, dangling sampled quote endings, 见前文.*\[[0-9]+\], table labels such as 合计|总计|利润分配|贷方收入, and source terms such as CR.|DR.|Profit on the edition. Repair by making each reader-facing unit complete without inventing missing source data, moving note markers to their note bodies, and splitting or labeling account subtotals so totals apply to the correct group. Broad quote and note-marker scans are candidates only; record normal multi-paragraph quotations, ordinary cross-references, and already complete poem interfaces as exceptions. Rebuild and open a fresh seed because the repaired round is not clean PASS.
• Round 051 setup-payload, license-tail, bibliography, and index addendum: Treat sampled paragraphs that end with an introducer such as 下面这封信：, 题名如下：, 图景：, or 把他说成： while the required title, letter, verse, or quotation sits in the next paragraph as audit-unit boundary defects. Also treat Gutenberg license tails ending at 只要遵守商标许可, raw bibliography entries such as NEWTON, Thomas, Bishop of Bristol，《Works》, and impossible collapsed index references such as 1422,170 as whole-book audit triggers. Start with exact sampled strings and compact scans for 只要遵守商标许可$|遵守商标许可……|NEWTON, Thomas|Bishop of Bristol，《Works》|1422,170|92, 1422; then inspect nearby bibliography and index entries before changing them. Repair confirmed setup defects by joining the setup and payload or making the setup sentence self-contained, repair truncated license boilerplate with the established complete wording, normalize only confirmed raw bibliography entries into target-language citation form, and split collapsed references only when source context supports the separation. Broad 《Works》 and colon-introducer scans are candidates, not automatic defects; record complete lists, normal retained source titles, and legitimate line numbers as exceptions. Rebuild and open a fresh seed because the repaired round is not clean PASS.
• Round 052 cross-file continuation, source-example interface, bibliography, and index-OCR addendum: Treat a reader-facing audit unit as defective when a dedication, letter, poem, or quotation begins in one generated unit and its essential payload or closing formula is stranded at the start of the next unit; this includes endings such as 同时既可爱又伟大。, 下面几行诗...：——, or 如下信件。 before the actual quoted/poetic content. Also treat retained English source examples as defects when the Chinese reader cannot see the point of the example in the same unit, and treat raw bibliography lines such as DRYDEN, John or OCR-like index references such as ii. 1I as audit triggers. Start with exact sampled strings, then scan compact candidates: dangling setup endings, next-unit quote openings, retained source lines such as For,——This is the sergeant, raw SURNAME, Name，《Title》 bibliography rows, and 1I|I1|1422,170 index fragments. Repair by joining or summarizing the missing payload in the same unit, adding an immediate Chinese gloss for retained source examples, normalizing only confirmed raw bibliography entries into target-language citation form, and correcting OCR references only when the source context supports the correction. Broad bibliography appendices and intentionally retained source quotations are exceptions only if they already carry a clear target-language reader interface. Rebuild, scan final Markdown and generated XHTML, and open a fresh seed because any repaired round remains FAIL_REPAIRED.
• Round 053 bibliography-candidate and audit-unit boundary addendum: When random spotcheck repeatedly finds raw bibliography rows, split the work into confirmed hits and candidate pools. Confirmed hits include sampled rows and direct cross-references such as ALMON, John, CHAMBERS, Ephraim, DOBLE, C. E., KING, William, Archbishop, SWIFT, Jonathan, TWINING, Rev. Thomas, and WARTON, Rev. Thomas; repair them with a Chinese reader-facing name/title plus source form in parentheses where useful. Broad bibliography appendices can contain many candidate rows and should be recorded with counts before further cleanup rather than silently treated as PASS. For setup/payload and letter-closing defects, repair sampled and same-family confirmed hits by joining the introduced poem, quotation, title, or closing signature into the same reader-facing unit, or by naming the payload in the setup sentence. Ordinary multi-paragraph quotation style is a candidate exception when the surrounding block remains readable and no introducer, payload, or signature is stranded. Rebuild, scan generated XHTML for exact failed strings, and open a fresh seed because this round is FAIL_REPAIRED.
• Round 054 account/table, setup bridge, index headword, and verse-shape addendum: Treat account-book remnants such as Caution not Repay'd / Battells not discharg'd, setup lines ending in 如下文字：, 想想原文：, 背诵道：, 插在这里：, 声音：, table bridges such as 额外开支，如下：, letter closings stranded before a signature, all-caps index parentheticals such as BURNEY, Mrs. / CAMPBELL, Rev. John of Kippen / CORBET, Andrew, OCR-like references such as iii. it, and verse flattened into prose by inline double spaces as current-run audit triggers. Start with exact sampled strings, glossary/title/index scans, and narrow rg candidate pools; repair confirmed hits by converting account rows to a table, making setup units self-contained or explicitly pointing to the following table/quote, joining closing formulas with signatures, normalizing sampled index headwords to Chinese reader-facing forms with source form in parentheses where useful, correcting source-supported OCR references, and preserving poem lineation with explicit Markdown line breaks. Broad index and bibliography scans are candidate pools with many legitimate exceptions; record candidate counts and fix only confirmed reader-facing defects. Rebuild, scan final Markdown and generated XHTML for exact failed strings, and open a fresh seed because this round is FAIL_REPAIRED.
• Round 055 repeated audit-unit boundary and corrupted Latin addendum: When consecutive random rounds keep finding paragraphs that end with a naked introducer, widen from sampled strings to a whole-book candidate scan for 开头是：, 轶事：, 警句：, 辩解：, 几句话：, 之后：, 撰写：, 结果：, 显著证据.*的：, and similar setup endings. Fix confirmed hits by either joining the short payload/signature into the same paragraph or making the setup paragraph self-contained with a concrete summary of the promised poem, anecdote, inscription, legal argument, or cross-file verse. For long letters, do not force every paragraph into one huge block; instead repair sampled open-letter units by adding a concise in-letter roadmap and joining true closing/signature splits. Treat corrupted-looking source quotations such as coronAc candescente ferreAc coronatur as P1 when they have no reader gloss; correct source-supported OCR/encoding damage and add an immediate Chinese meaning. Rebuild, scan both final Markdown and generated XHTML, and use a fresh seed; repaired boundary rounds never count toward consecutive PASS.
• Round 056 boundary persistence and index candidate-pool addendum: If a fresh seed still finds setup endings such as 我的理由简要如下：, 原先在这一节之前有一节放纵的诗：, 内容也实在够傻：, or 加里克回信如下：, repair both the sampled line and obvious same-local setup line in the same passage by naming the payload content in the setup unit. For quoted songs or poems split across several paragraphs, join the short lyric into one reader-facing unit when the split itself is the blocker. For source-index headwords, fix sampled all-caps residue such as DEGENERACY OF MANKIND to a quieter source form, but treat broad volume-6 all-caps parenthetical scans as candidate pools unless the Chinese headword is unclear or the source form dominates the reader surface; record the candidate count and confirmed repairs. Continue fresh-seed sampling until a whole round is clean, because repeated boundary persistence means the current run still has 0 consecutive PASS rounds.
• Round 057 bibliography, verse-lineation, foreign-quote, and index-boundary addendum: When a fresh seed samples bibliography rows such as ORTON, Job, PILKINGTON, James, THACKERAY, W. M., LANDOR, Walter Savage, O'LEARY, Rev. Arthur, or BOWEN, Emanuel as P2 source-interface residue, repair confirmed rows into a Chinese reader-facing citation: Chinese name/title first, with the source name/title preserved in parentheses where useful. If the appendix still has broad raw-bibliography candidates, record the candidate pool separately from confirmed defects. Treat poems collapsed into one prose paragraph, retained English/Latin quotations without same-unit Chinese meaning, and volume-index lines long enough to be truncated by sampling as current-run audit triggers. Repair poem lineation with Markdown hard breaks, add immediate Chinese glosses for short retained source quotations, and split only index/reference lines into shorter complete reader-facing units; do not split ordinary prose merely because it is long. Rebuild and open a fresh seed because this round is FAIL_REPAIRED.
• Round 058 cross-file setup, stale-sample, and encoding-marker addendum: If a setup paragraph is the last paragraph of a file and the payload begins in the next unit, repair by summarizing the next unit's payload in the setup paragraph rather than making the reader wait across files. For short quoted payloads, join the quote into the setup unit; for long poems or letters, add a concise content bridge and keep the full source payload nearby. If a sampled bad string no longer exists in current chapters/final, chapters/translated, or generated XHTML, record it as stale sample text with exact zero-hit scans rather than editing unrelated text. Treat visible source-name mojibake such as DAILL脡 as P1/P2 and repair to a source-supported form such as Daillé; broad mojibake-marker scans are candidate pools and require context before bulk editing. Rebuild and use a fresh seed because this round is FAIL_REPAIRED.
• Round 059 reader-unit aggregation, bibliography-interface, and foreign-quote addendum: If repeated random samples expose setup paragraphs ending in ： or ：——, quoted-letter paragraphs that are only one part of a continuous letter, or overlong volume-index paragraphs, first check whether the sampler is using true reader-facing audit units rather than isolated Markdown paragraphs. Repair the extraction rule so short setup+payload, continuous quoted text, and long index strings become intelligible review units; do not hide real defects by sampler changes alone. Separately scan confirmed text defects: raw bibliography rows that start with all-caps source-name order, title-only English bibliography rows, retained Latin/Italian/French snippets without same-unit Chinese meaning, table rows where English account labels dominate, poem lines flattened into prose, and broken metaphor literalizations. Fix confirmed hits with a Chinese reader interface while preserving source forms in parentheses where useful, add immediate target-language glosses for retained foreign quotes, remove table source-label residue, and restore poetic lineation. Rebuild and open a fresh seed because this round is FAIL_REPAIRED.
• Round 060 quoted-block structure addendum: When a sampler or review unit normalizes multi-paragraph letters, diary extracts, document headings, or short verse into a single run-on line, it can create false quote-boundary blockers and can also reveal real layout defects. Audit both layers. The reader-unit extractor should keep Markdown line breaks in sample text and merge consecutive paragraph blocks until Chinese quotation marks balance, so an addressee, date, salutation, signature, or document title is not reviewed alone. Then still repair confirmed target-text issues such as English month-order calques (第十月, 第一月) or verse lines flattened after a colon. Rebuild, smoke-test the sampler with the failed seed, remove temporary sample output, and open a fresh seed because the repaired round is not a clean PASS.
• Round 061 foreign-gloss, dangling setup, and inscription-layout addendum: When a random sample exposes a retained Latin/French/English quote, source-title form, variant phrase, or monument/verse block, verify that the same reader-facing audit unit contains the Chinese meaning and enough payload to stand alone. Scan exact failed strings first (Dum memor, quel sot ose, in company, title variants, paragraphs ending in 如下版本： or cross-unit wording such as 下一单元开头即引), then inspect small contexts against the source before changing. Repair confirmed hits by adding immediate Chinese glosses, turning missing quote payloads into self-contained note prose, normalizing reader-facing title forms in the active glossary, and restoring long inscriptions into readable paragraphs with punctuation instead of flattened source-layout residue. If the issue is caused by sampling an English quote apart from its following 其意为 paragraph, repair the reader-unit extractor as well as the text surface. Rebuild, smoke-test the failed seed, and treat the repaired round as FAIL_REPAIRED; only a later fresh seed can count toward PASS.
• Round 062 verse-lineation and bibliography-interface addendum: If a sampled poem or translated verse is flattened into a paragraph with inline double spaces, scan the full book for non-table inline double-space poetry candidates, then separate confirmed poems into Markdown-visible lines while recording tables, errata, account rows, and foreign-quote-plus-gloss lines as exceptions. For setup sentences that promise a following answer, letter, or verse but end before the payload, either merge the payload into the reader unit or rewrite the setup as a self-contained summary. For bibliography appendices, distinguish confirmed sampled entries that deserve Chinese-facing name/title translations from broad original-language bibliography rows; the latter should be explicitly labelled as original bibliographic records for version checking rather than left as 作者原名 prose. Rebuild and open a fresh seed because the repaired round is not a clean PASS.
• Round 063 foreign-verse lineation extension: Do not treat a retained source-language poem as closed merely because a Chinese gloss follows. If the source witness has verse line breaks, preserve a visible source-line shape in the same reader-facing unit before the gloss. Add scans for quoted Latin/Italian/French/English verse blocks that contain a gloss marker such as 意为 but still flatten multiple source lines into one run-on line; repair by restoring line breaks and keeping the Chinese meaning adjacent.
• Round 064 cross-file setup and long-epitaph extension: A paragraph at the end of a unit that points to the next unit with 原话如下, 读者可自行判断, or a quoted compliment ending in a colon is still a blocker even if the payload exists in the next file. Rewrite the current unit so it contains a concise summary of the promised payload, and leave the full quote in the next unit. For long Latin epitaphs or inscriptions, retaining the source layout is acceptable only when the same reader-facing unit also gives a compact Chinese summary or translation.
• Round 065 index/verse interface extension: Treat index units and short verse quotations as reader-facing prose, not harmless apparatus. Scan sampled strings plus exact same-family triggers such as iii. iii., malformed proverb variants (Quos Deus null, Quem Deus vult perdere), abnormal spaces after Chinese punctuation in index entries, and poems flattened into one paragraph with inline spaces. Repair confirmed hits by normalizing citation markers for the target reader, restoring source-supported foreign quote forms or preserving source oddities with clear Chinese context, removing spacing noise from index prose, and preserving visible verse lineation. Record source-text oddities as reader-interface normalizations rather than silently reproducing confusing forms; rebuild and open a fresh seed because the repaired round is not a clean PASS.
• Round 066 letter/bibliography/place-name interface extension: If a sample exposes only a letter heading, letter closing, or setup paragraph (致...先生, 我永远是，我亲爱的先生，, 我在这里给出一篇：, 他说，他——, or a colon introducing a poem/excerpt), make the current reader unit self-contained by adding a concise payload summary or by merging the closing formula with the signature. For retained Latin poems or epitaphs, add a target-language summary before the original as well as any following full gist, so a sampled source block is not foreign-only. For bibliography rows and index headwords, put the Chinese reader label first (伏尔泰, 普鲁士腓特烈二世, 挑翻) and keep the source headword in parentheses only as evidence. For source spellings that are historical variants of an established place (Islam for Ilam), keep one target name and explain the source spelling as 原文作 ... rather than creating a second Chinese place name.
• Round 067 high-impact style, polysemy, and index-interface extension: If a fresh seed catches stiff preface prose, false-friend headwords, or long index units, treat them as book-wide candidate families, not isolated polish. Scan exact failed target phrases such as 这样一种状态, 人类恩主, 颇为少见的小书, 领主的合理性, 喂坏, 滴温酒, slash-flattened verse, and mechanical index fragments such as 缺少温和, 做手势，反感, 幽灵：见幽灵, or 于某事多于另一事. Repair confirmed hits by rewriting high-impact opening prose into natural target-language paragraphs, choosing context-bound legal/food/emotional terms (酌定, 喂养不良, 慰心甘露), preserving verse lineation with XHTML-safe poem blocks when Markdown hard breaks collapse, and splitting or smoothing only the sampled long index units whose Chinese interface is unclear. Merge dangling letter closings with signatures in the same reader unit. Rebuild, scan generated XHTML for the rejected strings, and open a fresh seed because the repaired round is not a clean PASS.
• Round 068 note/index boundary extension: If a random sample begins with an index continuation or ends with a source quote/setup before the payload, repair the reader-facing unit rather than treating it as harmless pagination. Scan exact failed fragments such as 由霍姆博士答复，同上, ^关于约翰生的轶事, 莫根，莫里斯，。, 传种（PROPAGATE）, and setup phrases like 下面两行诗 or retained source quotes such as Nihil tetigit quod non ornavit. Fix confirmed hits by restoring the missing headword in the current unit, moving cross-file continuation material back under its headword, replacing damaged punctuation-only entries with complete source-backed entries, choosing a non-crude target headword (传播理解力 rather than biological 传种), and adding an immediate Chinese meaning for retained Latin or other source quotes. Rebuild and rescan final Markdown plus generated XHTML for the rejected fragments before opening a fresh seed.
• Round 069 literary/interface addendum: If a fresh seed exposes opening-prose stiffness, dangling setup/payload splits, retained foreign verse without same-unit target meaning, long index continuations without headwords, or flattened poem/song lines, treat them as one reader-facing interface family. Scan exact failed target strings first, including setup endings such as 说道：, 以下诗行——, 鲍斯威尔注——, foreign quote witnesses such as Et variis albae, title/OCR forms such as Dietionarius, index continuation starts such as ^v. 349 or 日期，忽视, metaphor false literals such as 用小册子打我, and verse slashes or inline song runs. Repair confirmed hits by moving short payloads into the same unit, adding immediate Chinese gist for retained Latin/foreign verse, restoring or prefixing the current index headword with （续）, splitting only overlong index/reference paragraphs at natural headword boundaries, rewriting high-impact opening prose into idiomatic target language, and preserving poem/song lineation with XHTML-safe poem blocks when needed. Record stale sample strings with exact zero-hit scans rather than editing unrelated current text; rebuilt rounds with these repairs remain non-PASS and require a fresh seed.
• Round 070 index-continuation and poem-block addendum: If a random sample begins inside a volume index entry, recover the governing headword from the previous paragraph or previous unit and prefix the sampled unit with 词头（SOURCE HEADWORD）（续）：; do this for both file-start continuations and later paragraph continuations in the same file. Use exact sampled starts such as 因惧怕斯摩莱特, 酿酒厂——利润, 安排地产继承, 弗格森是, 士兵，像, 只排除一种祸患, 衣着朴素, and iii. 271；v. 150 as low-token scan triggers, then inspect only the nearby index flow before changing. In the same audit pass, treat one-line songs or verse quotations and raw bibliography rows as reader-interface defects: convert confirmed flattened poems to XHTML-safe poem blocks, change dangling 他说—— bridges to complete quote introductions, and normalize source-first rows such as BRITISH MUSEUM MSS. into target-language bibliography prose while preserving the source form in parentheses. Rebuild, scan final/translated/generated outputs for the rejected starts and old raw rows, and open a fresh seed because any repaired round is not a clean PASS.
• Round 071 bracket-note sampler, title-gloss, and raw-index addendum: If a random sample aggregates a quoted letter with bracket-numbered notes such as [80], first fix the reader-unit extractor so bracket-number note paragraphs are caption_note units instead of being swallowed by quote-balancing. Still repair confirmed surface defects in the same round: damaged mixed-language titles such as La Crédulité des incré不信者的轻信（dules）, setup paragraphs whose promised verse sits in another unit, flattened verse quotations, false-negative constructions such as 并不比任何别的人更不可接受, raw italic index headwords such as _lady-like_ woman, _law_, or _Sartum tectum_, and source titles split across line breaks. Scan exact sampled strings first, then related _[A-Za-z][^_]{2,}_ index residues and setup endings before deciding. Repair with a Chinese reader label first, source form in parentheses when helpful, XHTML-safe poem blocks for verse, and sampler smoke tests that prove the failed bracket note is no longer sampled as a paragraph. Rebuild and open a fresh seed because sampler or text repairs keep the round out of the clean PASS streak.
• preservative polysemy guard: In doctrinal or polemical contexts, preservative often means a protective formula, antidote, or preventive treatise, not a literal 防腐剂. Scan 防腐剂|preservative|antidote|remedy|防腐 and repair only confirmed doctrinal or polemical false friends; keep real body-preservation or embalming contexts as exceptions. 引文释义桥防护：源语引文可以保留简短原文片段，但目标语释义必须读成完成态注释。应在终稿、译稿、生成 XHTML 和抽样 manifest 中扫描 ：意思是：、 意思是：、说：意思是：、表达 意思是 等制作痕迹；确认命中后改成 其意为：、直接引语或正常书目句。这句话的意思是 这类普通表述，只有在整条读者单元已经流畅可懂时才作为合理例外保留。

When a concrete book discovers a new quality family, write the immediate evidence in that book's retrospective_lessons.md, qa/chapter_controls/, or reviews/random_spotcheck/round_XXX/fixes/fix_log.md, then backfill the reusable lesson here. 具体书发现新的质量问题族时，先把即时证据写入该书的 retrospective_lessons.md、qa/chapter_controls/ 或 reviews/random_spotcheck/round_XXX/fixes/fix_log.md，再把可复用经验回填到本 skill。

Backfill only the general lesson: how to find it, how to classify it, how to audit similar cases, how to fix it, and how to recheck it. 回填只写通用经验：如何发现、如何归纳、如何审计同类、如何修复、如何复查。

Round 390 hard-translation and reader-unit closure addendum

When samples repeatedly expose source-order calques, metaphor literalization, missing subjects, unit wording, or commentary inserted between narrative clauses, first build low-token candidate pools from the sampled strings, glossary, source/final pairs, and note markers. Confirm each hit against the source, then audit the whole book for the same family before editing. Keep source-backed dual note references as explicit exceptions; do not merge them merely because their markers are adjacent. For setup lists and quotations, preserve translation-coverage paragraph blocks while making the reader-facing unit self-contained; if a merge lowers structural coverage, restore paragraph boundaries and use a concise date or payload label rather than deleting content. Rebuild, rerun coverage/publication/reader/EPUB gates, and open a fresh seed; repaired rounds never count toward the clean PASS streak.

Domain Metaphor False Literalization / 领域隐喻误作字面动作

• Symptom: A source term from bodily process, craft, cookery, medicine, law, or another technical domain is used metaphorically in criticism or anecdote, but the target text turns it into a literal physical action. This creates a comic or misleading image not present in the source, such as treating dramatic concoction as digestion rather than initial composition. 症状：源语把身体过程、工艺、烹调、医学、法律等领域词借入批评或轶事语境，译文却把它落成字面动作，造成原文没有的滑稽或误导图像。例如戏剧评论里的 concoction 应指最初构思，不应变成消化过程。
• Find by: Search exact source triggers and repaired target candidates together. Useful patterns include concoction|first concoction|irons in the fire|Quare adhaesit pavimento|pavimento plus target candidates such as 第一道消化|调制|最初构思|铁在火里|人行道. Inspect the sentence domain before deciding whether the image is literal, idiomatic, legal, or a source-quote interface. 发现方式：源语触发词和译文候选并查。可用模式包括 concoction|first concoction|irons in the fire|Quare adhaesit pavimento|pavimento，并联查 第一道消化|调制|最初构思|铁在火里|人行道 等译文候选。判定前先看句子领域，区分字面图像、习语、法律笑话和源语引文接口。
• Evaluative-verb extension: Historical notes often use swallowed, embraced, or espoused for credulous acceptance of a claim. Scan target residues such as 贪婪吞下|吞下.*说法|吞下.*指控|急切吞下; if the object is a thesis, calumny, tradition, or opinion, translate the evaluative relation as 急切采信, 轻信接受, 照单全收, or 偏执地采纳. Record literal eating or birds/beasts devouring bodies as exceptions. 评价动词扩展：历史注释常用 swallowed、embraced、espoused 表示轻信或采信某种说法。可扫目标语残留 贪婪吞下|吞下.*说法|吞下.*指控|急切吞下；若宾语是论点、诽谤、传统或意见，应译成“急切采信”“轻信接受”“照单全收”“偏执地采纳”等评价关系。真正吃食、鸟兽吞食尸体等字面场景应记录为例外。
• Sacred or classical image extension: Terms like fiery chariot belong to biblical or classical imagery, not modern transport. Scan source triggers such as fiery chariot with target candidates 火车|火焰车|火焰战车; in modern Chinese, prefer 火焰战车 or another context-bound image, and record literal modern vehicles as defects unless the source scene is actually modern. 圣经或古典意象扩展：fiery chariot 这类词属于圣经或古典意象，不是现代交通工具。并查源文触发词 fiery chariot 与译文候选 火车|火焰车|火焰战车；现代中文通常应译作 火焰战车 或其他贴合语境的意象，除非源文场景本身是现代交通，否则现代车辆字面义应视为缺陷。
• Fix by: Translate the function in the active domain first. For dramatic judgement, use 最初构思 or another composition term; for a joke that intentionally literalizes an idiom, keep the literal image only if the Chinese sentence also makes the joke intelligible. For short Latin legal phrases, keep the Latin when it carries the joke or evidence, but add an immediate Chinese meaning. 修复方式：先按当前语域翻译功能。戏剧判断用 最初构思 等写作/构作术语；若原文故意把习语字面化，则仅在中文也能读出笑点时保留字面图像。短拉丁法律语若承担笑点或证据，可保留拉丁原文，但附近必须有中文含义接口。
• Recheck: A round that fixes this family cannot PASS. Rerun a fresh whole-unit review, then rescan both the rejected literal wording and the accepted domain term in translated drafts, final Markdown, generated XHTML when available, glossary rows, and QA notes. 复查：凡修复此问题族的轮次不能 PASS。必须重新做一轮整章复查，并在译稿、终稿、已生成 XHTML（如已有）、术语表和 QA 记录中同时扫描被拒译法与已采纳的领域术语。