saberongo/public-domain-books-translation/quality-framework

Simplified Chinese Target Quality / 简体中文目标语言质量

Use this skill before translating full chapters. Do not begin batch translation until a sample passes the quality gate.

Required Workflow

1. Read references/research_digest.md.
2. Read references/quality_standard.md.
3. Read references/workflow.md.
4. Create metadata/book_specific_translation_research.md with templates/book_specific_research_prompt.md.
5. Create metadata/style_profile.md with templates/style_profile_prompt.md.
6. If the user provides lawful private benchmark excerpts, evaluate them with templates/private_benchmark_compare_prompt.md.
7. Run pretranslation trials with templates/pretranslation_trial_prompt.md.
8. Run the sample test described in tests/benchmark_protocol.md and templates/sample_translation_test_prompt.md.
9. Only after pretranslation and sample tests pass, translate chapters with templates/chapter_translation_prompt.md.
10. Revise failed chapters with templates/revision_prompt.md.
11. Use skills/expert-translation-quality/SKILL.md during chapter translation, chapter controls, review, reader-feedback fixes, and final artifact review when expert-level prose, context-dependent word choice, translation-stage polysemy handling, or downstream polysemy back-checking matters.
12. Audit image-bearing word choices with templates/image_word_audit_prompt.md.
13. Gate every chapter with templates/chapter_quality_gate_prompt.md.
14. Score every final chapter with templates/evaluation_rubric.md.
15. When a recurring translation-quality defect family is found, use skills/translation-quality-defect-families/SKILL.md, close the family in the current book, and merge reusable lessons into that skill.

Hard Rules

• Never optimize only for literal correspondence.
• Never keep English syntax when natural Chinese would say it differently.
• Never preserve a metaphor word-for-word when it becomes obscure or ugly in Chinese; preserve the effect instead.
• Never use copyrighted reference translations as source material unless the user provides lawful excerpts for private evaluation. Do not store long copyrighted excerpts in the project.
• If a sentence sounds like a translation, rewrite it unless the stiffness is intentional in the original.
• During the translation call itself, use a slim prompt and output only the translation. Do not ask the model to write QA, release notes, lint records, or workflow explanations in the same call.
• Natural Chinese is the first hard constraint in translation. Faithful but breathless long sentences, English clause skeletons, or compliance-looking prose that does not read like Chinese fail the gate.
• Polysemous or context-dependent source wording must be actively handled during translation when local context can resolve it, then revisited after downstream context is available when needed. A smooth but wrong word sense is a blocking mistranslation.
• If a passage is historically sensitive, keep the historical fact and tone, but add translator notes instead of silently modernizing.
• Do not translate a chapter without a book-level style profile and a passing sample-test report.
• In body text, prefer accurate image-bearing words over flat explanatory words when the source is symbolic, scenic, physical, or memorable.
• Do not start batch chapter translation without a passing qa/pretranslation/pretranslation_report.md.
• If a review finds a recurring quality pattern, do not only fix the sampled sentence. Classify the defect family, audit similar cases with low-token methods first, and update skills/translation-quality-defect-families/SKILL.md with reusable lessons.
• 发现反复出现的质量模式时，不得只修抽中的句子。必须归纳问题族，先用低 token 方法审计同类，并把可复用经验回填到 skills/translation-quality-defect-families/SKILL.md。
• 专家级译文、多义词翻译阶段主动判义与后文回看使用 skills/expert-translation-quality/SKILL.md 执行；具体复发坏例再沉淀到 skills/translation-quality-defect-families/SKILL.md。

Quality Gate

A chapter can become eligible for deterministic materialization into chapters/final/ only if the canonical generation satisfies all of the following. Reviewers never move or directly edit projection files:

• No major meaning errors.
• No missing paragraph.
• No unresolved proper noun or technical term inconsistency.
• Readability score is at least 4/5.
• Literary/narrative effect score is at least 4/5.
• Chinese-only readability score is at least 4/5 before source calibration.
• qa/chapter_controls/ records expert_translation_skill_used: true, expert_level_review_status: "PASS", polysemy_translation_stage_review: "PASS", polysemy_context_review: "PASS", and polysemy_unresolved_count: 0.
• A 20-sentence read-aloud sample has at most one clearly awkward sentence, and no key sentence is breathless.
• The reviewer cannot point to a sentence that feels mechanically translated.
• A gate report exists under qa/gates/ and says PASS.
• qa/samples/sample_test_report.md exists and says PASS.
• qa/pretranslation/pretranslation_report.md exists and says PASS.
• An imagery audit exists under qa/imagery/ for chapters with symbolic, scenic, physical, or memorable language.
• Any discovered translation-quality defect family has book-local evidence, confirmed fixes or exceptions, and a reusable merged lesson in skills/translation-quality-defect-families/SKILL.md.

Expert Translation Quality / 专家级译文质量

Translation, chapter controls, independent review, random spot-checks, final artifact review, private-use artifact review, and reader-feedback fixes must use skills/expert-translation-quality/SKILL.md when expert-level prose, context-dependent word choice, translation-stage polysemy handling, or downstream polysemy back-checking matters. Recurring concrete defects still belong in skills/translation-quality-defect-families/SKILL.md.

翻译、章节 control、独立审校、随机抽检、最终产物审阅、private-use 产物审阅和读者反馈修复中，只要涉及专家级译文、上下文依赖选义、翻译阶段多义词处理或多义词回看，必须使用 skills/expert-translation-quality/SKILL.md。具体复发坏例仍沉淀到 skills/translation-quality-defect-families/SKILL.md。