Calibrated translation prompt v4 contract
经校准的翻译提示词 v4 约定
Automated counterpart generation needs a stable prompt that reproduces the register and corrections established by human-reviewed translations. Injecting a general-purpose instruction document changes that calibrated model input whenever human or agent guidance changes, while an unframed response cannot carry a draft, its self-review, and the corrected document separately. Plain XML-like section tags also collide wit
English
Problem
Automated counterpart generation needs a stable prompt that reproduces the register and corrections established by human-reviewed translations. Injecting a general-purpose instruction document changes that calibrated model input whenever human or agent guidance changes, while an unframed response cannot carry a draft, its self-review, and the corrected document separately. Plain XML-like section tags also collide with valid Markdown that documents those same tags.
Decision
The committed translation prompt is the calibrated pipeline asset. Its renderer injects only the source language, target language, and current terminology table, and rejects unknown, missing, or malformed placeholder syntax before assembling a request. The request assembler retains the source basename outside the model-visible prompt and places each reviewed whole-document pair into one bare-text user/assistant example turn before the real source document. The template may carry model-specific calibration rules, but those rules remain subordinate to the repository's pairing, terminology, structure, and emphasis contracts.
The v7 calibration retains that v4 protocol and makes the instruction priority explicit: source meaning and protected structure, then the terminology table, then whole-document gold-pair voice, then general guidance and embedded examples. It directs the model to draft as a native technical author and then compare clause by clause, preserving actors, conditions, negation, modality, lifecycle conditions, direction, result channels, ownership, and quantities. Style guidance cannot invent an actor or vary a terminology-table form, defined concept, or contract verb merely for variety. Unresolved terminology stays unchanged in the translation and is reported only as pending review.
The response has three ordered top-level sections: translation, review, and final. The response consumer derives the target basename from the retained source context, preserves optional leading YAML frontmatter, and mechanically inserts or corrects the language switcher after the first H1 in final. The parser requires each section exactly once, rejects content outside the envelope, and tolerates one outer xml Markdown fence because models sometimes echo the prompt's example fence.
Response framing
Section delimiter lines are reserved by the wire format. When a Markdown body line consists of a delimiter tag, possibly preceded by backslashes, the serializer and model add one leading backslash; the parser removes exactly one. This count-preserving escape round-trips both a literal delimiter and an already escaped delimiter without changing inline tag mentions.
The executable contract lives in the renderer, request assembler, parser, and response consumer. Unit tests cover both directions, request order, placeholder validation, target-path validation, strict section order and cardinality, fenced responses, inline tag mentions, delimiter lines inside Markdown bodies, and frontmatter-preserving new-pair switcher correction. A keyless subprocess snapshot pins the assembled prompt and five reviewed example turns together with a frontmatter-bearing recorded response consumed through the target-path correction.
Alternatives considered
Inject translation-rules.md into every request. That document governs humans and agents as well as the automated pipeline. Injecting it couples each editorial clarification to model behavior and displaces the manually calibrated prompt constraints; the pipeline instead injects the binding terminology table and verifies its own asset directly.
Use a strict CDATA XML document. CDATA provides general XML framing but adds a nested protocol, an additional ]]> escape, and XML-parser behavior that the three-section contract does not otherwise need. Reserving and escaping six delimiter lines keeps the calibrated response sections while preserving arbitrary Markdown.
Return only the final translation. A single body is simpler to parse but discards the explicit correction pass used to catch tone, structure, terminology, and punctuation defects before publication.
Replace the calibrated v4 asset wholesale with a later experimental prompt. Later experiments clarified useful general rules but did not pass strict whole-document evaluation as complete replacements. The production asset adopts only the improvements that preserve the established examples and executable protocol.
Require a strict draft-to-final change ledger. Requiring every final edit to appear in a free-text review ledger adds output burden without proving that the review catches local semantic drift. The review records actual corrections, while the final translation remains subject to deterministic structure checks and human review.
Consequences
Prompt wording is executable behavior and receives code review, a translation-prompt verifier, and a runnable request/response snapshot. Focused tests pin the embedded examples and selected v7 safeguards. The translation-prompt-v4 snapshot directory names the stable renderer/parser protocol series rather than the current calibration revision. The calibrated asset and the general translation rules can evolve for their different audiences, but review must reject contradictions with repository contracts. The line escape is visible only when source documentation contains a wrapper tag on its own line, and parser tests pin its lossless behavior.
中文
问题
自动生成对侧文件需要一份稳定的提示词，能够复现经人工评审的译文所确立的语体和修正方式。注入通用说明文档，会让这份经校准的模型输入随着面向人类或 agent（智能体）的指导发生变化，而未经封装的响应无法分别承载草稿、自检内容和修正后的文档。普通的类 XML 分段标签还会与用于说明这些标签的合法 Markdown 内容发生冲突。
决策
提交入库的翻译提示词是经过校准的流水线资源。其渲染器仅注入源语言、目标语言和当前术语表，并在组装请求前拒绝未知、缺失或语法格式错误的占位符。请求组装器在模型可见的提示词之外保留源文件基本名，并在真正的源文档之前，将每组经评审的整篇文档对编排为一个纯文本 user/assistant 示例轮次。模板可以包含针对特定模型的校准规则，但这些规则必须服从仓库的配对、术语、结构与强调格式约定。
v7 校准保留这套 v4 协议，并明确指令优先级：先保持源文含义与受保护结构，再遵循术语表，然后以整篇金标校准语体，最后应用一般指导与内嵌示例。模型先以母语技术作者的方式起草，再逐分句对照源文，保留执行主体、条件、否定、情态、生命周期条件、方向、结果通道、所有权和数量。文体指导不得虚构执行主体，也不得仅为丰富措辞而改换术语表词形、已定义概念或约定动词。无法裁定的术语在译文中保持不变，只在评审段报告为待评审项。
响应包含三个有序的顶层分段：translation、review 和 final。响应消费方根据保留的源文件上下文推导目标文件基本名，保留文件开头可选的 YAML frontmatter，并以机械方式在 final 中第一个 H1 之后插入或校正语言切换行。解析器要求每个分段恰好出现一次，拒绝封套之外的内容，并允许响应最外层有一层 xml Markdown 围栏，因为模型有时会照抄提示词中的示例围栏。
响应封装格式
分段定界行由协议格式（wire format）保留。当 Markdown 正文中的某一行仅包含定界标签（前面可以带反斜杠）时，序列化器和模型会在行首再添加一个反斜杠；解析器则只移除一个。这种保留计数的转义方式让字面量定界标签与已转义的定界标签都能无损往返，同时不会改动行内提及的标签。
可执行约定由渲染器、请求组装器、解析器和响应消费方实现。单元测试覆盖两个翻译方向、请求顺序、占位符校验、目标路径校验、严格的分段顺序与数量约束、带围栏的响应、行内提及标签、Markdown 正文中的定界行，以及保留 YAML frontmatter 的新配对语言切换行校正。一个无密钥子进程快照锁定组装后的提示词、五个经评审的示例轮次，以及带 YAML frontmatter 的录制响应经目标路径校正后的消费结果。
考虑过的替代方案
在每个请求中注入 translation-rules.md。 该文档既约束人类与 agent，也约束自动翻译流水线。注入它会让编辑规范的每次澄清都与模型行为耦合，并挤占经过人工校准的提示词约束；因此流水线仅注入具约束力的术语表，并直接校验自身资源。
使用严格的 CDATA XML 文档。 CDATA 提供通用的 XML 封装，但会引入一层嵌套协议、额外的 ]]> 转义规则，以及三段式约定原本不需要的 XML 解析器行为。预留并转义六种定界行，既能维持经校准的响应分段，也能保留任意 Markdown 内容不变。
只返回最终译文。 单一正文更易解析，却会丢弃显式修正步骤；这个步骤用于在发布前发现语气、结构、术语和标点缺陷。
用较新的实验提示词整体替换经校准的 v4 资源。 后续实验澄清了一些有价值的通用规则，但作为完整替代方案未通过严格的整篇文档评估。生产资源只吸收能够保留既有示例与可执行协议的改进。
要求严格的草稿到定稿变更账本。 要求定稿中的每项修改都出现在自由文本评审账本中，会增加输出负担，却不能证明评审能够发现局部语义漂移。评审段记录实际修正，定稿仍需接受确定性结构检查与人工评审。
影响
提示词措辞属于可执行行为，因此需要经过代码评审、翻译提示词校验器校验及可运行的请求/响应快照验证。聚焦测试固定内嵌示例与选定的 v7 保护规则。translation-prompt-v4 快照目录命名的是稳定的渲染器／解析器协议系列，而不是当前校准修订号。经校准的资源与通用翻译规则可以针对各自的受众分别演进，但评审必须拒绝任何与仓库约定冲突的改动。只有当源文档中的封装标签独占一行时，行转义才会显现；解析器测试锁定这一无损行为。