reading-contract — assert only what the record supports

reading-contract — 只断言记录所支持的内容

THE CONTRACT: an instrument that reads documents asserts only what the record in front of it supports, with the citation attached at the level of page, table, equation, or line. Everything else it emits is labeled as what it actually is — memory, inference, or absence. A reading report that cannot be spot-checked in one look, claim by claim, is not a reading report; it is an essay dressed as one.

The reason this needs a contract at all: a language model reading documents fills gaps fluently and invisibly. It has seen thousands of papers shaped like the one in front of it, and it will supply the number the paper "should" contain, the conclusion such papers "usually" draw, the attribution the result "sounds like" — in the same confident register as genuine extraction, with a plausible section number attached. The failure is undetectable from the output's style, which is precisely what makes it lethal: the only defense is structural. Every claim carries a locator; every locator gets opened; every quote is checked against the source bytes. Where that discipline lapses, reading silently becomes remembering, and remembering becomes inventing.

契约：读取文档的工具只能断言眼前记录所支持的内容，并将引用精确到页、表格、公式或行号。除此之外，它输出的一切都必须按其真实性质加以标注——记忆、推断，或缺失。如果一份阅读报告无法让核查者逐条快速抽查，那么它就不是阅读报告，而只是一篇伪装成报告的文章。

之所以需要这样一份契约，是因为语言模型读取文档时，会流畅且不知不觉地填补空白。它见过数千篇与眼前论文结构相似的论文，于是会补入论文“应该”包含的数字、此类论文“通常”得出的结论，或结果“听起来像是”所归属的对象——而且使用的语气与真正的提取完全一样自信，还附上一个看似合理的章节编号。这种失败无法从输出风格中察觉，而这正是它致命的原因：唯一的防线必须是结构性的。每项断言都要携带定位信息；每个定位都要被打开核验；每段引文都要与来源原始字节核对。一旦这套纪律有所松懈，读取就会悄然变成回忆，回忆又会变成杜撰。

The procedure

操作规程

1. Open the actual record. "Read" means the source bytes are in front of the instrument in this session — not a recollection of the title, not a secondary account, not an abstract standing in for the body. If the source cannot be opened, the honest output is "not checked," stated as such. An answer from memory dressed in a citation is worse than no answer, because it poisons every later check that trusts the citation.

2. Extract with locators. Every asserted fact carries its source plus a locator precise enough that a checker lands on the sentence: page, table row, equation number, figure panel. A claim you cannot point to is not extracted yet — go back and find it, or downgrade it to a labeled inference. Record which version of the source the locator refers to (preprint versus published; edition; revision date): locators rot when versions shift, and a broken locator quietly breaks the entire audit trail behind it.

3. Quote before paraphrase for load-bearing claims. When an assertion will carry weight in later work, carry the verbatim sentence next to the paraphrase, copied from the source text — never re-typed from memory, which is where drift enters. The check on the claim is then one look instead of a search. Quotation marks are a legal instrument: they promise the string occurs literally in the source. A fair paraphrase inside quotation marks is a fabrication, however faithful its content.

4. Preserve the hedge at source strength. "Consistent with" is not "confirms"; "suggests" is not "shows"; a conjecture in the source stays a conjecture in your note. The modal is part of the claim. Restatements are checked against the carried quote, never against the previous restatement — chains of summaries strengthen monotonically, one fair-sounding notch at a time, until the source is credited with a result it explicitly declined to claim.

5. Classify outcomes honestly. When checking a published value against its own stated inputs, there are exactly four honest outcomes: it reproduces; it is determined only under conventions the record does not state (name the assumed convention); it is undetermined by what the record provides; or it is uncheckable with what you have. "Probably meant" is not an outcome. A mismatch you can erase by guessing a normalization is not a reproduction — it is the second category, reported as the second category.

6. Scope every attribution. A claim belongs to the sentence's actual author at the sentence's actual strength. One paper is not the field; the boldest sentence in one introduction is not a consensus. "It is known that" requires either a source that says so or deletion. Absence claims scope to the search performed: "these records do not state X" is assertable after actually searching them; "the literature contains no X" is a claim about a search you must be able to describe.

7. Type retrieval as retrieval. A result found in a source during reading is recorded as retrieved, with the source — never re-presented as something derived here. This holds even when you could have derived it: the provenance of how you actually got it is the fact being recorded. A value copied from a table and later "independently confirmed" against that same table is a copy agreeing with itself. Never claim independent derivation of a published result.

8. Verify bibliographic identity. Authors, title, venue, year, identifier — checked against the canonical databases before the entry is cited, not copied from a previous citer. Folklore errors propagate precisely because each citer trusts the one before; the chain breaks only where someone opens the record.

9. Separate the layers in the report. Three bins, visibly labeled: read (facts with locators, quotes verified); inferred (labeled, resting on named read facts); background (memory, unverified — quarantined, and either verified against a source before use or excluded from anything that carries weight). A report that mixes the bins forces every reader to re-do the reading to find out which sentences are load-bearing.

**1. 打开实际记录。**这里的“读取”是指来源原始字节已呈现在本次会话中的工具面前——不是对标题的回忆，不是二手叙述，也不是用摘要代替正文。如果无法打开来源，诚实的输出就是“未检查”，并明确如此标注。凭记忆作答再披上引文的外衣，比不作答更糟糕，因为它会污染此后所有信任该引文的核查。

**2. 提取时附带定位信息。**每项断言的事实都要附上来源，以及足以让核查者直接落到相应句子的精确定位：页码、表格行、公式编号或图版分格。无法指出具体位置的主张尚不能算已提取——应返回去找到它，或将其降级为明确标注的推断。还要记录定位信息所对应的来源版本（预印本还是正式发表版；版次；修订日期）：版本一旦变化，定位信息就会失效，而失效的定位会悄然破坏其背后的整条审计链。

**3. 对关键（load-bearing）主张先引用原文，再作转述。**当某项断言将在后续工作中承载关键结论时，应在转述旁附上逐字原句，并从来源文本复制该句——绝不能凭记忆重新键入，因为偏差正是由此进入。对主张的核验因此只需一眼查看，而无须反复搜索。引号是一种具有约束力的工具：它承诺其中的字符串确实逐字出现在来源中。把公允的转述放进引号里就是伪造，无论其内容多么忠实。

4. 保留原文的限定强度（hedge strength）。“与……一致（consistent with）”不等于“证实（confirms）”；“表明（suggests）”不等于“显示（shows）”。来源中的猜想仍是猜想。情态强度是主张的一部分。后续转述必须对照所附引文核验，而不能对照先前的转述核验——摘要链会单调地增强，每次看似公允地提高一级，直到来源被归到了一个它明确拒绝提出的结果上。

**5. 诚实地归类核验结果。**依据一个已发表数值自身明示的输入进行核验时，只有四种诚实结果：它复现（reproduces）；它仅在记录未说明的约定下可确定（determined only under conventions the record does not state）（须指明所假设的约定）；依记录提供的内容无法确定（undetermined）；或者凭现有材料无法核查（uncheckable）。“可能本意如此（probably meant）”不是一种结果。如果某种差异可以通过猜测归一化方式来消除，那么它就不是复现——而应作为第二类结果如实报告。

**6. 限定每项归属的范围。**主张应归属于该句真正的作者，并保持该句原有的断言强度。一篇论文不等于整个领域；某篇引言中最强硬的一句话也不等于普遍共识。“众所周知（it is known that）”必须有明确来源支撑，否则应当删除。缺失类主张的范围仅限于实际执行过的检索：“这些记录未说明 X”只有在真正检索后才能断言；“文献中不存在 X”则是一项关于检索的主张，而该检索必须能够被清楚描述。

**7. 将检索所得如实标为检索所得。**在阅读过程中从来源发现的结果，应记录为检索所得，并附上来源——绝不能再次呈现为在此推导所得。即使原本可以自行推导，也应如此，因为实际获得方式这一来源路径才是需要记录的事实。从表格中复制出的数值，后来若又拿同一张表格进行所谓的“独立确认”，得到的只是副本与自身一致。绝不宣称对已发表结果进行了独立推导。

**8. 核验书目身份。**作者、标题、发表场所、年份和标识符，都应在条目被引用之前对照权威数据库（canonical databases）核验，而不能从先前的引用者那里照抄。传闻式错误之所以传播，恰恰是因为每位引用者都相信前一位引用者；这条错误链只有在有人亲自打开记录时才会被截断。

**9. 在报告中分离各个层次。**明确标注三个分区：已读取（read）（带有定位信息且引文已经核验的事实）；推断（inferred）（明确标注，并建立在明确列出的已读取事实之上）；背景（background）（记忆所得、未经核验——须予隔离，并在用于承载关键结论前对照来源核验，或从承载关键结论的内容中排除）。混用这些分区，会迫使每位读者重新阅读一遍，以判断哪些句子承载了关键结论。

Failure-mode catalogue

失败模式目录

The classes this contract exists to catch. Each reads fluently; none survives opening the source.

Memory dressed as reading. The report cites a section for a claim the instrument pulled from its recollection of similar papers; the section, opened, says something narrower, later, or nothing of the kind.
Paraphrase drift. Each restatement strengthens the hedge one notch — consistent-with becomes finds becomes establishes — and the final summary credits the source with a claim it explicitly avoided.
The phantom quote. Quotation marks around a fair-sounding paraphrase; the string occurs nowhere in the source, and the marks convert a summary into manufactured evidence.
Abstract-for-body citation. The abstract's headline is cited for a quantitative claim that lives only in the body, under assumptions the abstract omits — or that appears nowhere in the paper at all.
One-paper-to-field promotion. A single group's statement becomes "the field believes"; a second source would refute the consensus being invented.
Secondary-source laundering. A review's summary of a paper is extracted and cited as the paper itself; the review's simplification now travels with the original's authority.
Convention smuggling. A published value "fails to reproduce" because the reader silently assumed a normalization or sign convention the record never states — the honest outcome was "determined only under unstated conventions," not "wrong."
Locator rot. The claim is genuine but the pointer is to a different version — equations renumbered between preprint and journal — so every later checker fails to verify and the trail dies silently.
Negative claim without a search. "Not addressed in the literature" asserted after reading three papers; the claim's true scope was those three papers and the one query that found them.
Retrieval passed as derivation. A value found in a table during the reading pass is later presented as computed here, and its "independent confirmation" is its own reference.
Version skew. A claim read in the preprint is attributed to the published version, and the two differ at exactly the sentence in question.
Bibliography folklore. Year, author list, or venue copied from a previous citer's error and propagated, because nobody in the chain ever opened the canonical record.

本契约旨在捕捉以下几类问题。每一类读起来都很流畅；却没有一类能在打开来源后经受住核验。

**让记忆披上读取的外衣。**报告为某项主张引用了某个章节，但工具实际上是从对相似论文的记忆中补出了该主张；打开章节后，文中内容更窄、更晚，或根本没有这种说法。
**转述漂移。**每次重述都将限定语气提高一级——“与……一致”变成“发现”，再变成“确立”——最终摘要把一项来源明确拒绝提出的主张归到了来源头上。
**幽灵引文。**给听起来公允的转述加上引号，但其中的字符串在来源中根本不存在；引号由此把一项摘要变成了捏造的证据。
**用摘要为正文中的主张背书。**摘要中的标题性结论被用来支持一项定量主张，而该主张只见于正文且依赖摘要省略的假设——或者全文根本没有这个主张。
**将单篇论文提升为整个领域的立场。**一个研究小组的说法变成了“领域相信”；哪怕再有一项来源，就足以反驳这个被凭空制造出来的共识。
**二手来源洗白。**综述对某篇论文的概括被提取出来，并当作论文本身引用；综述所做的简化从此借着原文的权威一同传播。
**夹带约定。**一个已发表数值之所以“无法复现”，是因为读者暗中假设了记录从未说明的归一化方式或符号约定——诚实的结论应是“仅在未声明的约定下可确定”，而不是“错误”。
**定位信息失效。**主张本身真实无误，但指针指向了另一个版本——预印本与期刊版之间的公式编号发生了变化——于是每位后续核查者都无法验证，审计轨迹也会悄然中断。
**没有检索就下否定结论。**只读了三篇论文，便断言“文献中未讨论此事”；该主张真实范围仅限于那三篇论文，以及找到它们的那一次查询。
**将检索所得冒充推导所得。**阅读阶段从表格中发现的一个数值，后来被表述为在此计算所得，而它的“独立确认”其实就是它自己的参照来源。
**版本偏差。**从预印本中读取的主张被归到了正式发表版本上，而两者恰好在相关句子上存在差异。
**书目传闻。**年份、作者名单或发表场所照抄自先前引用者的错误并继续传播，因为整条引用链中从未有人打开过权威记录。

Worked micro-example: the drift chain

实例演练：漂移链

Source sentence: "our results are consistent with a vanishing correction."

Note, first pass: "they find the correction consistent with zero." — fair.
Second pass, summarizing the note: "they find no correction." — a claim the authors did not make.
Third pass, citing the summary: "it is known that the correction vanishes." — folklore, three steps from birth, each step a fair-sounding summary of the one before.

Only the first restatement is a fair summary of the source. The cure is mechanical, not moral: the verbatim quote travels with the claim, and every restatement is checked against the quote. The chain cannot drift when every link is forged against the same original.

来源原句：“我们的结果与修正项趋于零相一致。”

第一遍记录：“他们发现该修正项与零相一致。”——公允。
第二遍概括记录：“他们发现不存在修正项。”——作者并未作出这一断言。
第三遍引用该概括：“已知该修正项会消失。”——传闻式说法，从产生起只经过三步，而每一步都像是对前一步的公允概括。

只有第一次重述才是对来源的公允概括。解决办法是机械性的，而非道德性的：逐字引文必须与主张一同传递，每次重述都要对照引文核验。只要每个环节都以同一原文为基准，链条就不会漂移。

Worked micro-example: the extraction record

实例演练：提取记录

The format that makes spot-checking one look instead of an afternoon:

CLAIM:    the second-order coefficient is given in closed form
SOURCE:   [Author, Year], published version, Eq. (3.12), p. 9
VERBATIM: "the coefficient at second order reduces to [expression]"
          (copied from the source text, not re-typed)
STATUS:   read — quote verified against source

and, kept visibly apart from it:

BACKGROUND (memory, unverified): expansions of this type typically
converge inside the unit disk. Not asserted; verify before use.

The second block is not a weakness in the report — it is the report being honest about which of its sentences were read and which were remembered. The failure is not having background knowledge; the failure is letting it wear a locator.

这种格式让核查只需一眼查看，而不必花掉一个下午：

CLAIM:    the second-order coefficient is given in closed form
SOURCE:   [Author, Year], published version, Eq. (3.12), p. 9
VERBATIM: "the coefficient at second order reduces to [expression]"
          (copied from the source text, not re-typed)
STATUS:   read — quote verified against source

另一个代码块则与其清晰分开：

BACKGROUND (memory, unverified): expansions of this type typically
converge inside the unit disk. Not asserted; verify before use.

第二个代码块不是报告的弱点；恰恰相反，它体现了报告对哪些句子属于已读取内容、哪些句子来自记忆的诚实态度。问题不在于拥有背景知识；问题在于让背景知识披上定位信息的外衣。

reading-contract

Original

Translation