验证部 · 深圳 / Verification Division, Shenzhen

我们发表全部内容。任何人都可以核查我们的工作。

We publish everything. Anyone may check our work.

本部门从事分类系统的独立复现与验证。我们发表代码、数据、参数、随机种子，以及全部未能复现的结果。未能复现的结果与其余部分同等重要，通常更重要。

This division performs independent replication and verification of classification systems. We publish our code, our data, our parameters, our random seeds, and everything we failed to reproduce. What failed to reproduce matters as much as the rest, and usually more.

复现不是指控。核查是一种尊重。没有人会去核查一件自己认为毫无价值的工作——不予理会即可，那只需要一个下午。

Replication is not an accusation. Checking is a form of respect. You do not check work you consider worthless; you dismiss it, which costs an afternoon.

研究Studies

本部门的唯一一条规则The division's only rule

凡发表，必附他人核查所需的全部材料。若我们不能附上，我们便不发表。

Whatever we publish, we publish with everything another person needs in order to check it. If we cannot supply that, we do not publish.

本站为静态页面。不使用访问统计，不设置 Cookie，不记录访问者。最后更新：2031 年。

This site is static. No analytics, no cookies, no visitor logging. Last updated 2031.

出版物 / Publications

论文Papers

三项研究。全部为预印本。全部附代码、数据与校验和。全部可自由再分发，唯一条件是署名。

Three studies. All preprints. All released with code, data, and checksums. All freely redistributable on one condition, which is attribution.

年份Year	题目Title	页pp.	评审状态Peer review
2026	某消费级排气语言学分类引擎的复现Reproduction of a Consumer Flatulinguistic Classification Engine	18	当日退回，未经评审returned same day, unrefereed
2030	某机构十年记录中的行政残差On an Administrative Residual in a Ten-Year Institutional Record	41	当日退回，未经评审returned same day, unrefereed
2031	对某消费级读取设备中一项未署名衍生关系的审计On an Unattributed Derivation in a Consumer Reading Device	23	未投稿not submitted

关于第三项为何未投稿：该领域仅有一种期刊。前两次投稿均于当日退回，未经评审，退稿单一字不差。我们认为第三次不会产生新的信息。

On why the third was not submitted: the field has one journal. Our first two submissions were returned the same day, unrefereed, on identical slips. We judged that a third would produce no new information.

被退回稿件的处置The disposition of a returned manuscript

两次退回均以同一份退稿单作出，一字不差，相隔五年。我们将其中一份装框，挂在验证部负责人办公室的墙上。理由见收讫。

Both returns were made on the same return slip, verbatim, five years apart. We have framed one of them. It hangs on the wall of the head of the division. The reason is given under Receipts.

预印本 · 2026-05-14 / Preprint

某消费级排气语言学分类引擎的复现Reproduction of a Consumer Flatulinguistic Classification Engine

HYAI-VD-2026-01 · 18 pp. · 用时九天assembled in nine days

林蔚然，陈砚，周汝成，以及验证部全体。通讯地址见 联系。

Lin Weiran, Chen Yan, Zhou Rucheng, and the Verification Division. Correspondence: see Contact.

摘要Abstract

我们复现了 The Fart Whisperer 应用之解释引擎的输出。系统由一个小型音频分类器与一张查找表构成，在本部门自行录制并标注的四千份样本上训练而成，历时九天。在该应用公开的分布上，一致率为百分之九十四点一。全部代码、训练数据、标注规程与随机种子随文发布。本文的目的不是评价该引擎，而是确定其输出是否可由外部材料重新产生。可以。

We reproduce the output of the interpretation engine of the application The Fart Whisperer. The system consists of a small audio classifier and a lookup table, trained in nine days on four thousand samples recorded and labeled by this division. Agreement with the published distribution is 94.1 per cent. All code, training data, labeling protocol, and random seeds are released with this paper. The purpose of this paper is not to evaluate the engine but to establish whether its output can be regenerated from outside materials. It can.

1 · 引言Introduction

该应用向用户提供一份“读数”：一段音频输入，经其解释引擎处理后，返回一段文字判读、一个可信度数值，以及一份“传承报告”——将该输入指派到八个祖源地区之一。该应用称其方法为专有，并未公布模型、特征、参数或参考语料。其公开材料中确有一项定量陈述：各地区的输出分布。

The application supplies its users with a "reading": an audio input is processed by its interpretation engine, which returns a written interpretation, a confidence value, and a Heritage Report assigning the input to one of eight ancestral regions. The application states that its method is proprietary and publishes no model, features, parameters, or reference corpus. Its public materials do contain one quantitative statement: the distribution of outputs across regions.

一项公开的分布是一项可检验的主张。本文即检验它。我们不询问该引擎内部如何工作；我们询问一个外部构建的系统，在不接触其任何内部材料的情况下，能否产生同样的输出。这是复现，而复现是科学唯一真正拥有的问题。

A published distribution is a testable claim. This paper tests it. We do not ask how the engine works internally; we ask whether a system built from outside, with no access to any of its internal materials, produces the same output. That is replication, and replication is the only question science actually owns.

我们在此明确一件事，以免其后被误解：本文不含任何指控。我们没有接触该公司的代码、模型、语料或雇员。我们所使用的每一份材料，均由本部门自行采集，并随本文一并公开。

We state one thing plainly here, to forestall misunderstanding later: this paper contains no accusation. We have had no access to the company's code, model, corpus, or employees. Every material used here was collected by this division and is published with this paper.

2 · 材料Materials

2.1 样本Samples

四千份音频样本，由成年志愿者在知情同意下提供，录制于 2026 年 3 月至 5 月。招募通过本部门公告与两所高校的校内公告进行，不设报酬。同意可随时撤回，撤回无需说明理由，亦不影响此前提供的其他样本。撤回记录见语料索引与附录 C。

Four thousand audio samples, supplied by adult volunteers under informed consent, recorded between March and May 2026. Recruitment was by division notice and by campus notice at two universities. No payment was offered. Consent is withdrawable at any time, without reason given, and without effect on other samples previously supplied. Withdrawals are recorded in the corpus index and in appendix C.

2.2 设备Equipment

录音设备为消费级手机与一支 20 元的驻极体话筒，取二者中较差的一路。这是刻意的：若复现只在实验室级设备上成立，则它并未复现一款在手机上运行的产品。

Recording used consumer telephones and one 20-yuan electret microphone, and we kept whichever channel was worse. This is deliberate: a replication that holds only on laboratory equipment has not replicated a product that runs on a telephone.

2.3 标注Labeling

每份样本由本部门两名成员独立标注，不一致处由第三人裁定，裁定记录保留。标注规程见附录 A，并作为独立文件发布。标注者在标注时不知晓该应用对同一样本的输出。

Each sample was labeled independently by two members of the division, with disagreements resolved by a third, and the resolutions retained. The labeling protocol is appendix A and is also released as a standalone document. Labelers did not know the application's output for a sample at the time of labeling it.

3 · 方法Method

系统有两个部件，合计四十一千字节。

The system has two parts and comes to forty-one kilobytes in total.

分类器。对每份录音，先以双阈值加迟滞确定端点，丢弃事件前后的静默，仅就发声段计算特征：时长、均方根电平、峰值频率、谱质心、谱通量，以及低频段与高频段能量比。特征共十一维。分类器为梯度提升树集成，单棵树深度不超过六。没有神经网络。没有需要训练的表示。

The classifier. For each recording, onset and offset are determined by a double threshold with hysteresis; the silence around the event is discarded, and features are computed on the sounded segment only: duration, RMS level, peak frequency, spectral centroid, spectral flux, and the ratio of low-band to high-band energy. Eleven feature dimensions in total. The classifier is a gradient-boosted tree ensemble of depth not exceeding six. There is no neural network and no learned representation.

查找表。分类器的输出并非文字。文字来自一张查找表：每一个特征桶对应一条写好的判读。表由本部门撰写，共九百四十条，随文发布。表中每一条均以冒号结尾，因为被复现的对象每一条都以冒号结尾。

The lookup table. The classifier's output is not text. The text comes from a lookup table: one written interpretation per feature bucket. The table was written by this division, contains nine hundred and forty entries, and is released with this paper. Every entry ends with a colon, because every entry in the object being replicated ends with a colon.

确定性。同一份录音必定产生同一条输出。随机性只进入训练，不进入推断，训练所用种子随文发布。这不是我们的设计选择，而是被复现对象的一项可观察性质：同一份录音在该应用上重复提交，返回相同结果。

Determinism. The same recording always produces the same output. Randomness enters training only, never inference, and the training seeds are published. This was not our design choice but an observable property of the object being replicated: the same recording submitted repeatedly to the application returns the same result.

4 · 评估规程Evaluation protocol

“一致率”在本文中的定义为：在同一批输入上，我们的系统与该应用返回同一地区判定的比例。判读文字不参与比对——两段不同的文字可以表达同一判定，而我们不愿以文本相似度冒充一致。可信度数值亦不参与比对，理由见 6.3。

"Agreement" in this paper means the proportion of inputs on which our system and the application return the same regional determination. The interpretation text is not compared — two different texts may express the same determination, and we decline to pass text similarity off as agreement. The confidence value is not compared either; see 6.3.

比对集为四千份样本中随机抽取的一千份，抽样种子公开。其余三千份仅用于训练，不参与评估，两者不重叠。

The comparison set is one thousand samples drawn at random from the four thousand, with the sampling seed published. The remaining three thousand were used for training only, take no part in evaluation, and do not overlap.

5 · 结果Results

一致率：百分之九十四点一。多数类基线：百分之七十一点四。

Agreement: 94.1 per cent. Majority-class baseline: 71.4 per cent.

5.1 按地区By region

地区Region	该应用公开占比Published share	本系统占比Ours
莱茵兰Rhineland	71.4%	71.4%
低地Low Countries	9.2%	9.3%
阿尔卑斯过渡带Alpine Transitional	6.1%	6.0%
伊比利亚沿海Iberian Coastal	4.8%	4.8%
潘诺尼亚盆地Pannonian Basin	3.3%	3.4%
波罗的海陆架Baltic Shelf	2.6%	2.5%
喀尔巴阡内陆Carpathian Interior	1.9%	1.9%
安纳托利亚边缘Anatolian Margin	0.6%	0.6%
合计Total	99.9%	99.9%

该应用公开的八项占比合计为百分之九十九点九。本部门就此不作解释，亦不认为需要解释；我们照录该数字；本系统亦得到同一合计，因为它正是为复现这一数字而构建的。

The eight published shares sum to 99.9 per cent. This division offers no explanation and does not consider one to be required. We record the figure, and our own system reproduces the same total, because it was built to reproduce this one.

5.2 残差的构成Composition of the residual

五点九个百分点的不一致中，五点八个百分点为常规误差：边缘样本、低能量样本、两个相邻地区之间的摇摆。这是任何一次诚实复现都会有的残差，我们对此不持异议。

Of the 5.9 percentage points of disagreement, 5.8 points are ordinary error: edge cases, low-energy samples, and wavering between two adjacent regions. This is the residual any honest replication carries, and we do not argue with it.

余下的部分不是误差。

The remainder is not error.

5.3 一项无法复现的参考读数One unreproducible reference reading

参考集中有一个读数，本系统的任何配置都无法产生——包括为产生它而专门构建的配置。我们将其记入脚注，置于第五节末尾，原文如下。

One reading in the reference set could not be produced by any configuration of our system, including configurations constructed for the purpose of producing it. We recorded it in a footnote at the end of section 5, reproduced here as published.

在该分布上，任何返回多数类的模型都能达到百分之七十一点四的一致率。我们的模型略好于此。我们不认为这是一个困难的结果。⁷

Any model that returns the majority class will achieve 71.4% agreement on this distribution. Ours does slightly better than that. We do not consider this a difficult result.⁷

本脚注发表于二〇二六年五月十四日，与论文正文同时，置于同一页的下方。此后四年，无人问及。

This footnote was published on 14 May 2026, at the same time as the body of the paper, at the bottom of the same page. In the four years that followed, no one asked about it.

该假定其后由本部门撤回。见更正。隐去索引的依据见匿名化规程第三条。

The assumption has since been withdrawn by this division; see Corrections. The basis for withholding the index is clause 3 of the anonymization protocol.

6 · 讨论Discussion

6.1 百分之九十四点一说明了什么What 94.1 per cent shows

它说明：该引擎的输出可由一个用九天、四千份样本和四十一千字节构建的外部系统重新产生。它不说明该引擎的内部结构与我们的相同。两个不同的系统可以产生同一个分布，这正是分布不能作为身份证明的原因。

It shows that the engine's output can be regenerated by an external system built in nine days from four thousand samples and forty-one kilobytes. It does not show that the engine's internals resemble ours. Two different systems can produce the same distribution, which is precisely why a distribution cannot serve as proof of identity.

6.2 百分之九十四点一不说明什么What it does not show

它不说明该引擎无效，也不说明该引擎有效。本文未评价任何一条判读的正确性，因为本部门不掌握判定其正确性所需的任何依据，而据我们所知，任何人都不掌握。

It does not show that the engine is invalid, and it does not show that it is valid. This paper evaluates the correctness of no interpretation, because this division holds nothing with which to determine correctness, and so far as we can establish, neither does anyone else.

6.3 关于可信度数值On the confidence value

该应用为每条读数附一个可信度百分数。我们未复现它，也未将其纳入比对，理由是：我们无法确定该数值所度量的对象。一个不知其所度量为何的数值，可以被匹配，但不能被复现，而两者的区别是本文的全部内容。

The application attaches a confidence percentage to each reading. We did not reproduce it and did not include it in the comparison, for the reason that we could not establish what the value measures. A number whose referent is unknown can be matched but not replicated, and the difference between those two things is the entire content of this paper.

6.4 分布的稳定性Stability of the distribution

莱茵兰占比自该应用公开发布以来保持在百分之七十一点四，精确到小数点后一位。分布会漂移。人口会变，设备会变，季节会改变人们的饮食。一个七年不动的分类占比，不是关于总体的事实，而是关于测量的事实：它意味着该系统始终以某一固定参考为基准作归一化。本文假定该参考为一份已被销毁或撤除的校准语料。该假定为我们所作。

The Rhineland share has held at 71.4 per cent, to one decimal place, since the application's public release. Distributions drift. Populations change, devices change, seasons change what people eat. A classification share that holds to one decimal for seven years is not a fact about the population; it is a fact about the measurement, and it implies that the system is being continuously normalized against some fixed reference. This paper assumes that reference to be a calibration corpus that has been destroyed or withdrawn. The assumption is ours.

7 · 局限Limitations

1. 本部门的四千份样本来自一个地理与年龄分布均较窄的群体。我们不主张其代表性，并已在附录 B 公布其构成。Our four thousand samples come from a population narrow in both geography and age. We claim no representativeness and have published the composition in appendix B.
2. 比对以该应用公开的分布为准，而非以其逐条输出为准。我们没有该应用的逐条输出，也没有请求过。The comparison is against the application's published distribution, not against its per-item output. We do not have its per-item output and have not asked for it.
3. 若该应用在本文发表后更改其引擎，本文的结果不因此失效，但也不再描述现状。本文描述的是二〇二六年五月的一个对象。If the application changes its engine after publication, our results do not become invalid, but they cease to describe the present. This paper describes an object as it stood in May 2026.
4. 本部门的标注者知道自己在参与一次复现。我们无法消除这一知晓，只能公布标注规程，以便他人在不知晓的条件下重做。Our labelers knew they were taking part in a replication. We cannot remove that knowledge; we can only publish the protocol so that others may redo the work without it.

8 · 结论Conclusion

该引擎的输出可以复现。复现它不困难。

The output of the engine can be reproduced. Reproducing it is not difficult.

附录、声明与参考文献Appendices, statements, and references

数据可用性Data availability

全部数据、代码、标注规程、种子与校验和自发表之日起公开，从未撤下。语料库的一次变更（二〇二七年移除十一份撤回样本）见数据与代码与语料索引。

All data, code, the labeling protocol, seeds, and checksums have been public since the day of publication and have never been withdrawn. The corpus’s one change — the removal of eleven withdrawn samples in 2027 — is described under Data & Code and Corpus index.

经费Funding

全部由本公司承担。金额与科目见经费。无外部资助，无委托方。

Borne entirely by this company. Amounts and line items are on the Funding page. No external sponsor, no commissioning party.

利益冲突Competing interests

无。本部门不持有该应用出版方的任何权益，不销售任何竞争产品，亦不销售任何产品。

None. This division holds no interest in the application's publisher, sells no competing product, and sells no product.

附录Appendices

• 附录 A · 标注规程（另作独立文件发布）Appendix A · The labeling protocol (also released as a standalone document)
• 附录 B · 语料构成：年龄、地域、设备、录制条件Appendix B · Corpus composition: age, geography, equipment, recording conditions
• 附录 C · 撤回记录，逐条，含日期Appendix C · Withdrawals, itemized, with dates
• 附录 D · 特征定义与计算式Appendix D · Feature definitions and formulae

参考文献References

1. 该应用之公开分布说明，发布方自述，访问日期 2026-03-02。The application's published distribution statement, as issued by the publisher, accessed 2026-03-02.
2. 该应用之公开方法说明，全文一句：“我们的方法为专有。”The application's published method statement, in its entirety: "Our method is proprietary."
3. 《排气语言学学报》，发行方即该应用之出版方，2019–2026 年各期。The Journal of Flatulinguistics, published by the application's publisher, issues 2019–2026.
4. 本部门标注规程 v1.0，2026-03。This division's labeling protocol, v1.0, March 2026.

本文于 2026-05-14 投稿至《排气语言学学报》，当日退回，未经评审。退稿单见收讫。

Submitted to the Journal of Flatulinguistics on 14 May 2026 and returned the same day, unrefereed. The return slip is under Receipts.

预印本 · 2030 年冬 / Preprint, winter 2030

某机构十年记录中的行政残差On an Administrative Residual in a Ten-Year Institutional Record

HYAI-VD-2030-01 · 41 pp. · 语料：2019–2029 年公开行政记录corpus: public administrative record, 2019–2029

林蔚然，陈砚，周汝成，罗有年，以及验证部全体。

Lin Weiran, Chen Yan, Zhou Rucheng, Luo Younian, and the Verification Division.

摘要Abstract

我们将某机构 2019 至 2029 年公开的行政记录作为语料处理，将每一件行政行为视为一个可分类事件，并按其记录是否标明发起方，分为已归属与未归属两类。语料含六万零四十五件行政行为。其中五万九千九百零四件可明确归属于该机构的五个下属机构。余下一百四十一件不能。本文对这一百四十一件制表（表四），检验其是否可由文书错误、分散的善意或机构自我纠正加以解释，并报告三者均不能。残差的年发生率在十年间的极差为一件，而没有任何人员编制模型能复现这一点——除一个模型之外。

We treat one institution's published administrative record, 2019–2029, as a corpus, classify each administrative act by whether the record identifies an originating party, and divide the corpus into attributed and unattributed classes. The corpus contains 60,045 administrative acts. Of these, 59,904 attribute cleanly to the institution's five constituent bodies. The remaining 141 do not. This paper tabulates those 141 (Table 4), tests whether they can be explained by clerical error, distributed goodwill, or institutional self-correction, and reports that none of the three accounts for them. The residual's annual rate has a range of one act across the decade, which no staffing model reproduces — except one.

1 · 引言Introduction

二〇二六年，本部门复现了该机构的解释引擎。本文与那项工作没有技术上的关系，只有方法上的关系：一份被完整发表的东西，应当被完整地读一遍。

In 2026 this division replicated the institution's interpretation engine. This paper bears no technical relation to that work, only a methodological one: a thing published in full ought to be read in full.

该机构自 2019 年起，逐年公开其行政记录：通信摘要、申领时间表、跨部门会议纪要、入藏目录、收讫确认。这些材料一直是公开的，任何人都可以取得，没有任何一份需要申请。据本部门统计，将其从头到尾完整读过一遍的，共有两人。本部门是第二个。第一个见于本文。

Since 2019 the institution has published its administrative record year by year: correspondence summaries, requisition schedules, cross-body session minutes, accession catalogues, and confirmations of receipt. These materials have been public throughout, obtainable by anyone, and none requires an application. By our count the record has had exactly two complete readers. This division is the second. The first is in this paper.

本文不是一次调查。本文是一次朗读。

This paper is not an investigation. It is a reading-aloud.

2 · 语料与处理Corpus and preparation

语料由十一个年度（2019 至 2029）的公开材料构成，合计八千一百二十四份文件。本文标题作“十年”。该称谓取自该机构本身对这一时期的叫法，本部门未予更改，并在此说明差额为一年。文件以原样保存，不作改写；每一份的取得日期、来源页面与校验和随文发布，以便他人取得同一份语料。若该机构此后修改或撤下任何一份，本部门所存版本的校验和将与之不符，而这一点本身即是一项可检验的事实。

The corpus consists of eleven calendar years of public materials, 2019 through 2029, 8,124 documents in total. The title of this paper says ten. The word is the institution’s own name for the period; we did not change it, and we note here that the difference is one year. Documents are retained as issued, without rewriting; the acquisition date, source page, and checksum of each are published so that others may obtain the same corpus. Should the institution later alter or withdraw any document, the checksum of our copy will no longer match, and that mismatch is itself a checkable fact.

该机构记录自身之完整，本文第二节称之为“堪称典范，且在若干处令人动容”。此语在本部门内部经过四稿。我们保留了它。

The institution documents itself with a completeness this paper calls, in section 2, "exemplary, and at points moving." The phrase survived four internal drafts. We kept it.

材料类别Document class	份数Docs	行政行为Acts
通信摘要Correspondence summaries	3,412	31,880
申领与采购时间表Requisition and procurement schedules	1,908	14,226
跨部门会议纪要Cross-body session minutes	1,144	6,015
入藏目录Accession catalogues	1,102	5,884
收讫确认与其他Confirmations of receipt and other	558	2,040
合计Total	8,124	60,045

3 · 归属规程Attribution protocol

我们将该机构 2019 至 2029 年间公开的行政记录作为语料处理，将每一件行政行为视为一个可分类事件。若记录标明发起的官员或主体，该行为记为“已归属”；若记录显示行为已执行而无发起方——有批准而无申请，有申领而无申领人，有维修而无工单，有致谢函而无立案文件——则记为“未归属”。我们对未归属类进行制表。

We treat the Institution's published administrative record, 2019–2029, as a corpus, and each administrative act as a classifiable event. An act is ATTRIBUTED where the record identifies an originating officer or body. An act is UNATTRIBUTED where the record shows execution without origination — an approval with no application, a requisition with no requesting officer, a repair with no work order, a letter of thanks with no triggering file. We tabulate the unattributed class.

归属由两名成员独立判定，不一致处由第三人裁定。判定的一致率为百分之九十九点二，即四百八十件存在不一致；全部不一致条目及其裁定见附录 B，逐条列出，包括裁定与初判相反的六十三件。

Attribution was determined independently by two members, with disagreements resolved by a third. Inter-rater agreement was 99.2 per cent, that is, 480 acts on which the two disagreed. Every disagreement and its resolution is listed itemized in appendix B, including the sixty-three in which the resolution went against the first reading.

一条排除规则值得指出，因为它其后受到质疑，而质疑是对的：凡记录本身注明“发起方待补”者，不计入未归属类。我们采用该规则，是因为待补是一种承诺而非一种缺失。此类条目共二十二件，单独列于附录 C，不计入表四。任何人若认为应当计入，可自附录 C 取得全部二十二件，自行加总。

One exclusion rule is worth naming, because it was later questioned and the question was fair: acts whose record itself carries the notation "originating party to follow" are not counted in the unattributed class. We adopted the rule because a notation of that kind is a promise rather than an absence. Twenty-two acts fall under it. They are listed separately in appendix C and are not in Table 4. Anyone who believes they belong in it may take all twenty-two from appendix C and add them.

4 · 结果Results

4.1 总体Overall

类别Class	件数Acts	占比Share
已归属Attributed	59,904	99.765%
未归属Unattributed	141	0.235%
合计Total	60,045	100%

4.2 表四 · 未归属行政行为，2019–2029，按类别Table 4 · Unattributed administrative acts, 2019–2029, by category

类别Category	件数Count
常规申领单上的会签，会签人身份无法确认（含每年一次的档案手套，2019–2029）Countersignatures on routine requisitions, countersigning officer unidentifiable (incl. archival gloves, annually, 2019–2029)	87
以声明未曾提交者之名义提交的申领单Requisitions filed in the name of officers who state they did not file them	3
邻近社会团体的会籍与订阅，经邮寄办理，仅具名，未具姓Memberships and subscriptions in adjacent civic bodies, entered by post, given name only	6
经邮寄在邻近团体投出的委托票Proxy votes cast by post in bodies adjacent to the Institution	4
向社会团体委员会的书面提名，申请人未披露Written nominations to civic committees, applicant undisclosed	1
无工单的维修与小型工程（窗户五金、椅脚毡等）Repairs and small works performed without work order (window furniture, chair felt, and similar)	9
无立案文件的致谢函（最早为 2027 年；积压系列）Letters of thanks issued without a triggering file (earliest 2027; the backlog series)	30
匿名提交并被搁置的政策文件Policy documents submitted anonymously and tabled	1
合计Total	141

请读者自行核对此处合计。这正是本文发表的目的。各行的工作底稿见工作底稿。

The reader is invited to check the addition. That is the entire point of the paper. The working papers behind each row are under Working papers.

4.3 表五 · 未归属行为按年度Table 5 · Unattributed acts by year

年度Year	2019	2020	2021	2022	2023	2024	2025	2026	2027	2028	2029	合计Total
件数Acts	12	13	13	13	13	12	13	13	13	13	13	141

十一个年度，最大十三件，最小十二件，极差一件。本部门检索了适用于该机构规模的全部标准人员编制模型，没有一个能产生极差为一的十一年序列——除一个之外：该模型含有一名额外的、连续的、无报酬的行政经办人。我们不主张该模型为真。我们报告的是：它是唯一能够拟合的模型。

Eleven years; maximum thirteen, minimum twelve, range one. We searched the standard staffing models applicable to an institution of this size and found none that produces an eleven-year series with a range of one — except one: a model containing one additional, continuous, uncompensated administrative agent. We do not assert that the model is true. We report that it is the only one that fits.

4.4 本文对表四的全部评述The paper's commentary on Table 4, in full

残差很小。残差是恒定的：其年发生率在整个十年中的波动小于每年一件，除“恰好多出一名连续的、无报酬的行政执行者”这一模型外，没有任何人员编制模型能复现这一点。残差不是随机的：其各项行为在类别层面共享一个特征——每一项都提供了一种功能，而五个下属机构均无对应的职能部门。我们考虑并排除了：文书错误（这些行为已被执行，而非被误记）；分散的善意（善意不是恒定的，见图六）；以及机构的自我纠正（在同一十年的记录上，该机构不作纠正，见该记录之全部）。我们由此得到一个可以陈述、却无法确定归属的发现。残差不是错误。残差是一种惯例。

The residual is small. The residual is constant: its annual rate varies by less than one act per year across the decade, which no staffing model reproduces except a model containing exactly one additional, continuous, uncompensated administrative agent. The residual is not random: its acts share a signature at the level of category — each supplies a function the five bodies possess no office for. We considered and rejected: clerical error (the acts are executed, not misrecorded); distributed goodwill (goodwill is not constant; see fig. 6); and institutional self-correction (the Institution, on the record of the same decade, does not correct; see the whole of the record). We are left with a finding we can state but not attribute. The residual is not an error. The residual is a practice.

本文不为该功能命名。命名它就需要确定其归属，而我们无法确定其归属。

The paper declines to name the function. Naming it would require attributing it, and we cannot attribute it.

5 · 备选解释Rival explanations

5.1 文书错误Clerical error

若这一百四十一件为记录错误，则相应的行为应当不曾发生。我们对其中可实地核验的三十九件逐一核验：手套已购入并使用，椅脚毡已装上，窗户五金已更换，会籍已生效并缴费，委托票已计入表决结果。行为已被执行，而非被误记。文书错误排除。

If the 141 were recording errors, the corresponding acts would not have happened. We checked the thirty-nine that admit of physical verification, one by one: the gloves were bought and used, the chair felt was fitted, the window furniture was replaced, the memberships were live and paid, and the proxy votes are counted in the recorded outcomes. The acts are executed, not misrecorded. Clerical error is excluded.

5.2 分散的善意Distributed goodwill

若这些行为出自多人各自的善意，则其年度序列应当随人员变动、工作压力与季节波动。图六以该机构同期的在职人数、离职率与会议密度作对照，三者均有可观波动，而残差序列的极差为一。分散的善意不是恒定的。我们没有见过恒定的善意。排除。

If these acts arose from the goodwill of many people acting separately, the annual series should move with staff turnover, workload, and season. Figure 6 plots the institution's headcount, attrition, and meeting density over the same period; all three move appreciably, while the residual series has a range of one. Distributed goodwill is not constant. We have never observed constant goodwill. Excluded.

5.3 机构的自我纠正Institutional self-correction

若该机构会自行填补其五个下属机构未覆盖的职能，则十年的记录中应当留下痕迹：一项动议、一次授权、一条编制。没有。在同一十年的记录上，该机构不作纠正。此项排除的依据是该记录之全部，而该记录是公开的。

If the institution corrected for functions its five bodies do not cover, the decade would show it: a motion, an authorization, an establishment line. There is none. On the record of the same decade, the institution does not correct. The basis for this exclusion is the whole of the record, and the record is public.

5.4 我们未能排除的What we could not exclude

我们无法排除该经办人不止一人。年度序列与“一人”相容，也与“一人及一名极少留下痕迹的协作者”相容。本文按最小假设写作，而最小假设不等于事实。此项局限其后被人指出，指出的方式见第七节。

We cannot exclude that the agent is more than one person. The annual series is consistent with one, and also with one plus a collaborator who leaves almost no trace. This paper is written on the smallest assumption, and the smallest assumption is not the same as the fact. The limitation was later pointed out to us; the manner of it is section 7.

6 · 局限Limitations

1. 本文只能看见留下纸面的行为。任何未留下纸面的行为，以本文的方法一件也看不见。这不是一项技术局限，而是本方法的定义。This paper can see only acts that left paper. Any act that left none is invisible to this method, not as a technical limitation but by definition of the method.
2. 语料为该机构自行发布的材料。若其发布不完整，本文的一百四十一件即为低估。我们无法检验其发布是否完整，因为检验的唯一材料就是其发布。The corpus is what the institution itself published. If its publication is incomplete, our 141 is an undercount. We cannot test the completeness of the publication, because the only material for such a test is the publication.
3. 排除规则见第 3 节末。二十二件“发起方待补”未计入。该规则可争，附录 C 使争论可以进行。The exclusion rule is at the end of section 3. Twenty-two "originating party to follow" acts are not counted. The rule is arguable, and appendix C makes the argument possible.
4. 本部门是外国机构，不熟悉该机构所在司法管辖区的行政惯例。凡我们可能因此误判之处，均在附录 B 标出。This division is foreign to the institution's jurisdiction and unfamiliar with its administrative conventions. Every point at which that may have caused a misreading is flagged in appendix B.

7 · 收到的更正The correction received

预印本流传九天后，本部门收到一个信封。普通白纸，无信头，无日期，无称谓，无签名。两段。

Nine days after the preprint circulated, this division received an envelope. Plain paper, no letterhead, no date, no salutation, no signature. Two paragraphs.

第一段就本文方法作出评述。它认可制表规程，质疑第三节末的那条排除规则，并在该段最后一句中自行让步。本部门此前从未从任何审稿人处收到过这样一段论证：两面都论得诚实，而两面都不具名。

The first paragraph engages the method. It endorses the tabulation protocol, questions the exclusion rule at the end of section 3, and concedes the exclusion rule in its own final sentence. This division has never received from any referee an argument that argues both sides honestly and signs neither.

第二段更正了一个数字。全文如下：

The second paragraph corrects one number. It is reproduced entire:

本部门依据公开记录重新制表四次，结果仍为一百四十一。无论第一百四十二件行为是什么，以我们掌握的任何方法都看不见它。本部门附于预印本的答复全文一句：

This division re-ran the tabulation four times against the public record and stands by 141. Whatever the 142nd act is, it is invisible to any method we possess. Our published response, appended to the preprint, is one sentence long:

我们接受这项更正——一项我们无法核实的更正，来自一个我们无法辨认的来源——理由是：被审计的领域显然一向如此运作，而它拥有我们所测量过的最好的复现记录。

We accept the correction, which we cannot verify, from a source we cannot identify, on the grounds that this is evidently how the field under audit has always worked, and it has the best replication record we have ever measured.

两个数字均予刊出。表四维持发表时原样，合计一百四十一。本节同时载明一百四十二。本部门不认为二者需要调和。来信原件已入本部门档案，依匿名化规程第五条保存；我们无法确认收到，因为我们没有可寄达的地址。

Both figures are published. Table 4 stands as published, at 141. This section also states 142. This division does not regard the two as requiring reconciliation. The original letter is held in our files under clause 5 of the anonymization protocol. We were unable to confirm receipt, having no address to confirm it to.

8 · 结论Conclusion

该机构的五个下属机构可以解释其全部职能，唯有一项除外，而该项仍在被履行。

The five bodies of the Institution account for the Institution's every function except one, which is nonetheless performed.

附录、声明与参考文献Appendices, statements, and references

数据可用性Data availability

语料、逐件归属判定、裁定记录、图六的底层数据与全部校验和均已公开。见数据与代码。

The corpus, the per-act attribution determinations, the resolutions, the underlying data for figure 6, and all checksums are public. See Data & Code.

利益冲突Competing interests

本部门与该机构之间存在一段公开的争议史：该机构于二〇二六年就本部门的复现工作提出盗用指控，并聘请了律师。本部门认为该事实应当由读者知晓，并据以对本文作更严格的审视。我们建议读者这样做。

There is a public history of dispute between this division and the institution: in 2026 the institution alleged misappropriation in respect of our replication and retained counsel. This division considers that the reader should know it and should read this paper more severely because of it. We recommend that the reader do so.

附录Appendices

• 附录 A · 语料清单：八千一百二十四份文件，含取得日期、来源与校验和Appendix A · Corpus manifest: 8,124 documents, with acquisition date, source, and checksum
• 附录 B · 逐件归属判定，含全部四百八十件不一致与裁定，及六十三件反转Appendix B · Per-act attribution determinations, with all 480 disagreements, their resolutions, and the sixty-three reversals
• 附录 C · “发起方待补”二十二件，逐条Appendix C · The twenty-two "originating party to follow" acts, itemized
• 附录 D · 图六底层数据：在职人数、离职率、会议密度Appendix D · Underlying data for figure 6: headcount, attrition, meeting density
• 附录 E · 人员编制模型检索：所检模型、拟合结果、被排除的理由Appendix E · The staffing-model search: models tested, fits obtained, grounds for rejection

本文投稿至《排气语言学学报》，当日退回，未经评审。退稿单与二〇二六年的一份一字不差。见收讫。

Submitted to the Journal of Flatulinguistics and returned the same day, unrefereed, on a slip identical word for word to the 2026 one. See Receipts.

审计 · 2031 年秋 / Audit, autumn 2031

对某消费级读取设备中一项未署名衍生关系的审计On an Unattributed Derivation in a Consumer Reading Device

HYAI-VD-2031-01 · 23 pp. · 用时九天 · 附全部代码与数据，与以往相同nine days · full code and data, as before

林蔚然，陈砚，周汝成，罗有年，以及验证部全体。

Lin Weiran, Chen Yan, Zhou Rucheng, Luo Younian, and the Verification Division.

摘要Abstract

安调健康公司（Attune Health, Inc.）所售 Auricle Pro 设备，其引擎为本部门二〇二六年公开发布之四千份样本语料库的重新加权。我们公布认定该衍生关系所依据的方法，以便他人复核。该语料库的许可要求署名，别无其他条件。产品、认证、备案文件与营销材料中均未出现署名。本审计另确定该设备之认证机构的所有权结构；该结构未被隐瞒，全部事实均见于公开备案文件，而备案文件是免费的。

The engine of the Auricle Pro device, sold by Attune Health, Inc., is a re-weighting of the 4,000-sample corpus this division released openly in 2026. We publish the method by which the derivation was established, so that others may check it. The corpus license requires attribution and nothing else. No attribution appears in the product, the certification, the filings, or the marketing. The audit also establishes the ownership structure of the device's certifying body. That structure is not concealed: every fact is in the public filings, and the filings are free.

1 · 引言Introduction

本部门发布语料库时附一项条件：署名。我们不追踪使用者以执行该条件，因为我们不打算执行它。我们追踪使用情况，是因为发表一个语料库而不追踪它的去向，等于发表了一半。

This division released the corpus on one condition: attribution. We do not track users in order to enforce it, because we do not intend to enforce it. We track use because to publish a corpus and not follow where it goes is to have published half of it.

二〇三一年春，一位来信者告知本部门，某消费级设备的输出与本语料库的若干特有性质相符。我们购入一台设备，自费，按零售价，历时九天。

In the spring of 2031 a correspondent informed this division that the output of a consumer device matched several idiosyncratic properties of our corpus. We bought a device, with our own money, at retail, and spent nine days.

2 · 认定The finding

该设备的引擎为本部门二〇二六年公开发布之四千份样本语料库的重新加权；该语料库的许可要求署名，别无其他条件；产品、认证、备案文件与营销材料中均未出现署名。

The engine is a re-weighting of the 4,000-sample corpus this division released openly in 2026 under a license requiring attribution and nothing else. No attribution appears anywhere in the product, the certification, the filings, or the marketing.

以上为本文第二节全文。本部门就此不加形容词。

The above is the whole of section 2. This division attaches no adjective to it.

3 · 方法：一个语料库如何在其衍生物中留下印记Method: how a corpus leaves a mark in what is derived from it

一个语料库不只承载信号，也承载它自己的意外。判定衍生关系，不看两者何处相同——相同可以来自同一个世界——而看两者何处以同样的方式出错。本部门采用三项印记，每一项都可由任何持有本语料库与一台设备的人重做。

A corpus carries not only signal but its own accidents. A derivation is established not by where two systems agree — agreement can come from sharing a world — but by where they are wrong in the same way. This division used three marks. Each can be redone by anyone holding our corpus and a device.

3.1 十九处反转The nineteen reversals

本语料库的标注由两人独立作出，不一致处由第三人裁定。裁定中有十九条与初判相反。这十九条是人的判断，不是世界的性质：换一位裁定者，可能得到不同的十九条。任何独立训练的系统，在这十九条上应当与初判、裁定各半。该设备在十九条中的十八条上与裁定一致。在独立性假设下，该结果的概率为零点零零零零七。

Our corpus was labeled independently by two people, with disagreements resolved by a third. In nineteen cases the resolution went against the first reading. Those nineteen are human judgments, not properties of the world: a different adjudicator might have produced a different nineteen. Any independently trained system should split roughly evenly between the first readings and the resolutions. The device agrees with the resolution in eighteen of the nineteen. Under an assumption of independence, the probability of that result is 0.00007.

3.2 被撤回的十一份The eleven withdrawals

本语料库有十一份样本经志愿者撤回，并自语料库移除，移除日期公开。撤回发生在二〇二七年；此前下载的副本仍含这十一份。该设备的行为与“含十一份”的版本一致，与“不含”的版本不一致。这确定的不是衍生关系本身，而是衍生发生的时间：早于二〇二七年。

Eleven samples in our corpus were withdrawn by their volunteers and removed, with the removal dates published. The withdrawals occurred in 2027; copies downloaded before then still contain the eleven. The device's behavior is consistent with the version that contains them and inconsistent with the version that does not. This establishes not the derivation but its date: before 2027.

本部门已就此另行去信，请求该公司自其训练材料中移除该十一份。此请求与署名无关，亦不在署名的四封信之内，是第五封，单独计。同样未获答复。撤回同意的人，其撤回应当到达它需要到达的地方。

This division wrote separately to ask the company to remove those eleven from its training materials. That request has nothing to do with attribution, is not among the four letters, and is counted separately as a fifth. It also received no reply. A person who withdraws consent is entitled to have the withdrawal arrive where it needs to arrive.

3.3 那支二十元的话筒The twenty-yuan microphone

本语料库以消费级手机与一支廉价驻极体话筒录制，并在两路中取较差的一路。该话筒在 3.2 kHz 附近有一处约 2 dB 的凸起，是这一型号的缺陷，并非世界的性质。该设备的响应在同一频段有同样的偏差。以另一支同型号话筒可复现此项；以任何其他话筒则不能。

Our corpus was recorded on consumer telephones and one cheap electret microphone, keeping whichever channel was worse. That microphone has a bump of about two decibels near 3.2 kHz — a defect of the model, not a property of the world. The device's response carries the same bias in the same band. The finding reproduces with another microphone of the same model and with no other.

4 · 认证机构The certifying body

该设备的包装载明其“经独立认证”。发出该认证的机构，以一位在世个人的姓名命名。本部门不在本文中印出该姓名。该个人与该机构无关，未曾同意，且已两次书面拒绝一项相关邀请；加于他的全部损害，正在于他的姓名被印出。本文以“〔姓名已隐去〕开放读取研究院”称之。

The device's packaging states that it is "independently certified." The body issuing that certification is named after a living individual. This division does not print his name here. The individual has no connection to the body, gave no consent, and has twice declined a related invitation in writing; the whole of the harm done to him consists in his name being printed. This paper calls it the [name withheld] Institute for Open Reading.

以下四项事实，均取自免费的公开备案文件。

The following four facts are taken from public filings, which are free.

#	事实Fact	来源Source
1	该研究院为该制造商的全资子公司，二〇三〇年设立。The Institute is a wholly owned subsidiary of the manufacturer, chartered in 2030.	设立登记文件Charter filing
2	设立后第十一日发出第一份证书，发予其母公司。It issued its first certificate eleven days later, to its parent.	证书登记簿Certificate register
3	其评审组由三名“经本院认证的独立审核员”组成，认证由本院发出。Its review panel comprises three "Institute-Certified Independent Auditors," certified by the Institute.	评审组名录Panel roster
4	其中两名受雇于一家鉴证服务公司，该公司与母公司共用注册代理人、通信地址与首席财务官。Two of the three are employed by an assurance company sharing a registered agent, a mailing address, and a chief financial officer with the parent.	年度报告Annual filings

以上没有一项被隐瞒。四项全部见于备案文件，而备案文件是免费的。本部门所做的，只是把它们读完了。

None of this is concealed. All four facts are in the filings, and the filings are free. What this division did was read them to the end.

该研究院所公布的标准，全文一页。本部门读过。本文不评价该标准，因为评价一项标准的前提是它规定了某件事。

The Institute's published standard runs to one page. This division has read it. This paper does not evaluate the standard, because evaluating a standard presupposes that it requires something.

5 · 去信记录Record of letters

四封，均为挂号信，均经签收。请求事项每次相同：在产品、认证或营销材料的任一处署名。未提出赔偿要求，未设定期限，未提出其他任何要求。完整记录见署名情况。

Four letters, all registered, all signed for. The request was the same each time: attribution, in any one of the product, the certification, or the marketing. No damages sought, no deadline set, nothing else asked. The full record is under Attribution.

许可只要求一件事。我们已去信四次。我们没有被拒绝。我们没有被受理。我没有想到自己会想念那个机构。

The license asked for one thing. We have written to them four times. We have not been refused. We have not been received. I did not expect to miss the Institution.

——验证部负责人，应本次发表之请求所供，三十一词。

— Head of division, supplied on request for this publication; thirty-one words.

6 · 本审计不做的事What this audit does not do

• 不主张损害赔偿。It claims no damages.
• 不请求救济。It requests no remedy.
• 不建议任何行动。It recommends no action.
• 不就任何人的动机作出认定。本部门不掌握动机的数据。It makes no finding as to anyone's motives. This division holds no data on motives.
• 不主张该设备无效。本审计确定其来源，不评价其性能。It does not assert that the device is ineffective. The audit establishes provenance, not performance.
• 不更改本语料库的许可条件。因一位使用者未予署名而收紧条件，等于惩罚全部已署名的使用者，以回应一件他们没有做的事。It does not change the corpus license. To tighten the terms because one user did not attribute would penalize every user who did, in response to something they did not do.

7 · 局限Limitations

1. 三项印记确定的是衍生关系，不是抄袭的路径。该公司可能自第三方取得本语料库，而该第三方已署名。若如此，则本文的认定不变，而应受此认定的对象改变。我们已在去信中说明这一可能，并请求告知。未获答复。The three marks establish derivation, not the route of it. The company may have obtained our corpus from a third party who did attribute. If so, the finding stands and the party to whom it applies changes. We said so in our letters and asked to be told. No reply.
2. 3.1 的概率计算假定十九条彼此独立。它们可能不完全独立。附录 A 以最保守的相关性假设重算，结果为零点零零四，结论不变。The probability in 3.1 assumes the nineteen reversals are independent. They may not be entirely. Appendix A recomputes under the most conservative correlation assumption; the result is 0.004 and the conclusion is unchanged.
3. 本部门只购入一台设备，并只在一个固件版本上测试。其他版本可能表现不同。设备序列号、固件版本与购买凭证均已公布。This division bought one device and tested one firmware version. Other versions may behave differently. The serial number, firmware version, and proof of purchase are published.

8 · 结论Conclusion

该引擎衍生自本语料库。该语料库仍然公开，条件不变。

The engine is derived from our corpus. The corpus remains public, on the same terms.

声明与附录Statements and appendices

数据可用性Data availability

测量脚本、原始录音、频响曲线、十九条反转的逐条比对、备案文件的取得记录与全部校验和均已公开。见数据与代码。

The measurement scripts, the raw recordings, the frequency-response curves, the item-by-item comparison of the nineteen reversals, the acquisition record for the filings, and all checksums are public. See Data & Code.

利益冲突Competing interests

本部门是被衍生语料库的发布方，因而不是一个中立的当事方。这是本文最重要的一项声明，我们把它放在读者会读到的位置。本文所依据的三项印记，均可由与本部门无关的第三人重做，而这正是我们公布方法而非仅公布结论的原因。

This division published the corpus from which the derivation was made and is therefore not a disinterested party. This is the most important statement in the paper, and we have put it where the reader will read it. All three marks can be redone by a third party unconnected with this division, which is why we publish the method and not only the conclusion.

附录Appendices

• 附录 A · 3.1 的概率计算，含保守相关性重算Appendix A · The probability calculation in 3.1, with the conservative recomputation
• 附录 B · 十九条反转，逐条，含两名标注者的初判与裁定Appendix B · The nineteen reversals, itemized, with both first readings and the resolution
• 附录 C · 频响测量：设备、话筒、方法、曲线Appendix C · Frequency-response measurement: device, microphone, method, curves
• 附录 D · 备案文件清单与取得记录Appendix D · Manifest of filings, with acquisition record
• 附录 E · 五封去信全文，含挂号回执Appendix E · The five letters in full, with registered-post receipts

本文未投稿。理由见论文。

Not submitted. The reason is under Papers.

数据与代码 / Data & Code

数据与代码Data & Code

全部材料自首次发表之日起公开，未曾撤下。语料库变更过一次：二〇二七年移除十一份经撤回的样本。变更有日期，旧校验和保留于 CHECKSUMS.txt，下表所列为现行版本的校验和。除此之外，未有任何修改。

Everything below has been public since the day of first publication and nothing has been withdrawn from publication. The corpus has changed once: eleven withdrawn samples were removed in 2027. The change is dated, the earlier checksums are retained in CHECKSUMS.txt, and the checksums below are those of the current release. Apart from that, nothing has been altered.

文件File	大小Size	SHA-256（前 16 位）SHA-256 (first 16)
corpus-4000.tar.zst	1.84 GB	4f1a90c37b6e2d05
labels-4000.csv	312 KB	a07d5c81e4b93f2a
labeling-protocol.pdf	96 KB	c93b02fa7e15d846
classifier/	41 KB	18e6b4d0aa27c9f3
lookup-table.tsv	7 KB	6b2ee5c1903a7d84
seeds.txt	142 B	0d47a3ce8f61b209
audit-2031/administrative-corpus/	88 MB	e5c1470b93da6f28
CHECKSUMS.txt	1 KB	—

许可Licence

署名。仅此一项。

Attribution. That is the whole of it.

可用于商业用途。可修改。可再分发。可用于训练。无需征得许可，无需通知我们，无需付费。请写上我们的名字。

Commercial use permitted. Modification permitted. Redistribution permitted. Training permitted. No permission required, no notification required, no payment required. Please write our name.

关于四千份样本的来源On the provenance of the four thousand samples

全部样本由成年志愿者在知情同意下提供，一次一份，附书面同意书，可随时撤回，撤回无需说明理由。撤回记录见语料附录 C。截至本次更新，共有十一份样本被撤回，并已自语料中移除；该语料库因此含三千九百八十九份样本，而我们仍称之为四千份语料，因为发表时的名称不应因后来的减少而改变。此说明即是差额本身。

All samples were supplied by informed adult volunteers, one at a time, under written consent, withdrawable at any time and without reason given. Withdrawals are recorded in appendix C of the corpus. As of this update, eleven samples have been withdrawn and removed; the corpus therefore contains three thousand nine hundred and eighty-nine. We still call it the 4,000-sample corpus, because a name given at publication should not change with a later subtraction. This note is the difference.

复现说明 / How to check our work

如何复现How to reproduce this

下列步骤在一台普通计算机上约需四十分钟，其中三十七分钟用于下载。不需要专用硬件。

The steps below take about forty minutes on an ordinary computer, thirty-seven of which are the download. No special hardware is required.

最后一行不是错误。见二〇二六年论文的脚注七。

The last line is not an error. See footnote 7 of the 2026 paper.

若你得到不同的结果If you get a different result

请来信。请附你的种子、你的机器、你的完整输出，以及你认为分歧所在之处。若你是对的，我们会更正，并写上你的名字；若你不愿具名，我们会照办，并注明更正来自一位要求不具名的通信者。两种情形都不必道歉。

Write to us. Include your seed, your machine, your full output, and where you believe the divergence lies. If you are right we will correct the record and write your name in it; if you would rather not be named, we will do that instead and record that the correction came from a correspondent who asked not to be named. Neither case requires an apology.

若你打算指控我们If you intend to accuse us

同样欢迎。请附上同样的材料。一项可核查的指控与一项可核查的更正，在本部门内部适用同一流程，由同一批人处理，并在同一页面公布。

Also welcome. Please attach the same material. Within this division a checkable accusation and a checkable correction follow the same procedure, are handled by the same people, and are published on the same page.

署名情况 / Attribution status

署名情况Attribution status

本部门追踪该语料库已知的下游使用情况，不是为了强制执行许可条款——我们不打算强制执行——而是因为发表一个语料库而不追踪它的去向，等于发表了一半。

This division tracks known downstream use of the corpus. Not in order to enforce the license — we do not intend to enforce it — but because to publish a corpus and not follow where it goes is to have published half of it.

使用者User	用途Use	署名Attributed
十四所大学与研究机构Fourteen universities and research institutes	教学、再分析、方法学Teaching, re-analysis, methods work	是yes
独立研究者（六十一位，累计）Independent researchers (sixty-one, cumulative)	各类Various	是yes
两家发表了竞争性复现研究的实验室Two laboratories publishing competing replications	复现，其中一项与我们不一致Replication, one of them disagreeing with us	是yes
安调健康公司（Attune Health, Inc.）Attune Health, Inc.	商用产品引擎Engine of a commercial product	否no

与我们不一致的那项复现，我们已在更正页列出，一并列出他们正确而我们错误的那一处。

The replication that disagrees with us is listed under Corrections, together with the one point on which they were right and we were not.

去信记录Record of letters

四封。全部为挂号信，全部经签收。请求事项每次相同：在产品、认证或营销材料的任一处署名。未提出赔偿要求。未设定期限。

Four. All registered, all signed for. The request was the same each time: attribution, in any one of the product, the certification, or the marketing. No damages sought. No deadline set.

序号No.	方式Method	送达Delivered	答复Reply
1	挂号信Registered post	已签收signed for	无none
2	挂号信Registered post	已签收signed for	无none
3	挂号信Registered post	已签收signed for	无none
4	挂号信Registered post	已签收signed for	无none

我们没有被拒绝。我们没有被受理。

We have not been refused. We have not been received.

拒绝是一种答复。拒绝会产生一份记录、一个相对方、一项事项，而事项会进入待办队列。我们所去信的地方没有队列。这不是一项指控；这是审计中最难写的一项发现，因为它不是任何人做了什么，而是没有任何地方发生任何事。

A refusal is a reply. A refusal creates a record, a party, a matter; and a matter enters a queue. Where we wrote, there is no queue. This is not an accusation. It is the hardest finding in the audit to write, because it is not something anyone did. It is nothing happening anywhere.

本部门就此事的立场The division's position

署名不花钱，不需要许可，不占用产品上任何一处显眼的位置。我们请求过四次。我们不会请求第五次。记录留在此页。

Attribution costs nothing, requires no permission, and occupies no prominent surface of any product. We have asked four times. We will not ask a fifth. The record stays on this page.

更正 / Corrections

更正Corrections

本页自二〇二六年起持续记录，从未删减。本页上的更正不作归档处理，因为一份已归档的更正就等于没有更正。

This page has run since 2026 and nothing has been removed from it. Corrections are not archived off this page, because a correction that has been archived is not a correction.

收讫 / Receipts

本部门确认收到一切This division confirms receipt of everything

这是一项惯例，不是一项政策。我们在理解它之前，就已在受审计对象处见其可取，此后予以采用。

This is a practice, not a policy. We admired it in our subjects before we understood it, and then we adopted it.

前两个数字相同。这不是一项成就，而是本惯例的定义。第三个数字少一。差额为一封：一个未署名、无回邮地址的信封，本部门无法答复，亦无法向其确认收到。该封的收讫记录即本页下方的最后一行。本页承诺在三者不同时刊出差额，此处即是。

The first two figures are identical. This is not an achievement; it is the definition of the practice. The third is short by one. The difference is a single item: an unsigned envelope with no return address, which this division could not answer and to which it could not confirm receipt. Its entry is the last line at the foot of this page. This page undertook to print the difference if the figures ever differed. This is the difference.

墙上的那一份The one on the wall

二〇二六年，我们向该领域的期刊投出一篇复现论文。该稿当日退回，未经评审。二〇三一年，我们投出了那份审计。该稿当日退回，未经评审，退稿单一字不差：

In 2026 we submitted a replication to the field's journal. It was returned the same day, unrefereed. In 2031 we submitted the audit. It was returned the same day, unrefereed, on a slip identical word for word:

该刊主编就此说明：在不同刺激下给出一致的编辑答复，正是标准的操作性定义。他们测试了我们两次，我们复现了。

The journal's editor has explained the position: consistency of editorial response under varying provocation is the operational definition of a standard. They tested us twice; we replicated.

本部门同意这一说法。该陈述在相隔五年的两次独立试验中逐字复现，本部门未见本领域有第二例。我们将该退稿单装框保存，与其他复现结果同列。

We agree with the position. The statement reproduced verbatim across two independent trials five years apart, and this division knows of no second instance of that in this field. We framed the slip and hung it with our other replications.

其他已确认收到的事项，选录Other receipts confirmed, selected

• 某地方俱乐部委员会来函一封，内容为其在审计中出现一事。已答复。该俱乐部的名称依匿名化规程隐去。A letter from the committee of a local club, on the subject of its appearance in the audit. Answered. The club's name is withheld under our anonymization protocol.
• 一位独立研究者来函一封，询问其二〇二四年的电路是否被本部门用于任何用途。未曾使用。已如实答复，并附一份我们确实使用过的材料清单。A letter from an independent researcher asking whether this division had used his 2024 circuit for anything. We had not. Answered, with a list of what we had used.
• 该应用出版方法务部门来函十一封。全部已确认收到，全部已答复，全部保存。Eleven letters from the legal office of the application's publisher. All receipted, all answered, all kept.
• 一个信封，素纸，无签名，两段。已确认收到，收件地址不详，无法寄出。此行即为该次确认。One envelope, plain paper, unsigned, two paragraphs. Receipt confirmed to an address we do not have and could not send to. This line is that confirmation.

常见问题（已答复） / Frequently asked questions, answered

常见问题Frequently asked questions

本页每一个问题都有答案。若你的问题不在此处，请来信，我们会作答，并将其补入本页。

Every question on this page has an answer. If yours is not here, write to us; we will answer it and add it.

你们是不是对他们的引擎做了逆向工程？Did you reverse-engineer their engine?

没有。我们录了四千份样本，做了标注，训练了一个模型。全部已公开。

No. We recorded four thousand samples, we labeled them, and we trained a model. It is all published.

94.1% 是不是很高？Is 94.1% a high number?

不高。每次都猜多数类可以得到 71.4%。

No. Guessing the majority class every time gets you 71.4%.

那你们为什么还要发表？Then why publish it?

因为一个不困难的结果，只有在有人把它做出来并给出数字之后，才被认为是不困难的。

Because a result is only known to be undifficult after somebody has done it and printed the number.

你们申请专利了吗？Have you patented any of this?

没有可申请的东西。一个小型分类器加一张查找表。

There is nothing to patent. A small classifier and a lookup table.

你们收费吗？Do you charge for any of it?

不收。

No.

谁资助这项工作？Who funds the work?

本公司。经费来源与金额列于每篇论文的首页，与作者姓名并列。

This company. Sources and amounts are on the first page of every paper, beside the authors' names.

你们讨厌那家公司吗？Do you dislike the company?

不。

No.

你们讨厌那个机构吗？Do you dislike the Institution?

不。他们称我们的工作为盗窃，并聘请了律师。他们也确认收到了我们寄去的每一封信，每一次，始终在三十日之内，并且至今仍保存着全部信件。这两件事都是真的，我们不认为需要在二者之间作出选择。

No. They called our work theft and retained counsel. They also confirmed receipt of every letter we ever sent them, every time, within thirty days, and they have all of it still. Both are true, and we do not consider ourselves obliged to choose between them.

脚注七是什么意思？What does footnote 7 mean?

我们不知道。我们记录了它，不作解释。

We do not know. We recorded it without explanation.

那第一百四十二件行政行为是什么？What is the 142nd act?

我们不知道。以我们掌握的任何方法都看不见它。我们仍然刊出这个数字。

We do not know. It is invisible to any method we possess. We publish the figure anyway.

如果我证明你们错了呢？What if I prove you wrong?

请来信。见如何复现与更正。

Write to us. See How to reproduce this and Corrections.

你们会回信吗？Will you reply?

会。

Yes.

关于 / About

关于本部门About this division

宏源应用智能验证部，深圳。成立于二〇二四年，现有九人。本部门不开发产品。本部门核查他人已发表之工作，并发表核查的全部过程。

The Verification Division of Hongyuan Applied Intelligence, Shenzhen. Established 2024. Nine people. The division builds no products. It checks published work, and publishes the checking.

人员People

林蔚然，研究员，验证部负责人。二〇二六年复现与二〇三〇年审计的第一作者。以电子邮件详尽作答，接受采访则简短，凡事均加脚注。

Dr. Lin Weiran, Senior Researcher and head of the division. Lead author of the 2026 replication and the 2030 audit. Answers email at length, interviews briefly, and footnotes everything.

其余八人的姓名与分工列于每篇论文的首页。本页不刊登人员照片。

The remaining eight are named, with their contributions, on the first page of every paper. This page carries no photographs of people.

我们不做的事What we do not do

• 不保留方法。We withhold no methods.
• 不认证任何设备。We certify no devices.
• 不主张损害赔偿。We claim no damages.
• 不接受保密协议。We enter into no non-disclosure agreements.

立场，一句The position, in one sentence

验证部负责人，2031 年，第三项研究发表之前head of division, 2031, before the third study was published

核查是一种尊重。你不会去核查一件你认为毫无价值的工作——你会不予理会，那只需要一个下午。我们已在该领域投入五年与两项研究。

Checking is a form of respect. You do not check work you consider worthless — you dismiss it, which costs an afternoon. We have now spent five years and two studies on this field.

联系Contact

来函请寄验证部信箱；地址、电子邮件与承办人姓名列于每篇论文的首页，与经费来源并列。我们确认收到一切。

Write to the division's post box. The address, the email, and the name of the person who will handle it are on the first page of every paper, beside the funding. We confirm receipt of everything.

本部门不设表单。

This division does not operate a form.

语料索引 / Corpus index

语料索引Corpus index

四千份样本语料库的构成、变更与撤回记录。语料库本身可自数据与代码取得，无需申请，无需登记，无需告知我们用途。

The composition, changes, and withdrawals of the 4,000-sample corpus. The corpus itself is available from Data & Code: no application, no registration, and no obligation to tell us what it is for.

发表时命名为“四千份样本语料库”，此名不因其后的减少而更改。差额即本页。

It was named the 4,000-sample corpus at publication, and the name does not change with a later subtraction. This page is the difference.

构成Composition

下列构成表为发表时（2026 年）的构成，合计四千。撤回发生于二〇二七年，见本页末。

The composition tables below are as at publication in 2026 and total four thousand. The withdrawals occurred in 2027; see the foot of this page.

年龄Age

年龄段Band	18–24	25–34	35–44	45–54	55–64	65+	合计Total
份数Samples	1,482	1,109	642	418	241	108	4,000

该分布明显偏向青年。这是招募方式的后果，不是设计。本部门在二〇二六年论文第七节将其列为第一项局限，并在此重述：任何以本语料库训练的系统，都继承这一偏向。

The distribution skews young. That is a consequence of how we recruited, not a design. This division listed it as the first limitation in section 7 of the 2026 paper and restates it here: any system trained on this corpus inherits the skew.

录制条件Recording conditions

条件Condition	份数Samples	说明Note
手机，单声道Telephone, single channel	2,614	七种机型，机型分布见附录 BSeven handset models; distribution in appendix B
廉价驻极体话筒Cheap electret microphone	1,386	同一型号，3.2 kHz 附近有约 2 dB 凸起One model, with a ~2 dB bump near 3.2 kHz
合计Total	4,000

该隆起是话筒型号的缺陷，不是世界的性质。它此后成为一项可用于判定衍生关系的印记，见2031 年审计第 3.3 节。本部门当初并未为此目的保留它；我们保留它，是因为它是真实发生的录制条件。

The bump is a defect of the microphone model, not a property of the world. It later became a mark by which a derivation could be established; see section 3.3 of the 2031 audit. We did not retain it for that purpose. We retained it because it is what the recording actually was.

撤回登记Withdrawal register

十一份。志愿者以不可逆编号标识，不具名，依匿名化规程第二条。撤回无需说明理由，登记表亦不设理由栏。

Eleven. Volunteers are identified by irreversible number and are not named, under clause 2 of the anonymization protocol. No reason is required for withdrawal, and the register has no column for one.

编号ID	撤回日期Withdrawn	已自语料库移除Removed	已通知下游Downstream notified
V-0288	2027-02-09	是yes	是yes
V-0417	2027-02-09	是yes	是yes
V-0631	2027-03-14	是yes	是yes
V-0902	2027-03-14	是yes	是yes
V-1155	2027-05-02	是yes	是yes
V-1478	2027-05-02	是yes	是yes
V-1802	2027-06-21	是yes	是yes
V-2240	2027-06-21	是yes	是yes
V-2799	2027-09-08	是yes	是yes
V-3311	2027-09-08	是yes	是yes
V-3746	2027-11-30	是yes	是yes

“已通知下游”指本部门已就该撤回，书面通知当时全部已知的下游使用者。二〇二七年共发出十三次通知，十三次均获答复并确认移除。二〇三一年，本部门查明另有一名此前不为我们所知的下游使用者，遂就同一事项补发第十四次通知，收件人为安调健康公司，未获答复。该封信不在署名的四封之内，单独计，见 2031 年审计第 3.2 节。

"Downstream notified" means this division wrote to every downstream user known to us at the time. Thirteen notifications were sent in 2027, and all thirteen were answered with removal confirmed. In 2031 we established the existence of a further downstream user not previously known to us, and sent a fourteenth notification on the same matter, addressed to Attune Health, Inc. It received no reply. That letter is not among the four attribution letters and is counted separately; see section 3.2 of the 2031 audit.

撤回不是一项手续。一个人有权改变主意，而改变主意应当到达它需要到达的地方。本部门无法保证这一点在所有下游都实现，只能保证已就每一次撤回去过信，并在此逐条列出结果。

A withdrawal is not a formality. A person is entitled to change their mind, and a change of mind should arrive where it needs to arrive. This division cannot guarantee that it does so everywhere downstream. We can guarantee that we wrote about every one, and we list the outcomes here, itemized.

匿名化规程 v1.2 / Anonymization protocol

匿名化规程Anonymization protocol

本部门发表全部内容。本页列出全部例外，以及每一项的依据。凡本部门未予发表者，均见于本页。

This division publishes everything. This page lists every exception and the grounds for each. Wherever this division does not publish, it is on this page.

条款Clauses

1. 个人不具名；机构具名。公司、社团、期刊、认证机构与政府部门在本部门的出版物中一律具名。个人不具名，除非该个人是本部门成员，或已就具名以书面表示同意。若某一机构的名称本身即为一位在世个人的姓名，而该个人未曾同意，则机构照常具名，其名称中属于该个人姓名的部分隐去。Individuals are not named; organizations are. Companies, associations, journals, certifying bodies, and government offices are named in our publications without exception. Individuals are not named unless they are members of this division or have consented in writing to be. Where an organization’s name is itself the name of a living individual who has not consented, the organization is named as usual and the part of the name belonging to that individual is withheld.
2. 志愿者以不可逆编号标识。编号与身份的对应关系不予保存。这意味着本部门无法应任何人的要求——包括志愿者本人、法院或本部门自身——重新建立该对应关系。代价是：撤回须由志愿者出示其编号。我们认为这一代价落在正确的一边。Volunteers are identified by irreversible number. The mapping from number to identity is not retained. This means this division cannot reconstruct it for anyone — the volunteer, a court, or ourselves. The cost is that a withdrawal requires the volunteer to present their number. We consider the cost to fall on the correct side.
3. 若一个条目的索引本身足以指认一个人，则索引隐去，条目保留。二〇二六年论文脚注七即依此条：该读数被完整记录，其索引未被印出。隐去索引不是隐瞒发现；发现是那个读数无法复现，而这一点已经印出。Where the index of an item would itself identify a person, the index is withheld and the item is kept. Footnote 7 of the 2026 paper is issued under this clause: the reading is recorded in full and its index is not printed. Withholding an index is not withholding a finding. The finding is that the reading could not be reproduced, and that is printed.
4. 撤回优先于完整性。一份被撤回的样本自语料库移除，不保留副本，不保留“仅供内部核对”的副本。若因此使某项此前发表的结果无法精确重现，则该情形在更正页说明，而样本不予恢复。Withdrawal takes precedence over completeness. A withdrawn sample is removed from the corpus. No copy is retained, including no copy "for internal verification only." Where this makes a previously published result impossible to reproduce exactly, the situation is described under Corrections and the sample is not restored.
5. 匿名来信原样保存，不作辨认。本部门不对匿名来件作笔迹、纸张、油墨、邮戳或书写习惯分析，不比对不同来件之间的一致性，亦不委托他人为之。一个选择不具名的人，已经说明了他的意思。Anonymous correspondence is kept as received and is not investigated. This division performs no analysis of handwriting, paper, ink, postmark, or style on anonymous items, does not compare items to one another for consistency, and commissions no one else to. A person who has chosen not to sign has stated their meaning.
6. 若一份文件的内容会指认其作者的一项持续做法，则其内容不予刊出，而其存在予以刊出。本条适用于二〇三〇年审计表四第八行所涉的那份文件。该文件真实存在，提交于二〇二二年三月，经邮寄，匿名，十一页，已登记，已搁置，现仍在收件机构存档。其标题与目次不予刊出。
   指认一件行为，即等于确定其归属；而本部门的发现恰恰是：该行为无法归属。刊出其内容，会使本部门自己的结论失效，并且对一个显然不愿被指认的人，做出对方已明确表示不愿发生的事。Where the contents of a document would identify an ongoing practice of its author, the contents are withheld and the existence is published. This clause applies to the document at row eight of Table 4 in the 2030 audit. The document is real: submitted March 2022, by post, anonymously, eleven pages, logged, tabled, and still held on file by the body that received it. Its title and table of contents are not printed.
   To identify an act is to attribute it, and this division's finding is precisely that the act cannot be attributed. Printing the contents would invalidate our own conclusion and would do to a person who has evidently chosen not to be identified the thing they have already indicated they do not want.
7. 本规程约束本部门，不约束他人。我们不要求任何人采用它，不以它为条件发放语料库，也不认为不采用它的人做错了什么。This protocol binds this division and no one else. We ask no one to adopt it, we do not condition the corpus on it, and we do not consider anyone who declines it to have done anything wrong.

本规程 v1.0 发布于 2026 年；v1.1 增列第五条（2029）；v1.2 增列第六条（2031）。历次版本均在此页保留，见附录。

v1.0 issued 2026; v1.1 added clause 5 (2029); v1.2 added clause 6 (2031). Every prior version is retained on this page; see the appendix.

工作底稿 / Working papers

工作底稿Working papers

表四（2030 年审计）给出件数。本页给出各行的细目，依匿名化规程处理。凡本页未刊出者，均注明所依条款。

Table 4, of the 2030 audit, gives counts. This page gives the particulars behind each row, handled under the anonymization protocol. Wherever something is not printed here, the clause under which it is withheld is named.

第一行 · 会签，87 件Row 1 · Countersignatures, 87

分布于十一个年度，平均每年七点九件。其中十一件为同一项：档案手套，每年一次，自 2019 至 2029 年不间断，规格相同，数量相同，会签人身份每年均无法确认。本部门为此项单列一节，理由是：一件连续十一年、每年一次、从不遗漏、从不重复申领的采购，在人员编制模型中是一个极强的约束条件。

Spread across eleven years, averaging 7.9 per year. Eleven of the eighty-seven are the same item: archival gloves, once a year, without a gap from 2019 to 2029, same specification, same quantity, countersigning officer unidentifiable every time. This division gave the item its own section, on the grounds that a purchase made once a year for eleven years, never missed and never duplicated, is a very strong constraint on any staffing model.

第三行 · 会籍与订阅，6 件Row 3 · Memberships and subscriptions, 6

六项会籍，分属六个不同的邻近社会团体：一家冰壶俱乐部、一个退伍军人会堂附属组织、一个图书馆之友协会、一个居民委员会、一个地方史学会，以及一个冬季咖啡合作社（2028 年加入，共十一名成员，自加入之年起无欠费）。六项均经邮寄办理。六份表格均只要求填写名，未要求填写姓；六份也都只填了名。

Six memberships in six different adjacent civic bodies: a curling club, a legion hall auxiliary, a library friends' association, a residents' committee, a local historical society, and a winter-coffee cooperative (joined 2028; eleven members; no unpaid balance since that year). All six entered by post. All six forms asked for a given name only, and all six received one.

六项会籍，六个不同的名字。全部不予刊出，依第一条。

Six memberships; six different given names. None printed, under clause 1.

本部门就这六个名作了检验，以确认其间是否存在结构：首字母、顺序、构词、与团体名称的对应。均无。这六个名在统计上不具任何特征——考虑到所审计的十年，本部门认为这一点本身值得记录。

This division tested the six names for structure: initials, ordering, morphology, correspondence with the bodies' names. There is none. The names are of no statistical interest, which, given the decade under audit, we found interesting enough to record.

第四行 · 委托票，4 件Row 4 · Proxy votes, 4

年Year	团体Body	动议Motion	投票方向Vote
2025	退伍军人会堂附属组织Legion hall auxiliary	社区团体周四场地费维持不变Hold the Thursday booking rate fixed for community groups	赞成for
2026	图书馆之友协会Library friends' association	开放时间Library hours	赞成for
2027	冰壶俱乐部Curling club	一条延长线An extension cord	赞成for
2028	居民委员会Residents' committee	一项分摊费用，用于“对一处停用喷泉的装饰性改善”An assessment funding "decorative improvements to a non-operational fountain"	反对against

四票均以邮寄投出，四票均计入表决结果，四票均无法归属。第四项的表决结果因该票而改变。

All four cast by post, all four counted in the recorded outcomes, all four unattributable. The fourth changed the outcome it was counted in.

第五行 · 书面提名，1 件Row 5 · Written nomination, 1

2029 年，向市政遗产委员会提名一处私人车库，书面提出。依该委员会惯例，申请人不予披露；本部门亦无从得知，并依第一条不作追查。提名结果不在本审计范围内。

In 2029, a written nomination of a private garage to the municipal heritage committee. Under that committee's practice the applicant is not disclosed; this division has no way of learning it and, under clause 1, does not pursue it. The outcome of the nomination is outside the scope of this audit.

第七行 · 致谢函，30 件Row 7 · Letters of thanks, 30

最早一封为 2027 年。三十封均无立案文件——即无任何可指明其触发事由的档案。三十封的收件人各不相同，分布于三个年度，发出顺序与该机构一份积压等候名单的顺序一致。该名单本身是公开的。本部门核对了顺序，一致，三十分之三十。

The earliest is 2027. All thirty lack a triggering file — that is, any record identifying what occasioned them. The thirty have thirty different recipients, fall across three years, and were issued in the order of a backlog waiting list published by the institution. That list is itself public. This division checked the ordering. It matches, thirty for thirty.

第八行 · 匿名提交并被搁置的政策文件，1 件Row 8 · Policy document submitted anonymously and tabled, 1

2022 年 3 月，某地区分类局经邮寄收到一份匿名政策文件，十一页。已登记，于下一次跨部门会议上搁置，此后未再提出。收讫确认已发往回邮地址；该回邮地址为一个邮政信箱，并于同年关闭。

In March 2022 a bureau of regional classification received by post an anonymous policy document of eleven pages. It was logged, tabled at the next cross-body session, and never raised again. Receipt was confirmed to the return address, which was a post-office box, and which closed the same year.

本部门在此记录一项方法上的困难，因为不记录它就等于假装它不存在：一个以“发表全部内容”为唯一规则的部门，在本行上没有发表全部内容。第六条是本部门为这一情形写下的规则，而规则是在遇到该情形之后写的。这一先后顺序应当由读者知晓。

This division records a methodological difficulty here, because not recording it would amount to pretending it does not exist: a division whose only rule is to publish everything has not published everything on this row. Clause 6 is the rule we wrote for the situation, and we wrote it after meeting the situation. The reader is entitled to know the order in which that happened.

伦理与同意 / Ethics and consent

伦理与同意Ethics and consent

本部门采集的是人体产生的声音。这属于个人信息。本页说明我们在采集时对志愿者说了什么，以及我们此后做到了什么。

This division collects sounds produced by human bodies. That is information about persons. This page states what we told volunteers at the time of collection and what we have done since.

同意书要点The consent form, in summary

1. 参与自愿，不设报酬，不参与不产生任何后果。Participation is voluntary and unpaid, and declining carries no consequence.
2. 样本将被公开发布，任何人可自由取得，包括商业使用者。这一点在采集前以口头与书面两次说明。Samples will be published openly and may be obtained by anyone, including commercial users. This was stated twice before collection, once verbally and once in writing.
3. 样本不附姓名、联系方式或任何可识别信息。志愿者取得一个编号，该编号与身份的对应关系不予保存。Samples carry no name, contact details, or identifying information. The volunteer receives a number, and the mapping from number to identity is not retained.
4. 同意可随时撤回，无需说明理由。撤回后样本自语料库移除，不保留副本。Consent is withdrawable at any time, without reason. On withdrawal the sample is removed from the corpus and no copy is retained.
5. 本部门无法收回已被他人下载的副本，并在采集前明确说明了这一点。这是本同意书中最重要的一句，也是唯一一句我们要求志愿者复述的。This division cannot recall copies already downloaded by others, and said so before collection. It is the most important sentence in the form and the only one we asked volunteers to repeat back.
6. 不接受未成年人的样本。年龄以本人声明为准，本部门不查验证件——查验证件需要留存证件信息，而留存证件信息与第三条相冲突。此项取舍在此说明，以便他人评价。No samples from minors are accepted. Age is by self-declaration; this division does not inspect documents, because inspecting documents requires retaining information from them, which conflicts with clause 3. The trade-off is stated here so that others may judge it.

此后我们做到的What we have done since

十一次撤回，十一次自语料库移除。二〇二七年向下游去信十三次，十三次获答复并确认移除；二〇三一年补发第十四次，收件人为安调健康公司，未获答复。逐条记录见语料索引。

Eleven withdrawals and eleven removals from the corpus. Thirteen letters to downstream users in 2027, all thirteen answered with removal confirmed; a fourteenth sent in 2031 to Attune Health, Inc., which received no reply. The itemized record is under Corpus index.

本部门未能做到的一件事，也在此说明：在二〇二六年公开发布之前，我们没有预见到有人会在未署名的情况下使用该语料库，因此在同意书中没有向志愿者说明这种可能。这不是一项法律缺陷，而是一项告知上的缺陷，并且是我们的。

One thing we did not do is also stated here: before the 2026 release we did not anticipate that anyone would use the corpus without attribution, and so we did not describe that possibility to volunteers in the consent form. That is not a legal defect. It is a defect of disclosure, and it is ours.

经费 / Funding

经费Funding

全部由本公司承担。无外部资助，无委托方，无附条件的捐赠。下表按研究列出直接支出，单位为人民币元。人员工资由本公司常规预算承担，不计入下表。验证部九人的年度工资总额为一百九十八万六千四百元（2031 年度）。

Borne entirely by this company. No external funding, no commissioning party, no conditional gift. The table gives direct expenditure by study, in yuan. Salaries are met from the company's ordinary budget and are not in the table. The annual salary cost of the division's nine members is 1,986,400 yuan (2031).

科目Line item	2026	2030–31	2031
录音设备Recording equipment	3,180	0	0
场地与志愿者交通Venue and volunteer travel	11,420	0	0
计算Compute	2,640	9,110	4,205
文件调取与备案查询Document and filing retrieval	0	6,875	3,940
设备购置（零售价）Device purchase, at retail	0	0	2,480
邮资（含挂号与回执）Postage, including registered post and receipts	318	742	596
数据托管，三年期Data hosting, three-year term	7,200	7,200	7,200
合计Total	24,758	23,927	18,421

邮资含挂号与回执费用。回执逐份编号并保存，清单见 2031 年审计附录 E。

Postage includes registered-post and return-receipt fees. Return receipts are numbered and retained; the manifest is appendix E of the 2031 audit.

本部门不接受的What this division does not accept

• 来自本部门可能审计之对象的资助。Funding from any party this division may audit.
• 附有发表前审阅条款的资助。Funding conditioned on pre-publication review.
• 咨询费、专家证人费用，以及任何形式的诉讼报酬。Consulting fees, expert-witness fees, and payment of any kind connected with litigation.
• 未经请求的款项。若收到，登记、入账、不动用、不退回——退回会开启一段通信，而通信会创设一个相对方。款项与登记记录一并保存。Unsolicited payments. If one arrives it is logged, booked, not spent, and not returned — returning it would open a correspondence, and a correspondence creates a party. The payment is kept with its log entry.

动态 / Notices

动态Notices

本页不删除条目。

Entries are not removed from this page.

日期Date	事项Entry
2031-11-04	匿名化规程更新至 v1.2，增列第六条。旧版保留。Anonymization protocol updated to v1.2 with the addition of clause 6. Prior versions retained.
2031-10-27	发表《对某消费级读取设备中一项未署名衍生关系的审计》。附全部代码与数据。Published On an Unattributed Derivation in a Consumer Reading Device, with full code and data.
2031-09-30	第四封署名请求函寄出，挂号。Fourth attribution request sent, registered post.
2031-07-15	撤回二〇二六年论文 6.4 节所作之假定。见更正。Withdrew the assumption made in section 6.4 of the 2026 paper. See Corrections.
2030-12-18	收到一份未署名的两段文件，更正表四之合计。本部门接受，并同时刊出两个数字。Received an unsigned document of two paragraphs correcting the total of Table 4. Accepted; both figures published.
2030-12-09	《某机构十年记录中的行政残差》预印本发布，四十一页。同日投稿，同日退回，未经评审。On an Administrative Residual released as a preprint, 41 pages. Submitted and returned the same day, unrefereed.
2030-03-11	行政复现工作开始。Work began on the administrative replication.
2031-03-20	查明一名此前不为本部门所知的下游使用者，补发第十四次撤回通知。未获答复。Established a downstream user not previously known to us and sent a fourteenth withdrawal notification. No reply.
2031-01-14	第一至第三封署名请求函寄出，挂号。Attribution requests one to three sent, registered post.
2029-11-05	匿名化规程更新至 v1.1，增列第五条。Anonymization protocol updated to v1.1 with the addition of clause 5.
2029-08-19	采用一家竞争性实验室的切分阈值。全部结果重新计算。Adopted a competing laboratory's segmentation threshold. All results recomputed.
2027-11-30	第十一次撤回登记完成。下游通知十三次，答复十三次。Eleventh withdrawal registered. Thirteen downstream notifications sent, thirteen answered.
2027-04-22	更正附录 B 中 dB 与 dBFS 混用一处，由一位来信人指出。两张图重新计算。Corrected the interchangeable use of dB and dBFS in appendix B, pointed out by a correspondent. Two figures recomputed.
2026-06-01	匿名化规程 v1.0 发布。Anonymization protocol v1.0 issued.
2026-05-14	《某消费级排气语言学分类引擎的复现》发布。同日投稿，同日退回，未经评审。Reproduction of a Consumer Flatulinguistic Classification Engine released. Submitted and returned the same day, unrefereed.
2024-01-08	验证部设立。The Verification Division was established.

引用 / Citation

如何引用本部门How to cite this division

本页供需要引用者取用。本部门更希望你核查我们，而不是引用我们；但引用是他人查到我们的方式，因此本页存在。

This page is here for those who need to cite. This division would rather you checked us than cited us; but citation is how others find their way to the work, so the page exists.

2026

2030 · 行政残差Administrative residual

2031 · 未署名衍生关系Unattributed derivation

引用语料库Citing the corpus

语料库的许可只要求署名。署名的形式不限：一行文字、一个脚注、一份材料清单中的一项，均可。不必通知我们，不必征得同意，不必付费。

The corpus license asks only for attribution. The form is not specified: a line of text, a footnote, an entry in a materials list — any of these. No notification, no permission, no payment.

若引用表四的合计，请注明为一百四十一，并注明本部门另刊出一百四十二。两个数字都属于这项工作。

If you cite the total of Table 4, cite it as 141, and note that this division also publishes 142. Both figures belong to the work.

人员 / People

人员People

验证部现有九人。每一篇论文的首页列出全部参与者及其分工，与经费来源并列。本页不刊登人员照片。

The Verification Division has nine members. The first page of every paper lists everyone who took part and what they did, beside the funding. This page carries no photographs of people.

姓名Name	职务Post	分工Contribution
林蔚然 · Lin Weiran	研究员，验证部负责人Senior Researcher, head of division	三项研究的第一作者Lead author, all three studies
陈砚 · Chen Yan	副研究员Associate Researcher	声学测量、频响、分段Acoustic measurement, frequency response, segmentation
周汝成 · Zhou Rucheng	副研究员Associate Researcher	统计、概率计算、图六Statistics, probability calculations, figure 6
罗有年 · Luo Younian	助理研究员Assistant Researcher	档案、公开备案文件、语料清单Archives, public filings, corpus manifest
苏敏之 · Su Minzhi	助理研究员Assistant Researcher	标注规程、裁定记录Labeling protocol, adjudication record
郑培 · Zheng Pei	工程师Engineer	数据发布、校验和、本站Data release, checksums, this site
高兰 · Gao Lan	研究助理Research Assistant	语料管理、撤回登记、下游通知Corpus management, withdrawal register, downstream notification
吴维明 · Wu Weiming	研究助理Research Assistant	复现测试、独立重跑Replication testing, independent re-runs
卫萍 · Wei Ping	行政专员Administrative Officer	收讫与通信Receipts and correspondence

最后一项为专职岗位，设立于二〇二四年，与本部门同时。

The last is a full-time post, established in 2024, at the same time as the division.

联系 / Contact

联系Contact

来函请寄验证部信箱。承办人姓名、通信地址与电子邮件列于每一篇论文的首页，与经费来源并列。本部门不设表单。

Write to the division's post box. The name of the person who will handle it, the postal address, and the email are on the first page of every paper, beside the funding. This division does not operate a form.

你可以期待的What you may expect

事项Item	本部门的做法Our practice
确认收到Confirmation of receipt	全部来函，无例外Every letter, without exception
答复A reply	全部来函，无例外Every letter, without exception
指出我们错了Telling us we are wrong	刊登于更正，署你的名，或依你的要求不署Published under Corrections, in your name or, if you prefer, not
要求撤回你的样本Withdrawing your sample	凭编号办理，无需理由，并向全部已知下游去信On presentation of your number, without reason, and we write to every known downstream user
匿名来信Writing anonymously	原样保存，不作辨认，依规程第五条Kept as received and not investigated, under clause 5 of the protocol

本部门设有一名专职人员负责收讫与通信。

This division employs one person whose post is receipts and correspondence.

我们确认收到一切。这一惯例并非本部门首创；我们在理解它之前，就已在受审计对象处见其可取，此后予以采用。

We confirm receipt of everything. The practice is not ours in origin: we admired it in a subject of audit before we understood it, and afterwards adopted it.