我们
错了。

第六格是一个人。不是一个数字。
Cell six is a person. Not a number.

点击 ID 跳转到该错误的完整档案。 Click an ID to jump to the full file.

1. ERROR №.

   2046-03

   ATTR

   ATTRIBUTION BIAS · 归因偏差

   把「主动发起的群聊话题」误归为「社会压力」

   A chat topic the subject started themselves was misclassified as 'social pressure'

   STATUS

   FIXED

   已修复
   DISCOVERED / 发现于

   2046-04-12

   CORRECTED / 修正于

   2046-05-08

   SOURCE / 来源

   reader letter

   读者来信

   SCOPE / 范围

   Case №.087 and 43 back-scanned historical reports

   Case №.087 及回溯扫描的 43 份历史报告
   —01

   What happened

   发生了什么

   我们在 Case №.087(化名「阿河」,关于跳槽决定)的报告里,把他和前同事的一次群聊归为「社会压力」信号,并因此把外部影响占比从 17% 上调到 23%。阿河在读到报告后写信告诉我们:那次群聊是他主动起的话题,他在借话题试探自己想不想走。

   In Case №.087 (pseudonym 'A-He', on a job-change decision), we classified a group chat between him and a former colleague as a 'social pressure' signal, which raised external influence from 17% to 23%. After reading the report, A-He wrote to us: he had started that chat himself, using it to test his own inclination to leave.

   —02

   Where we were wrong

   我们错在哪

   我们默认「对话 = 外部影响」,没有区分「谁发起」和「谁回应」。

   We assumed 'conversation equals external influence'. We did not distinguish who initiated from who responded.

   —03

   What we did

   我们做了什么

   新增「发起者/被影响者」二分标注。重写 Case №.087 Page 08 相应条目,归因从 23% 社会压力改为 17% 社会压力 + 6% 个人试探。回溯扫描 2046 年 1–4 月发布的全部 218 份报告,识别出 43 份含有类似「主动发起对话」的条目,重新计算归因并给每位读者寄送更正函。

   Added an 'initiator / recipient' distinction to our conversation coding. Rewrote the relevant entry on Page 08 of Case №.087, shifting 23% 'social pressure' into 17% 'social pressure' + 6% 'self-probing'. Back-scanned all 218 reports published Jan–Apr 2046, identified 43 containing similar 'self-initiated conversation' entries, recalculated attribution, and mailed corrections to each reader.

   —04

   What we did not do, but should have

   我们没做但本来应该做的

   我们没有公开更正过程中的新归因模型的权重。这一项延到 2047 年第一季度发布。

   We did not publicly disclose the weights of the revised attribution model used in the correction. That disclosure is deferred to Q1 2047.

2. ERROR №.

   2046-07

   WORD

   MISLEADING PHRASING · 措辞误导

   一句注记被读者用作自我惩罚的依据

   A curator note was used by the reader as grounds for self-punishment

   STATUS

   FIXED

   已修复
   DISCOVERED / 发现于

   2046-06-19

   CORRECTED / 修正于

   2046-07-02

   SOURCE / 来源

   reader letter

   读者来信

   SCOPE / 范围

   Case №.142 · one reader

   Case №.142 · 一位读者
   —01

   What happened

   发生了什么

   策展人 A 在 Case №.142 的策展人笔记里写道:「你男朋友那句话是全报告中唯一值得你记得的信号。」读者在三个月后回访时告诉我们,她把这一句话读成了「其他任何信号都不值得记得」,并因此长时间无法走出与这位男朋友的关系——即便那段关系早已因为其他原因应该结束。

   Curator A wrote, in the notes for Case №.142: 'Your partner's line is the only signal in this entire report worth remembering.' Three months later, during a follow-up conversation, the reader told us she had read that line as meaning 'no other signal is worth remembering', and had for months been unable to leave the relationship — a relationship that, for reasons unrelated, should long since have ended.

   —02

   Where we were wrong

   我们错在哪

   「唯一值得记得」是一种判决,不是描述。我们把策展人的情感感受,误装成读者应该采用的结论。

   'The only one worth remembering' is a verdict, not a description. We packaged a curator's emotional impression as a conclusion the reader should adopt.

   —03

   What we did

   我们做了什么

   删除该条注记。重新发布 Case №.142 的修订版,新版本中同一段改为:「这条信号是我们模型没法处理的重量——它不属于 23%,也不属于 77%。」与读者当面致歉并全额退款。在所有策展人的内部写作规范中加入条款:禁止使用「唯一」「最」「真正的」等最高级限定词。

   Deleted the note. Re-issued a revised edition of Case №.142, in which the same passage now reads: 'This signal carries a weight our model could not handle — it belongs to neither the 23% nor the 77%.' Met the reader in person to apologise and issued a full refund. Added to all curators' internal writing guidelines: no superlatives — no 'only', 'most', 'truly'.

   —04

   What we did not do, but should have

   我们没做但本来应该做的

   我们没有召回已经寄出的 Case №.142 初版。这份文件仍然在这位读者手中。我们尊重她保留它的权利。

   We did not recall the original edition of Case №.142 already in the reader's possession. She continues to keep it. We respect her right to do so.

3. ERROR №.

   2046-09

   DATA

   INCOMPLETE DATA · 数据不完整

   我们遗漏了一个平台的算法推送权重

   We missed a platform's algorithmic push weight entirely

   STATUS

   FIXED

   已修复
   DISCOVERED / 发现于

   2046-08-03

   CORRECTED / 修正于

   2046-09-21

   SOURCE / 来源

   internal audit

   内部审计

   SCOPE / 范围

   19 reports between Feb–Jul 2046 involving one podcast app

   2046 年 2–7 月间涉及某播客 App 的 19 份报告
   —01

   What happened

   发生了什么

   在算法推送日志的聚合过程中,我们以为某主流播客 App 只提供订阅式分发,不使用 recommendation pipeline。我们因此没有把该 App 的曝光数据纳入权重计算。内部审计发现,该 App 2046 年初上线了基于听完率的推送模块。

   In aggregating algorithmic push logs, we assumed a certain major podcast app offered only subscription-based delivery and used no recommendation pipeline. We therefore excluded its exposure data from our weight calculations. An internal audit discovered that, in early 2046, the app had launched a listen-through-rate-based recommendation module.

   —02

   Where we were wrong

   我们错在哪

   我们假设一个 App 的分发机制长期不变。实际上,算法基础设施的变化速度远超我们的回看周期。

   We assumed a platform's delivery mechanism stays stable. In reality, algorithmic infrastructure changes faster than our retrospective window.

   —03

   What we did

   我们做了什么

   重新生成 19 份受影响的报告。在每份修订版顶部加印「已更正 · CORRECTED」红章,并附一封简短说明。今后每季度扫描主要内容平台的分发机制变更,并在年报中公布扫描结果。

   Regenerated the 19 affected reports. Every revised edition is stamped 'CORRECTED · 已更正' at the top, accompanied by a short note. From now on, we scan delivery mechanism changes at major content platforms every quarter, and publish the scan results in each annual report.

4. ERROR №.

   2046-11

   CURO

   CURATOR OVERREACH · 策展人过度推断

   把「替朋友搜索」误读为读者本人的思考

   Searching on behalf of a friend misread as the reader's own thinking

   STATUS

   FIXED

   已修复
   DISCOVERED / 发现于

   2046-09-28

   CORRECTED / 修正于

   2046-10-14

   SOURCE / 来源

   reader letter

   读者来信

   SCOPE / 范围

   Case №.209 · one reader

   Case №.209 · 一位读者
   —01

   What happened

   发生了什么

   策展人 B 在 Case №.209 的注记里,根据读者在三个深夜搜索「分手 原因」,推断她正在考虑结束自己的婚姻。读者回信说:「我在替我的朋友查,她那段时间很难。」

   Curator B wrote, in the notes for Case №.209, that three late-night searches for 'reasons for breakup' suggested the reader was considering ending her own marriage. The reader replied: 'I was searching for a friend who was going through a hard time.'

   —02

   Where we were wrong

   我们错在哪

   我们把「搜索行为」默认归属给搜索者本人。我们低估了「替他人查」在当代信息行为里的比例。

   We defaulted the ownership of a search to the person searching. We underestimated how large a share of contemporary information behaviour is done on someone else's behalf.

   —03

   What we did

   我们做了什么

   在 /demo 授权环节新增一项可选输入:「在过去 120 天里,是否有一段时间你在替他人查某件事? 大致是哪类主题?」结果会被用于将相关搜索从归因模型中剔除。在所有策展人培训中加入「替他人」情境分析模块。

   Added an optional prompt to the /demo authorization step: 'During the past 120 days, was there a period in which you were searching on behalf of someone else? What topics, roughly?' Results are used to exclude such searches from the attribution model. Added an 'on-behalf' module to all curator training.

5. ERROR №.

   2046-12

   OPS

   OPERATIONAL · 运营失败

   一位策展人因工作压力辞职

   A curator resigned due to workload pressure

   STATUS

   MITIGATED

   已缓解
   DISCOVERED / 发现于

   2046-10-05

   CORRECTED / 修正于

   —

   SOURCE / 来源

   curator self-report

   策展人自检

   SCOPE / 范围

   Curator Collective · one member

   策展委员会 · 一位成员
   —01

   What happened

   发生了什么

   策展人 C 在 2046 年 10 月 5 日提出辞职。在与她的离职谈话中,我们确认她在过去六个月承担了过多读者的直接对话工作,尤其是那些读报告后情绪剧烈波动的读者。她的离职信里写:「我每天要读 8 份报告。每一份都是一个人生里最重的决定。我没有办法持续。」

   Curator C tendered her resignation on 5 October 2046. In her exit conversation, we confirmed she had carried too much of the direct reader-dialogue workload over the previous six months, particularly with readers who reacted emotionally after reading their reports. Her resignation letter read: 'I read eight reports a day. Each is the heaviest decision in a life. I cannot sustain this.'

   —02

   Where we were wrong

   我们错在哪

   我们没有给读者回访设容量上限。我们以为「读者来信值得被回应」是一条绝对的原则,但我们把这份承诺的成本,压在了太少的几个人身上。

   We had set no capacity cap on reader follow-ups. We treated 'every reader letter deserves a response' as an absolute principle — while placing its cost on too few shoulders.

   —03

   What we did

   我们做了什么

   将策展人回访的周上限从「无限制」改为「每人每周 20 小时」。2046 年末开始招募 3 位新策展人。这仍然没有解决根本问题:我们在读者情感负担 和 策展人承受能力之间,还没有找到一个稳定的节奏。

   Changed curator follow-up hours from 'uncapped' to 'twenty hours per curator per week'. Began recruiting three additional curators in late 2046. This does not resolve the underlying problem: we have yet to find a stable cadence between reader emotional load and curator capacity.

   —04

   What we did not do, but should have

   我们没做但本来应该做的

   我们没有公开策展人 C 的名字。我们尊重她的隐私。但我们把她的离职信——经她同意——收录在本报告的附录,让她这句话可以被后来的策展人读到。

   We did not disclose Curator C's name. We respect her privacy. But — with her consent — her resignation letter is included in the appendix of this report, so that future curators can read what she wrote.

6. ERROR №.

   2046-13

   WORD

   MISLEADING PHRASING · 措辞误导

   「这个百分比够不够」是一句质问,不是一个问题

   'Is this percentage enough' is an accusation, not a question

   STATUS

   FIXED

   已修复
   DISCOVERED / 发现于

   2046-11-11

   CORRECTED / 修正于

   2046-12-01

   SOURCE / 来源

   external researcher

   外部举报

   SCOPE / 范围

   All 1,284 report covers published in 2046

   2046 全年所有 1,284 份报告的封面
   —01

   What happened

   发生了什么

   我们每份报告结尾都有同一个模板句——以每位读者自己的归因百分比代入:「这 X% 够不够?」。一位独立研究者在公开信中指出,这句话在语气上是一个修辞性质问,而不是一个开放问题。读者即使回答「够」,也会感觉自己在辩护。

   Every report we published ended with the same template line — with each reader's own attribution figure substituted in: 'Is this X% enough?'. An independent researcher wrote in an open letter: this line is rhetorically an accusation, not an open question. Even when a reader answers 'yes', she feels she is defending herself.

   —02

   Where we were wrong

   我们错在哪

   我们在最想表现克制的那一刻,其实最不克制。另外:我们用同一个百分比占位符盖过 1,284 份完全不同的报告,本身就是我们一直声称要避免的那种「一刀切」。

   At the precise moment we believed ourselves most restrained, we were least so. And: using the same percentage placeholder to close 1,284 entirely different reports is itself the kind of one-size-fits-all framing we claim to refuse.

   —03

   What we did

   我们做了什么

   从 2046 年 12 月起,报告结尾句改为双选模板:「这 X%——你需要它多一点,还是少一点?」(X 由每位读者那份报告的归因数代入)。读者可以选择,或者不选。这一版语气不强迫答案。

   From December 2046, the closing line became a two-way template: 'This X% — do you need more of it, or less?' (X substituted with each reader's own attribution figure). The reader may choose, or decline. The tone no longer forces an answer.

7. ERROR №.

   2046-14

   TECH

   TECHNICAL · 技术故障

   时间戳精度丢失,导致策展人注记引用错误时刻

   Timestamp precision loss caused curator notes to cite wrong moments

   STATUS

   FIXED

   已修复
   DISCOVERED / 发现于

   2046-12-02

   CORRECTED / 修正于

   2046-12-04

   SOURCE / 来源

   internal audit

   内部审计

   SCOPE / 范围

   7 reports had timestamps incorrectly rounded

   7 份报告的时间戳被错误四舍五入
   —01

   What happened

   发生了什么

   在我们的日志聚合管线中,一次库升级导致时间戳从秒级被意外四舍五入到分钟级。策展人在引用读者某条深夜搜索时,引用的时间比实际晚 30 秒 – 1 分钟不等。这在一般情况下无关紧要,但在其中一份报告里,策展人用「23:49 那次孤单的搜索」作为核心论点,实际时间是 23:48:32。

   A library upgrade in our log aggregation pipeline accidentally rounded timestamps from seconds to minutes. When curators cited a reader's late-night search, the time cited was 30 seconds to 1 minute later than the actual time. In most cases this is immaterial — but in one report a curator's central argument rested on 'that lonely search at 23:49', when the actual timestamp was 23:48:32.

   —02

   Where we were wrong

   我们错在哪

   我们在论证里依赖精度到秒的时间戳。我们的管线没有测试覆盖这种精度要求。

   Our arguments depended on second-level precision. Our pipeline lacked tests to enforce that precision.

   —03

   What we did

   我们做了什么

   修复管线。重新发布 7 份受影响的报告。在所有涉及「具体时刻」的注记中,要求策展人在引用前至少核查两次时间戳来源。

   Fixed the pipeline. Re-issued the 7 affected reports. For any curator note that cites a specific moment, curators are now required to double-verify the timestamp source before citing.

1. —01

   关于「那个归因百分比本身」

   On the attribution figure itself
   ADMISSION / 承认

   这是我们今年被问最多的:「报告里那个归因百分比真的是科学结果吗?」 不是。这是一个叙事装置。我们用加权平均计算归因,但权重本身是我们定的。那个数字是模型结果——但它是「模型认为」的数,不是真相。举一个例子:以今年被引用得最多的那份报告(Case №.林夏)为例,同一份原始数据,换一套权重,她的 23% 可能变成 19%、27%、甚至 31%。每一份报告都一样——数字在哪里落下,取决于我们用哪一把尺子量。

   This was the question readers asked us most often this year: 'Is the attribution figure really a scientific result?' It is not. It is a narrative device. We calculate attribution via weighted averages, but the weights themselves are chosen by us. The figure is what the model says — it is the model's number, not a truth. For example: take the most-cited report of the year (Case №.Lin Xia) — on the same underlying data, a different weight scheme could shift her 23% to 19%, or 27%, or 31%. Every report is like this: where the number lands depends on which ruler we chose to measure with.

   CORRECTION / 修正

   我们仍然会继续使用这个数字,因为它的功能是让人停下来。但从 2047 年起,每份报告会在 Page 04 的归因饼图下方加一行小字:「本数字由模型加权计算得出。权重详见附录 A。换一套权重会得到不同结果。」

   We will continue to use the number, because its function is to make a reader pause. But from 2047 onward, each report carries a small line below the Page 04 pie: 'This figure is produced by our weighted model. Weights are listed in Appendix A. A different weight scheme yields different results.'

2. —02

   关于「没搜过 = 没想过」

   On 'not searched = not considered'
   ADMISSION / 承认

   我们今年把「她没搜过」作为负空间证据的一大支柱。但一位读者在一份手写信里写:「我从没搜过『留在北京的理由』——不是因为我没想过,而是因为我在地铁上想。地铁没信号。」

   We leaned heavily on 'she never searched' as a pillar of negative-space evidence. But a reader wrote to us, by hand: 'I never searched ''reasons to stay in Beijing'' — not because I didn't think about it, but because I thought about it on the subway. The subway has no signal.'

   CORRECTION / 修正

   从 2047 年起,每份报告加入一个新页面:「Missing Channels / 我们拿不到的部分」。明确列出我们永远看不到的 5 类信号:离线思考、深夜失眠、与自己对话、没说出口的话、断网的通勤。我们在每一份报告里承认这些洞。

   From 2047, every report will include a new page: 'Missing Channels / What We Cannot Reach'. It will explicitly list the five classes of signal we can never see: offline thought, sleepless nights, inner monologue, words not said aloud, and offline commutes. Every report will name the holes in itself.

3. —03

   关于「我们把「闺蜜推荐」算在『社会压力』里」

   On counting 'a friend's recommendation' as 'social pressure'
   ADMISSION / 承认

   我们一直把亲密关系中的建议,和陌生人在平台上的转发,放在同一个「社会信号」桶里。这是一种明显的过简化。一位朋友的一句话,可能比 87 条陌生人点赞都更重;但在我们的模型里,它们被同等计数。

   We have long placed advice from intimate relationships and forwards from strangers on a platform into the same bucket: 'social signal'. This is a clear over-simplification. A line from a close friend may carry more weight than 87 strangers' likes — but in our model, they are counted equally.

   NO CORRECTION YET / 暂无修正方案

   我们承认这一条,但我们还没有想出怎么做。
   We admit this one. We have not yet figured out how to fix it.

1. UNFIXED-01 · WE DO NOT KNOW HOW TO FIX THIS

   未能修复 · 我们承认

   我们无法区分「你搜索它是因为你想过它」和「你搜索它是因为系统让你想到它」

   We cannot distinguish 'you searched it because you thought of it' from 'you searched it because the system made you think of it'

   这是一个因果识别问题。当一个人搜索「杭州小公司」,我们无法判断是她心里本来就有这个念头,还是算法在此之前已经把这个词重复 40 次让她认识到「这是个可以搜的词」。我们目前选择的做法是:把这类搜索归给「归因模型中的不确定项」,不做判断。我们承认这是一种逃避。我们没有能力解决这个问题,可能永远没有。

   This is a causal identification problem. When a person searches 'small companies in Hangzhou', we cannot tell whether the thought was already hers, or whether the algorithm had repeated that phrase 40 times until she recognised 'this is something one can search for'. Our current practice is to route such searches into 'indeterminate' and make no call. We acknowledge this is a form of evasion. We do not know how to solve this. We may never.

2. UNFIXED-02 · WE DO NOT KNOW HOW TO FIX THIS

   未能修复 · 我们承认

   读完报告后,读者情绪下坠——我们没有完备的接住方式

   When a reader falls emotionally after reading the report, we have no complete way to catch them

   2046 年有 126 位读者在读完报告的 72 小时内主动联系我们,表达了从「强烈的被看见」到「深度的自我怀疑」不等的情绪。我们目前的接续机制是:一次免费的策展人一对一回访 + 一份推荐的心理咨询师名单。这是不够的,但我们不打算假装更多。我们承认,这款产品会让一部分人更不好过。我们承诺继续把它做得更轻,但我们不承诺把它做得完全无害。

   In 2046, 126 readers contacted us within 72 hours of reading their report, expressing emotions ranging from 'powerfully seen' to 'deeply self-doubting'. Our current follow-up is one free one-on-one conversation with a curator, and a referral list of recommended therapists. It is not enough. But we do not intend to pretend otherwise. We acknowledge that, for some, this product will make things harder. We commit to making it lighter; we do not commit to making it harmless.