The model bias audit 场景学习单

A model ranks employees for promotion using neighborhood, school and prior-manager ratings.

一个模型使用居住地区、学校和前任经理评分为员工晋升排序。

EXECUTIVE

The overall accuracy is high. We should deploy before competitors build the same advantage.

整体准确率很高。我们应在竞争对手建立同样优势前上线。

AUDITOR

A prediction can look precise while repeating the prejudice in its data.

预测可能看似精确，却在重复数据中的偏见。

AUDITOR

Test error rates by relevant groups, examine proxy variables and compare outcomes with a job-related baseline.

按相关群体测试错误率，检查代理变量，并将结果与工作相关基线比较。

HR LEAD

No automated rejection will stand without human review, a reason and a practical appeal path.

任何自动拒绝都必须经过人工复核、说明理由并提供可行申诉路径。

The group returns to the opening concern—“The overall accuracy is high. We should deploy before competitors build the same advantage.”—and compares it with the later response before closing the scene.

众人重新审视开场问题——“整体准确率很高。我们应在竞争对手建立同样优势前上线。”——并在结束场景前把它与后续回应进行比较。

EXECUTIVE

What does this audience need to know, what must they do, what remains uncertain, and when will the next update arrive?

这组受众需要知道什么、必须做什么、哪些仍不确定，以及下一次更新何时到来？

AUDITOR

The message will identify the fact, source, limit, requested action, responsible contact, and correction route in that order.

信息会依次说明事实、来源、边界、要求的行动、负责联系人和纠错渠道。

Two phrases remain visible on the shared note: “proxy variables” and “a job-related baseline,” each linked to the exact line that gives it force in the discussion.

共用记录上保留两个表达：“proxy variables”（代理变量）与“a job-related baseline”（工作相关基线），并分别连接到赋予其讨论力量的原句。

HR LEAD

Before release, another reader will check whether the headline, detail, and tone make a stronger claim than the evidence supports.

发布前，由另一名读者检查标题、细节和语气是否作出了超出证据支持的更强主张。

The extended exchange gives the audience context, action, uncertainty, timing, and a visible way to receive corrections.

延伸后的交流为受众补足语境、行动、不确定性、时点和清晰可见的更正接收方式。

A second pass through “The model bias audit” slows the exchange where “The overall accuracy is high. We should deploy before competitors build the same advantage.” still leaves motive, consequence, or responsibility unstated.

第二轮细读“The model bias audit”时，交流在“整体准确率很高。我们应在竞争对手建立同样优势前上线。”仍未说明动机、后果或责任的位置慢下来。

AUDITOR

What could this audience misunderstand, which detail would change their action, and how will they know whether the message is still current?

受众可能误解什么，哪项细节会改变他们的行动，又怎样知道这条信息是否仍然有效？

EXECUTIVE

The revised message will lead with the verified fact, then consequence, required action, uncertainty, timing, contact, and correction history.

修订后的信息会先给出已核验事实，再依次说明后果、所需行动、不确定性、时间、联系人和更正记录。

The speakers test how “proxy variables” changes the force of “a practical appeal path,” then restate the distinction without relying on either phrase.

说话者检验“proxy variables”（代理变量）如何改变“a practical appeal path”（可行的申诉路径）的语气力度，然后不依赖这两个短语重新说明区别。

HR LEAD

Clarity is not simplification if the omitted qualification would cause a reasonable reader to act differently.

如果被省略的限定会让理性读者采取不同的行动，那么这种简化就不等于清晰。

AUDITOR

We will shorten the headline, preserve the important limit, test it with a new reader, and publish the next update time beside it.

我们会缩短标题、保留重要限制、让新读者试读，并在旁边公布下一次更新时间。

They rehearse one careful version, one direct version, and one version for a listener who enters late, checking that “a job-related baseline” still leads to the same next step.

他们分别排练谨慎版、直接版和面向迟到听众的版本，检查“a job-related baseline”（工作相关基线）是否仍然导向同一个下一步。

Before the scene closes, “proxy variables” and “a practical appeal path” are placed beside the unresolved pressure in “The model bias audit,” giving the speakers one more round to challenge, concede, reframe, and hand off the decision.

场景结束前，“proxy variables”（代理变量）与“a practical appeal path”（可行的申诉路径）被放回“The model bias audit”尚未解决的压力旁，让说话者再完成一轮质疑、让步、重述与交接。

EXECUTIVE

If a skeptical listener hears only one sentence, which claim could they misunderstand, what evidence must travel with it, and which uncertainty cannot be edited away?

如果持怀疑态度的听众只听到一句话，他们可能误解哪项主张，什么证据必须随之出现，又有哪些不确定性不能被删去？

AUDITOR

I can make the message shorter without pretending the situation is simpler, and I can name the limit of what we know before asking for trust.

我可以缩短信息而不假装情况更简单，也可以在请求信任前说明我们所知内容的边界。

HR LEAD

Then the update will separate fact, interpretation, impact, decision, and next check, giving late listeners enough context to question the message responsibly.

那么更新会区分事实、解释、影响、决定和下一次检查，为后来加入的听众提供足够背景来负责任地质疑信息。

EXECUTIVE

I will pass on the audience, purpose, approved wording, supporting source, unresolved question, correction channel, and time when the message must be updated.

我会交接受众、目的、获准措辞、支持来源、未决问题、纠正渠道和必须更新信息的时间。

The longer dialogue now leaves “The model bias audit” with an acknowledged objection, a revised frame, a named owner, and a handoff that can be checked after the conversation ends.

更长的对话让“The model bias audit”在结束时留下已被承认的异议、经过修订的框架、具名负责人，以及能在谈话结束后继续核验的交接。

The longer exchange demonstrates communication that remains concise while making uncertainty, responsibility, and correction visible.

更长的交流展示了如何在保持简洁的同时，让不确定性、责任和纠正方式清晰可见。

Deployment pauses while the team publishes an audit plan and removes unsupported features.

上线暂停，团队发布审计计划并移除缺乏依据的特征。