The disparate-impact audit 场景学习单

Acceptance rates and later job performance appear in separate columns across demographic groups.

不同群体的录取率和后续工作表现显示在不同栏中。

DIRECTOR

The genetic threshold is the same for every applicant, so the system is neutral.

基因门槛对每名申请者相同，因此系统是中立的。

AUDITOR

A rule can look identical for everyone and still exclude one group without predicting the job.

一条规则可以对所有人看似相同，却在无法预测工作表现时排除一个群体。

AUDITOR

One group is rejected three times more often, while accepted workers from all groups perform at the same level.

一个群体的拒绝率高出三倍，但各群体获录用员工的表现处于同一水平。

DIRECTOR

Pause automatic rejection, remove non-predictive markers and rerun prior denials through independent human review.

暂停自动拒绝，删除无预测力标记，并对过去拒绝进行独立人工复核。

The group returns to the opening concern—“The genetic threshold is the same for every applicant, so the system is neutral.”—and compares it with the later response before closing the scene.

众人重新审视开场问题——“基因门槛对每名申请者相同，因此系统是中立的。”——并在结束场景前把它与后续回应进行比较。

DIRECTOR

Which statement is a direct observation, which is an inference, and which independent source could test the difference?

哪一项是直接观察，哪一项是推论，又有哪个独立来源可以检验两者的差别？

AUDITOR

We will record the claim, source, counterexample, and confidence level separately so a later summary cannot turn one into another.

我们会把主张、来源、反例和置信程度分别记录，避免后续摘要把它们混成同一种证据。

Two phrases remain visible on the shared note: “without predicting the job” and “rejected three times more often,” each linked to the exact line that gives it force in the discussion.

共用记录上保留两个表达：“without predicting the job”（却无法预测工作表现）与“rejected three times more often”（拒绝率高出三倍），并分别连接到赋予其讨论力量的原句。

DIRECTOR

Before acting, one reviewer will name the missing evidence, the threshold for changing course, and the time of the next check.

行动前，由一名复核者写明缺失证据、改变方向的门槛和下一次检查时间。

The extended exchange shows how an initial claim becomes a testable record instead of a premature conclusion.

延伸后的交流展示了最初主张如何变成可检验记录，而不是过早结论。

A second pass through “The disparate-impact audit” slows the exchange where “The genetic threshold is the same for every applicant, so the system is neutral.” still leaves motive, consequence, or responsibility unstated.

第二轮细读“The disparate-impact audit”时，交流在“基因门槛对每名申请者相同，因此系统是中立的。”仍未说明动机、后果或责任的位置慢下来。

AUDITOR

If the conclusion is wrong, which assumption failed first, and what observation would expose that failure before the decision becomes costly?

如果结论有误，最先失效的是哪个假设，又有什么观察能在决定代价变高前暴露这一点？

DIRECTOR

The record will distinguish what happened, who observed it, when it was recorded, what conflicts with it, and how confident each speaker is.

记录会区分发生了什么、由谁观察、何时记下、有哪些冲突信息，以及每位发言者有多大把握。

The speakers test how “without predicting the job” changes the force of “rerun prior denials,” then restate the distinction without relying on either phrase.

说话者检验“without predicting the job”（却无法预测工作表现）如何改变“rerun prior denials”（重新审查过去的拒绝）的语气力度，然后不依赖这两个短语重新说明区别。

DIRECTOR

A convincing story is still only a hypothesis until the missing source and strongest contradiction are examined.

在缺失来源和最强反证得到检查之前，再有说服力的叙述也仍然只是假设。

AUDITOR

Then we will write one provisional conclusion, one alternative explanation, and a specific test that can move confidence in either direction.

随后我们会写下一项暂定结论、一种替代解释，以及一个能让置信度向任一方向变化的具体检验。

They rehearse one careful version, one direct version, and one version for a listener who enters late, checking that “rejected three times more often” 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“rejected three times more often”（拒绝率高出三倍）是否仍然导向同一个下一步。

Before the scene closes, “without predicting the job” and “rerun prior denials” are placed beside the unresolved pressure in “The disparate-impact audit,” giving the speakers one more round to challenge, concede, reframe, and hand off the decision.

场景结束前，“without predicting the job”（却无法预测工作表现）与“rerun prior denials”（重新审查过去的拒绝）被放回“The disparate-impact 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。

DIRECTOR

Suppose the most persuasive claim is wrong. Which detail would fail first, who can inspect it without relying on us, and what decision can remain reversible meanwhile?

假设最有说服力的主张是错的，哪个细节会最先失效，谁能不依赖我们独立检查，又有什么决定可以暂时保持可逆？

AUDITOR

I can accept that my interpretation fits several facts, but I cannot present it as settled until the missing record and strongest contradiction have been reviewed.

我承认自己的解释符合若干事实，但在缺失记录和最强反证得到复核前，不能把它说成已经定论。

DIRECTOR

Then our conclusion will carry a confidence level, an alternative explanation, and a named test instead of borrowing certainty from the speaker’s confidence.

那么结论会附带置信程度、替代解释和具名检验，而不是借用说话者的自信制造确定感。

DIRECTOR

I will give the next reviewer the source list, disputed point, decision threshold, and deadline, so they can change our position without defending our pride.

我会把来源清单、争议点、决策门槛和期限交给下一位复核者，让对方可以修正立场而不必维护我们的面子。

The longer dialogue now leaves “The disparate-impact 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 disparate-impact 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。

By the end of the longer exchange, evidence is not decoration for a decision; it becomes the mechanism for revising it.

到这段更长的交流结束时，证据不再只是决定的装饰，而成为修订决定的机制。

The model remains available for study but loses authority to close an application.

模型仍可用于研究，但失去关闭申请的权限。