TeachOps：四个 Agent 跑完，审计仍然给了 FAIL

TeachOps 是个人参赛项目，实验对象是多个 Agent 的协作，不代表有四位人类队友。我负责场景收敛、角色边界和验收要求，AI 编程工具完成代码与配置。

我们把教学设计预审收窄到一节合成课例：Evidence 整理课标与学情，Design 修订教案，Audit 独立核查，Manager 负责调度和传递文件。导师角色保留最后的决定。

先有预期样例，再检查真实运行的差异

早期预置样例能展示想要的协作流程，却不能证明 Agent 真正完成了工作。材料因此区分设计、fixture 回放与 live 运行，输入中的学情也标明为合成数据。

8 月 16 日的 live 运行并没有照着预期样例给出一份全绿结果。Audit 指出修订稿漏掉了一个严重误解的显式处理，并将对应规则 R-003 判为 FAIL，课时结构 R-005 判为 WARN。

问题涉及把“分母越大，分数越大”当成规则的误解。修订稿有活动、有引用，仍缺少针对它的明确教学处理。若只检查文件齐不齐、引用编号在不在，就可能放过这个缺口。

人的批准需要保留条件

导师角色随后在房间内作出附条件批准：教学实施前必须补入处理该误解的环节。Manager 记录并传递这个要求。

这里不能把“附条件批准”改写成问题已经解决。记录能够证明有人根据审计结果作出了具体决定；它没有证明下一版教案完成修改，更没有证明真实课堂效果。导师也是演示中的真人审批角色，不能据此推断已经获得一线教师试用。

缺资料时，下游要真正停止

另一条异常样例故意缺少关键课标材料。Evidence 返回机器可读的 BLOCKED 与缺失清单，Manager 停止修订，不让 Design 用模型记忆补写依据。

这是角色权限的验收：调度者不替导师批准，审计者不自行修改教案，设计者不能越过证据门槛。把这些限制写进角色说明之后，还需要实际运行记录来核查。

本文核对固定版本审计报告和人工决定，未重跑 Agent。当前支持的是单一合成课例中的协作、审计与阻断；下一步应请真实导师用授权课例检查规则是否有用，并确认人工条件能否追到下一版。

TeachOps: four agents completed the run, and the audit still failed

TeachOps is my individual competition project. Its four agents are experimental software roles, not four human teammates. I set the scenario, responsibilities, and acceptance requirements; AI coding tools implemented the code and configuration.

We narrowed lesson-plan review to one synthetic case. Evidence organizes curriculum and learner information, Design revises the lesson, Audit checks it independently, and Manager coordinates files. A human mentor role retains the decision.

Compare expectations with an actual run

Fixtures showed the intended workflow but could not establish live collaboration. The records distinguish design, fixture replay, and live execution, with synthetic learner data labeled.

The August 16 live run did not produce a clean pass. Audit gave R-003 a FAIL because the revision omitted explicit treatment of a serious misconception; lesson timing received a WARN.

The misconception concerned assuming a larger denominator means a larger fraction. Activities and valid-looking citations did not ensure the revision addressed it. File completeness alone would miss that gap.

Preserve the condition attached to approval

The human mentor role gave conditional approval, requiring an additional teaching step before classroom use. Manager recorded and relayed it.

That decision does not establish that the next revision satisfied the condition. Nor does a real person acting as mentor in the demonstration establish adoption by practicing teachers or classroom effectiveness.

Verify that missing evidence stops downstream work

A separate case omitted critical curriculum material. Evidence produced a machine-readable BLOCKED state and missing-item list; Manager stopped revision rather than asking Design to fill gaps from model memory.

These checks examine authority boundaries: the coordinator cannot approve, the auditor cannot silently rewrite, and the designer cannot bypass the evidence requirement.

I reviewed the fixed-version audit report and human decision without rerunning agents. The evidence covers this synthetic case. A useful next step is an authorized lesson reviewed by a practicing mentor, including follow-up on approval conditions.

不是每件事都需要一个标签。

探索线索