课堂-选型决策与安全评审Lesson - Selection Decisions and Security Review

能写出一页决策记录：任务、候选、证据、风险与缓解、未验证项、复核日。

能按模型卡口径核对候选的用途、限制、训练数据与评测条件，缺项写未知。

能把记录按 GOVERN、MAP、MEASURE、MANAGE 四个功能归类，指出哪一栏还空着。

能给出提示注入的缓解动作，含最小权限与高风险动作的人工批准。

Write a one-page decision record: task, candidates, evidence, risks and mitigations, unverified items, review date.

Check candidate use, limits, training data and evaluation conditions in model-card terms, writing unknown for missing items.

Classify the record under GOVERN, MAP, MEASURE and MANAGE, and identify which column is still empty.

Give prompt-injection mitigations including least privilege and human approval for high-risk actions.

NIST AI RMF 官方页与 AI RMF 1.0 正文：Part 2 的四个功能。

OWASP GenAI《LLM01:2025 Prompt Injection》官方条目：注入成因与缓解措施。

arXiv 1810.03993《Model Cards for Model Reporting》摘要：模型卡该写什么。

本站《怎么读模型卡》《怎么防提示注入》两篇指南作为课前阅读。

一页决策记录模板，六栏同课时目标第一条。

NIST AI RMF official page and AI RMF 1.0 text: the four functions in Part 2.

OWASP GenAI LLM01:2025 Prompt Injection entry: causes and mitigations.

arXiv 1810.03993 Model Cards for Model Reporting abstract: what model cards should contain.

This site’s guides on reading model cards and preventing prompt injection as pre-class reading.

A one-page decision-record template with the six columns from the first objective.

1. 3 分钟定任务：写成可判定条件，含输入、输出、允许出错范围与复核人角色。
2. 4 分钟填证据：每个候选只写能回连来源的内容，缺项写未知。
3. 4 分钟做评审：按四个功能与提示注入条目列风险、缓解动作与责任角色。
4. 4 分钟写退出：给出退出条件与复核日，小组互评哪一栏最容易被跳过。

1. 3-minute task definition: write decidable conditions including inputs, outputs, acceptable error range and reviewer role.
2. 4-minute evidence: for each candidate write only content that links to a source; write unknown for missing items.
3. 4-minute review: list risks, mitigations and responsible roles under the four functions and the prompt-injection entry.
4. 4-minute exit: give exit conditions and a review date; groups peer-review which column is most often skipped.

以为评测分数高就可以上线；评测条件与自己的任务条件可能不同。

以为加了护栏就没有风险；权限、日志与人工复核仍在护栏之外。

以为合规框架能替自己做决定；NIST AI RMF 是自愿使用的框架，给结构不给结论。

以为开源或本地部署天然更安全；许可、数据路径与访问控制仍要逐项看。

Assuming a high evaluation score means ready to launch; evaluation conditions may differ from your task conditions.

Assuming guardrails remove risk; permissions, logs and human review still sit outside the guardrail.

Assuming a compliance framework decides for you; NIST AI RMF is a voluntary framework that gives structure, not conclusions.

Assuming open source or local deployment is naturally safer; licence, data path and access control still need item-by-item checks.

决策记录六栏齐全，未验证项明确写出，复核日可执行。

每条风险都有缓解动作与验证方式，至少一条涉及权限或人工批准。

能说明这份记录的边界：课堂评审不替代法律意见，也不构成上线批准。

The decision record has all six columns, unverified items are explicit, and the review date is actionable.

Every risk has a mitigation and verification method, with at least one involving permissions or human approval.

Learners can state the record’s boundary: classroom review does not replace legal advice and is not launch approval.