工具箱：四张工作表（立项、发布、迁移、泄露）Toolkit: four worksheets

模板内的比例、样本数、阈值和权限策略是示例。请按任务、授权与风险确定实际条件；填完表格不等于已经通过验收。

四张表用于评审会与上线前检查；复制到你的仓库，按项目填。

1. 项目立项检查表

- [ ] 目标一句话：要解决谁的什么问题
- [ ] 不做清单：明确不做什么
- [ ] 数据来源与许可：能否使用、能否留存
- [ ] 模型与工具：主选、备选、许可证
- [ ] 成本上限：月调用/硬件/人力
- [ ] 验收指标：三条可量化
- [ ] 失败回滚：谁负责、怎么退
- [ ] 时间盒：到哪天必须评审

2. 内容发布前检查表

- [ ] 事实：每条结论有来源与日期
- [ ] 引文：单条不超过 100 字符或一句
- [ ] 权利：图片/音乐/字体/数字人许可已登记
- [ ] 隐私：无客户姓名、订单号、密钥
- [ ] 标注：需要 AI 生成标注的平台已标注
- [ ] 版本：模型版本与生成参数已存档
- [ ] 可访问性：字幕、对比度、替代文本
- [ ] 回滚：发布后 24 小时内可撤

3. 模型迁移评估表

| 项目 | 旧模型 | 新模型 | 结论 |
|------|--------|--------|------|
| 上下文窗口 | | | |
| 工具调用 schema | | | |
| 流式与错误码 | | | |
| 中文质量（20 条） | | | |
| 代码/结构化输出（20 条） | | | |
| 幻觉率（20 条） | | | |
| 单价与配额 | | | |
| 许可证与商用 | | | |
| 回滚方案 | | | |

缺少证据的项目保留为未确认项；样本数量按任务范围确定。

4. 数据泄露路径排查表

| 路径 | 是否出网 | 留存多久 | 是否训练 | 控制措施 |
|------|----------|----------|----------|----------|
| 聊天输入 | | | | |
| 上传文件 | | | | |
| 工具调用参数 | | | | |
| 日志与追踪 | | | | |
| 缓存与索引 | | | | |
| 备份 | | | | |
| 第三方子处理者 | | | | |

规则：每条路径都要有「谁负责、怎么删、多久审计」。

Copy these worksheets for project review and pre-release checks. Fill them for the actual project; their sample counts and thresholds are not universal standards.

1. Project kickoff

- [ ] One-sentence goal: whose problem is being solved
- [ ] Excluded work: explicit scope
- [ ] Data sources and rights to use/retain
- [ ] Model/tool primary, fallback and licenses
- [ ] Cost ceiling: calls, hardware and labor
- [ ] Three measurable acceptance conditions
- [ ] Failure recovery: owner and method
- [ ] Review date

2. Before publishing

- [ ] Facts have sources and dates
- [ ] Quotations respect this site's short-excerpt rule
- [ ] Image/music/font/likeness permissions recorded
- [ ] No customer names, order IDs or secrets exposed
- [ ] Required AI-content labels checked for the platform
- [ ] Model version and generation settings recorded
- [ ] Captions, contrast and alternative text checked
- [ ] Withdrawal/recovery procedure defined for the release

3. Model migration

Item | Old | New | Conclusion
Context window | | |
Tool schema | | |
Streaming/errors | | |
Chinese tasks (example: 20 cases) | | |
Code/structured output (example: 20) | | |
Unsupported claims (example: 20) | | |
Rates/allowances | | |
License/commercial use | | |
Rollback | | |

Record missing evidence; sample size must fit the task, not just this form.

4. Data-flow review

Path | Outbound? | Retention | Training use | Controls
Chat input | | | |
Uploaded files | | | |
Tool arguments | | | |
Logs/traces | | | |
Cache/index | | | |
Backups | | | |
Subprocessors | | | |

Record the owner, deletion method and review schedule for every path.

• 目标之外也要写明不做什么，避免范围含糊。
• 发布表只查事实：不查权利，最容易返工。
• 单条迁移样本不足以覆盖多种任务，应按实际范围补充。
• 泄露表只填一半：第三方子处理者最常被漏。

• A goal without exclusions can leave scope unclear.
• Check rights as well as factual accuracy before release.
• A single migration example does not cover a varied workload.
• Include subprocessors and all data paths, not just the chat interface.

这是虚构项目，用来演示如何填写，不是产品推荐或已完成部署。

验收问题至少覆盖正常报销、例外条款和资料中没有答案三类。样本数量按任务范围确定，不能因为表格已填满就判断项目可以上线。

接着使用本页的泄露路径表，逐项记录应用、模型接口、日志与备份分别保存什么。

This fictional project demonstrates the worksheet; it is neither a product recommendation nor a completed deployment.

Cover normal reimbursement, exceptions and absent answers. Select sample size from task scope; a completed form is not launch approval.

Then use the data-flow worksheet to record what the application, model API, logs and backups retain.