怎么估算一次上线的安全成本How to Estimate the Safety Cost of a Launch

列支出项，不列形容词：评测样本准备、人工复核轮次、灰度期双跑、返工窗口、回滚演练，每项写清由谁做、多久做一次。

基数取自己的历史数据：同类任务的平均处理时长、驳回率、工单量。取不到就明确写「假设值」，把假设连同理由放进表里，别只留一个数字。

区分一次性与经常性：读卡片、设计权限是一次性投入；抽检、监控、日志留存是经常性支出，要按月计。

给不确定性留位置：说明哪几项区间宽、为什么宽（新场景、缺历史数据、依赖外部审批），而不是统一加一个缓冲系数。

对比「不做的代价」要能追溯：误发处理时长、合同违约条款、监管整改要求，各自标注来源；找不到来源就留空，不编。

举个例子（虚构数据）：输入是「上线 AI 客服摘要，预计每天 300 条」的量级（虚构前提）→ 处理是四项列表，其中抽检按 10%（每天 30 条，每条假设 2 分钟，假设值）计 → 结果判断是代入假设得到每天约 1 小时人工，与团队可投入工时比较，判断是否需要降范围 → 若超出可投入工时，下一步是明确写出「降抽检比例带来的风险」或「缩小首批场景」，而不是把假设值当承诺写进方案。

List spending items, not adjectives: sample preparation, review rounds, parallel running during a staged rollout, rework windows, rollback drills - each with who does it and how often.

Take the base from your own history: handling time for comparable tasks, rejection rate, ticket volume. If the data is missing, write "assumed value" and keep the assumption plus its reason in the sheet instead of a bare number.

Separate one-off from recurring: reading cards and designing permissions are one-off; sampling, monitoring and log retention recur monthly.

Leave room for uncertainty: state which items have wide ranges and why (new scenario, no history, external approval dependency) instead of applying one blanket buffer factor.

Make the "cost of not doing it" traceable: incident handling time, contractual penalty terms, regulatory remediation requirements - each with a source. Leave the cell blank rather than inventing a figure.

Worked example (fictional data): input is a launch of AI-generated support summaries at "300 items a day" (fictional premise) -> processing lists the four items, with sampling at 10% (30 items a day, assumed 2 minutes each) -> the check substitutes the assumptions to get roughly one hour of human work per day and compares it against capacity to decide whether to narrow scope -> if capacity is exceeded, next step is to state plainly either the risk of lowering the sampling rate or the reduced first-batch scope, rather than writing an assumption into the plan as a commitment.