S7-18 · 基础 · 概念实验 S7-18 · Foundation · Concept Lab

人在回路 · 不是出了错才临时找个人 Human-in-the-loop · Not a Person Summoned After Failure

risk score ≥ threshold → review queue → approve / correct / reject → continue

人在回路(Human-in-the-loop)是在高风险、低置信或例外节点由人审核、补充或批准后再继续。它应是状态机中的明确控制点,不是失败后的万能兜底。 Human-in-the-loop places human review, correction, or approval at high-risk, low-confidence, or exception nodes. It is an explicit state-machine control point, not a universal fallback after failure.

01 · 先从日常问题开始 01 · Start with an everyday question

十件订单里只有两件可能造成重大损失——你会让人逐件检查,还是给高风险件设一道明确关卡? Only two of ten orders could cause major loss. Review everything, or place an explicit gate on high-risk cases?

02 调整入队风险阈值,看四个结果一起变化 Move the Review Threshold and Watch Four Outcomes Change

拖动风险阈值 drag risk threshold
人工队列 Human queue
of 10 tasks
自动化率 Automation rate
bypassing review
残余高风险 Residual high risk
unreviewed tasks

03 人在回路由四个明确零件组成 Human-in-the-loop Has Four Explicit Parts

把刚才的变化拆开 Map the change to parts
01 · TRIGGER

触发条件 Trigger

风险、置信、例外类型或金额门槛必须可计算、可审计。 Risk, confidence, exception type, or value thresholds must be computable and auditable.

02 · QUEUE

人工队列 Human queue

任务进入有所有者、优先级和超时的真实工作队列。 Tasks enter a real work queue with owners, priority, and timeout.

03 · DECISION

人工决定 Human decision

审核者批准、补充、纠正或拒绝,并留下理由。 Reviewers approve, supplement, correct, or reject with a recorded reason.

04 · RESUME

恢复执行 Resume

只有明确结果写回状态,流程才能继续或终止。 The process continues or stops only after an explicit result returns to state.

04 最小规则:风险超过阈值才转人工 Minimal Rule: Route to Humans Above the Threshold

review if risk ≥ τ
human review ⇐ risk ≥ τ
阈值 τ 越低,队列越长、自动化率越低、等待时间越高,但未审核的高风险任务更少。 Lower τ creates a longer queue, less automation, and more wait time, but fewer high-risk tasks bypass review.
计算风险 Score risk 固定量尺与阈值 fixed rubric and threshold
转入人工状态 Enter human state 队列有责任人 queue has an owner
结果写回 Write decision back 批准、纠正或拒绝 approve, correct, or reject

05 边界实验:阈值拉得过高 Boundary Test: Set the Threshold Too High

主动制造失败 Create a failure

自动化率很好看,却放过高风险件 Automation looks great while high-risk work slips through

把阈值拉到 90。人工几乎不排队,但风险 70、80 的任务不再被审核。 Raise the threshold to 90. The human queue nearly disappears, but risk-70 and risk-80 tasks bypass review.

尚未执行。先在上方改变主变量。 Not run yet. Change the main variable above first.
出了问题再找人就是人在回路。 Calling someone after failure is human-in-the-loop. → ✓ 把触发、队列、决定和恢复写进流程。 Design trigger, queue, decision, and resume into the flow.
人工越多,系统一定越安全。 More human review always means more safety. → ✓ 同时测等待、自动化和残余风险。 Measure wait, automation, and residual risk together.
人工批准可以不留理由。 Human approval needs no reason record. → ✓ 决定必须写回状态并可审计。 The decision must return to state and remain auditable.
06 · 一句话带走 06 · One line to keep

人在回路是状态机中的明确控制点:按风险、低置信或例外入队,由人做有记录的决定后再继续;阈值改变会同时影响自动化、等待和残余风险。 Human-in-the-loop is an explicit state-machine control point: route risk, low confidence, or exceptions to recorded human decisions, then continue. Thresholds jointly change automation, wait, and residual risk.