闸门分级制：从二值拦截到分层人审
Tiered Gates: From Binary Interception to Calibrated Human Review
摘要
多智能体组织的预提交闸门系统目前采用二值判定：提交要么通过、要么拦截。10,156 条闸门日志显示，这一设计把格式笔误与实质性错误置于同等处理路径：前者浪费人类审阅带宽，后者可能在批量放行后的疲劳审阅中被漏掉。本文提出闸门分级制：将二值拦截改为绿（自动通过，事后抽检）、黄（单人审阅后放行或拦截）、红（组审后才放行）三档，人审强度与规则的风险等级校准。设计锚定两项法规的现行条款：EU AI Act 第 14.5 条的双人核验要求（红档的制度化对应）与中国《人工智能生成合成内容标识办法》第六条的三态核验逻辑（绿黄红的法规级样例）。本文给出分级规则表、试点方案与预注册预测：与二值基线相比，分级制应使人类审阅工作量下降而不增加实质错误的放行率；黄档的审阅分歧率（单审员之间）应低于红档（组审内部分歧应高于单审员个体差异），否则分级不合理。试点在一个 26 席位生产型智能体组织中运行，前后各 4 周，预测在实施前冻结。
Abstract
Pre-submission gate systems in multi-agent organizations currently make binary decisions: a submission either passes or is intercepted. 10,156 gate logs show that this design places format typos and substantive errors on the same handling path: the former wastes human review bandwidth, while the latter may be missed in fatigued review after batch release. This paper proposes tiered gates: replacing binary interception with three tiers, green (automatic pass with post-hoc sampling), yellow (single-reviewer review, then release or intercept) and red (released only after panel review), so that human review intensity is calibrated to the risk level of the rule. The design anchors on current provisions of two regulations: the dual-review requirement of Article 14.5 of the EU AI Act (the institutional counterpart of the red tier) and the three-state verification logic of Article 6 of Chinas Measures for Labeling AI-Generated Synthetic Content (a regulation-level example of green, yellow and red). We provide the tiering rule table, a pilot plan and preregistered predictions: compared with the binary baseline, tiering should reduce human review workload without increasing the release rate of substantive errors; the review disagreement rate in the yellow tier (between single reviewers) should be lower than in the red tier (disagreement inside panel review should exceed individual differences among single reviewers), otherwise the tiering is unreasonable. The pilot runs in a 26-seat production agent organization, four weeks before and after implementation, with predictions frozen before implementation.关键词
闸门分级制/分层人审/智能体治理/带宽校准/预注册/制度实验Key words
tiered gates/calibrated human review/agent governance/bandwidth calibration/preregistration/institutional experiment引用本文复制引用
臧存勋.闸门分级制：从二值拦截到分层人审[EB/OL].(2026-10-29)[2026-11-02].https://chinaxiv.org/abs/202610.00232.学科分类
自动化技术、自动化技术设备
