三个情报台，一套核查标准Three desks, one verification standard
每本登记册单独记账：条条有来源、有证据等级、有最后核查日期，全部公开。文字、图表与数据登记册采用 CC BY 4.0，引用注明出处即可。Every register keeps its own ledger: each entry carries a source, an evidence grade and a last-checked date, all public. Text, charts and registers are CC BY 4.0; cite the source and reuse freely.
证据等级Evidence grade
一手已核Primary-verified
多源已核Multi-source verified
媒体转述Media-reported
待核Pending
数据总览Data Overview
三个情报台，一张图看懂。这里是各登记册之上的视觉总览：数字全部取自各册本身，与登记册对不上时，以登记册为准。Three desks, one picture. This page is the visual overview above the registers: every number comes from the registers themselves, and where this page and a register disagree, the register wins.
一号台 · 企业 AI 架构情报Desk 1 · Enterprise AI Architecture
判例按主失效层归档，治理措施按「关」登记。另有准入册 3 条、模型变更 34 条、架构图谱 13 个、人物册 114 位，明细在各册。Cases are filed by primary failure layer, and governance measures by gate. Also on this desk: an access register of 3 entries, 34 model changes, 13 reference architectures and 114 people in the leaders register. Details live in each register.
判例落在哪一层Which layer the cases land in
门禁层占近半：多数事故不是模型不行，是没有一道关拦住。另有 1 条为受控实验，未计入分层。The Access Control layer holds nearly half: most incidents are not a model failing, they are the absence of a gate that would have stopped it. A further 1 entry is a controlled experiment and is not counted in the layers.
治理措施落在哪一关Which gate the governance measures land in
十关里五关已有措施入册；编排、知识、记忆、工具、模型五关尚无企业公开措施被登记，不等于市场上没有，是公开口径暂缺。Five of the ten gates have measures on record. Orchestration, Knowledge, Memory, Tool and Model have no publicly disclosed enterprise measure registered yet. That does not mean none exists in the market; it means none is on the public record yet.
判例披露月份（2026 年）Case disclosure months (2026)
按判例的公开披露月份统计。判例库 2026-10 上线并回溯收录，此前年份另有 6 条（2023 至 2025 年），未画入本图。Counted by each case's public disclosure month. The case register launched in October 2026 and collected earlier material retrospectively; a further 6 cases from earlier years (2023 to 2025) are not drawn here.
二号台 · AI 就业与职业情报Desk 2 · AI Jobs and Careers
脊柱是三句：暴露 ≠ 替代、使用 ≠ 流失、宣布 ≠ 落地。下图量的是任务暴露，不是失业预测。The spine is three lines: exposure is not replacement, use is not loss, and an announcement is not an outcome. The chart below measures task exposure, not unemployment forecasts.
职业观测暴露度（册内前列）Observed occupational exposure (top of the register)
共登记 11 行，其中可比职业 10 个，暴露度 ≥50% 的 8 个；OCC-011 为零暴露职业示例行，不计入比较。The register holds 11 rows, of which 10 are comparable occupations; 8 sit at 50% exposure or above. OCC-011 is an illustrative zero-exposure row and is not counted in the comparison.
就业研究册 · 6 篇Employment research register · 6 studies
每篇附方法与限度，混引纠正见《三个数字，三回事》。Each study is filed with its method and its limits. For the misquotation corrections, see Three Numbers, Three Different Things.
招聘信号时间线 · 3 例Hiring signals timeline · 3 cases
信号以「个案观察」呈现：宣布 ≠ 落地，单家公司不外推全经济。Signals are presented as case observations: an announcement is not an outcome, and one company is never extrapolated to the whole economy.
三号台 · 医疗 AI 情报Desk 3 · Medical AI
脊柱：获批 ≠ 上线 ≠ 付费 ≠ 有效。先看拿到什么许可，再看有没有付费路径。The spine: cleared or approved is not deployed, deployed is not paid, and paid is not effective. First look at what permission a device actually holds, then at whether a payment path exists.
FDA 对照构成 · 20 台FDA crosswalk composition · 20 devices
510(k) 是「清除（cleared）」，不是「批准（approved）」；PMA 才是批准。本批覆盖 8 个临床领域，放射科 12 台占六成。A 510(k) is a clearance, not an approval; only a PMA is an approval. This batch covers 8 clinical areas, and radiology accounts for 12 devices, six in ten.
付费对照 · 16 行 / 6 台器械Reimbursement crosswalk · 16 rows / 6 devices
按付费覆盖状态分：仅住院场景 7 行、未查明 8 行、逐案裁量 1 行。「未查明」本身就是结论的一部分：获批了，也可能没人付钱。Split by payment coverage status: inpatient settings only 7 rows, not yet established 8 rows, case-by-case discretion 1 row. Not yet established is itself part of the finding: a device can be cleared and still have nobody paying for it.
最近活动Recent activity
五关评审 EVA-001《Manus 过五关》研究页上线Five-Gate Review EVA-001, The Gauntlet: Manus, research page published 企业 AI 架构评估示范报告（方法示范）上线Demonstration enterprise AI architecture assessment report (method demonstration) published 判例库回溯第一批八条入库（INC-046 至 INC-053）The case library backfill batch 1 entered: INC-046 to INC-053 判例库新增 INC-045 恶意 AGENTS.md 劫持编程 Agent（知识层，一手已核，发现方自述）The case library adds INC-045, a malicious AGENTS.md hijacking a coding agent (Knowledge layer, primary-source verified, discoverer’s account) 架构图谱新增 ARC-013 Gemini Agent 治理栈（门禁与盯梢，一手已核）The architecture atlas adds ARC-013, the Gemini Agent governance stack (access control and observability, primary-source verified) 判例库新增 INC-042 至 044（含中文判例 Kimi 额度误扣），更正 INC-041，补记 INC-003；外部 Agent 准入册新增 AGT-003 嘉信理财Case library adds INC-042 to 044 (including the Kimi quota case), corrects INC-041 and updates INC-003; the External Agent Access Register adds AGT-003 (Charles Schwab) 判例库补齐至 42 条；监管动作册公开Case library completed at 42 cases; Regulatory Actions Register published 招聘信号册新增美国运通个案；《三个数字》纠正篇发布Hiring Signals adds an American Express case; the Three Numbers correction published数据来源：一号台判例与各册导出、二号台三册登记表、三号台器械与付费对照表；各源最后核查日见页首。本页为总览层，引用数字请回各册核对原文。本页数值由构建脚本自源数据生成，逐项对账表随包附上。Sources: the Desk 1 case and register exports, the Desk 2 register tables, and the Desk 3 device and reimbursement crosswalks. Each source was last checked on the date shown at the top of this page. This page is an overview layer: to cite a number, check it against the original register. Every figure here is generated from the source data by the build script, and an item-by-item reconciliation table ships with the package.