Oversight that cannot override无法否决的人工监督
A person is nominally in the loop but lacks the information, time, competence or stopping authority to change the outcome.名义上有人参与，实际上却没有信息、时间、能力或停止权限去改变结果
Independent assessment for consequential decisions面向重大决策的独立评估
Govern the joint system, not only the model 治理人机联合决策
Phoenix examines how AI actually enters a decision: who may challenge it, which evidence is verified, how conflicts are disclosed, and whether human oversight still works under pressure. 菲尼克斯研究AI如何真正进入决策：谁有权质疑，哪些证据经过验证，利益冲突如何披露，以及人工监督在压力之下是否依然有效
Evidence before assurance先有证据，再谈保障Every material finding stays traceable to its source.每一项重大结论都可追溯到证据来源
Compatible with the standards you already use兼容通行标准NIST AI RMF and ISO/IEC 42001 as the base, with a human-systems lens on top.以NIST AI RMF与ISO/IEC 42001为基础，叠加人机系统视角
Independent by design以制度保障独立性Conflicts, limitations and dissent are disclosed, never hidden.利益冲突、限制与分歧一律披露，从不隐藏
The overlooked risk被忽略的风险
A safe model can still produce an unsafe decision模型安全，决策未必安全
Failures rarely live inside the model alone. They emerge between the system and the organization: interfaces that invite over-reliance, reviewers without time or authority, contrary evidence that quietly disappears, and vendors who control both the evidence and its interpretation.失败很少只发生在模型内部，更多出现在模型与组织之间：界面诱导过度依赖，复核者缺少时间或否决权，反面证据悄然消失，供应商同时掌握证据与解释
Evidence without contestability无法质疑的证据
Decision-makers see supporting signals but cannot reconstruct the sources, assumptions or counter-evidence behind them.决策者看到的是支持性信号，却无法重建其来源、假设与反证
Governance captured by dependency被依赖关系俘获的治理
Whoever controls the data, tooling or budget quietly shapes the review, escalation and remedy process.掌握数据、工具或预算的一方，无形中主导了复核、升级与补救
Benefits centralized, burdens displaced收益集中，风险外移
Efficiency gains accrue centrally while errors, surveillance, skill loss and correction costs fall on others.效率收益集中获得，而错误、监控、技能退化和纠错成本由他人承担
What we do我们的服务
Four focused services for high-consequence systems四类专业服务
We assess one defined decision process, not an entire organization by assertion. Scope, evidence and limitations are agreed before the work begins.我们评估的是一条明确的决策流程，而不是凭印象评价整个组织，项目开始前先约定范围、证据与限制
01Flagship旗舰服务
Human–AI Decision Assurance Review人机联合决策保障评估
A four-layer review of whether one consequential AI-assisted decision remains challengeable, traceable and accountable under real operating pressure.围绕一条重大的AI辅助决策链开展四层评估，检验它在真实运营压力下是否仍然可质疑、可追溯、可问责
02
AI Governance & Risk AssessmentAI治理与风险评估
Independent review of system purpose, evidence, monitoring, human oversight, supplier dependency and governance controls, mapped to NIST and ISO.独立审查系统用途、证据、监测、人工监督、供应商依赖与治理控制，并映射到NIST与ISO框架
03
Executive Decision Risk高管决策风险工作坊
Closed-door board and executive exercises that surface automation reliance, motivated reasoning, evidence selection and escalation failures, without turning anyone into a psychological subject.通过闭门的董事会与高管演练，识别自动化依赖、动机性推理、证据筛选和升级失效，不对任何个人做心理诊断
04
Scenario Analysis & Systemic Risk情景分析与系统性风险
Decision-focused scenario trees, sensitivity analysis and simulation for supply chains, infrastructure, investment and policy, used only when the data and mechanisms justify the model.面向供应链、基础设施、投资与政策的情景树、敏感性分析与仿真，只有在数据与机制充分时才使用复杂模型
Assessment architecture评估架构
Four layers, one accountable decision chain四层评估架构
We combine documentary review, interviews, sampled reconstruction of past decisions and controlled behavioral testing. Each layer produces its own evidence and its own output, and the four are cross-checked before anything is graded.我们结合文件审查、访谈、历史决策抽样复建与受控行为测试，每一层都有自己的证据与输出，四层交叉核对之后才进入分级
Norms & obligations规范与义务
Applicable law, contracts, internal policy and standards are mapped to concrete controls.把适用法律、合同、内部政策与标准映射为具体控制
Output · obligation-to-control map输出 · 义务与控制映射Technology & operations技术与运营
Purpose, data, performance limits, monitoring and supplier dependence are tested against logs and records.依据日志与记录测试系统用途、数据、性能边界、监控与供应商依赖
Output · technical risk register输出 · 技术运营风险登记Human factors & decisions人因与决策
Automation reliance, trust calibration, interface effects, dissent and incentives are observed in controlled scenarios.在受控情景中观察自动化依赖、信任校准、界面影响、团队异议与激励
Output · behavioral risk and control effectiveness输出 · 行为风险与控制有效性ThreeCo & anti-capture三同与反俘获
Control rights, participation, appeal, conflicts of interest and distributional effects are challenged.检验控制权、参与、申诉、利益冲突与分配影响
Output · capture resilience and governance gaps输出 · 俘获韧性与治理缺口Every material statement carries an evidence label每项重大陈述都有证据标签
V
Verified fact已验证事实Supported by logs, documents, observation or multiple sources.由日志、文件、观察或多源证据支持
C
Client statement客户陈述Provided by the client or a vendor and not yet independently verified.由客户或供应商提供，尚未独立验证
I
Analytical inference分析推断Professional judgment resting on evidence that is listed.基于已列明证据的专业判断
S
Scenario assumption情景假设Used for stress testing; not a prediction of reality.用于压力测试，不代表对现实的预测
How a control is graded控制如何分级
0Not assessed未评估Out of scope or insufficient evidence; never read as a pass.不在范围或证据不足，不得推断为通过
1Missing缺失No control, or one that does not match the risk.没有控制，或控制与风险不匹配
2Designed已设计Documented, but not shown to operate.文档存在，但未证明实际运行
3Partly effective部分有效Runs in most scenarios with notable exceptions.多数情景运行，但有显著例外
4Effective有效Operates consistently across the agreed sample and period.在约定样本与期间内持续运行
Grades are given per control, never as a single organizational score. A document that exists is not a control that runs.分级针对每一项控制，而不是给组织打一个总分，文档存在不等于控制在运行
ThreeCo Principles
A practical lens for power, participation and distribution三同治理框架
ThreeCo does not replace law, NIST or ISO. It adds the questions those frameworks leave open: who controls the system, who can contest it, and who receives the benefits or bears the burdens.三同不取代法律、NIST或ISO，而是补上它们未必回答的问题：谁控制系统，谁可以质疑，谁获得收益，又由谁承担代价
同有Co-ownership
Rights, stakes and control权利、利益与控制
Clarify who controls data, models, interfaces, evidence and exit options, and who carries the material risk.厘清数据、模型、接口、证据与退出选择由谁控制，以及谁承担实质风险
Can affected stakeholders know, access, refuse, correct or exit?受影响者是否拥有知情、访问、拒绝、纠错或退出的权利
同治Co-governance
Contestability and remedy质疑、复核与补救
Build multi-party challenge, review, escalation, appeal and remedy into the operating system itself, not only into the policy.把多方质疑、复核、升级、申诉与补救写进实际运行机制，而不只停留在政策文件
Does human oversight have the competence, time and authority to say no?人工监督是否有能力、时间与权限对AI说不
同享Co-sharing
Benefits, burdens and capability收益、负担与能力
Examine how productivity gains, error costs, surveillance, skill change and remediation capacity are distributed across the people involved.审视生产率收益、错误成本、监控、技能变化与整改能力如何在相关群体之间分配
Who gains the efficiency, and who pays when the system is wrong?谁获得效率，系统出错时又由谁付出代价
NIST AI RMFISO/IEC 42001EU AI Act mapping欧盟AI法案映射Client internal policy客户内部政策Human factors science人因科学Anti-capture analysis反俘获分析
How an engagement works项目如何开展
From a bounded question to a testable remedy从明确问题到可复测的整改
Five stages, each with a defined client input and a defined output. Clients confirm facts at every stage; they do not control conclusions.五个阶段，每一阶段都有明确的客户投入与输出，客户在每个阶段确认事实，但不控制结论
Define the decision界定决策
System boundary, the role AI plays, harm scenarios, materiality and stakeholders.明确系统边界、AI角色、损害情景、重大性与利益相关者
Plan the evidence规划证据
The documents, logs, interviews, observations and samples each conclusion will require.为每项结论列出所需的文件、日志、访谈、观察与样本
Test behavior测试行为
Controlled scenarios of uncertainty, injected error, dissent, time pressure and escalation.在受控的不确定、错误注入、异议、时间压力与升级情景中观察反应
Challenge the findings挑战结论
Client fact-check, separated independent review, and documented disagreement.客户事实核对、隔离的独立复核，并记录分歧
Remediate and retest整改与复测
Named owners, deadlines, operating evidence and the residual risk management accepts.明确负责人、期限、运行证据，以及管理层接受的残余风险
A typical assurance review includes一次保障评估通常包含
- Up to 12 key-person interviews and two process observations最多12次关键人员访谈与两次流程观察
- Sampled reconstruction of up to 50 past decisions or incident records最多50个历史决策或事件记录的抽样复建
- One controlled behavioral stress test with up to 20 participants一项受控行为压力测试，最多20名参与者
- NIST, ISO and client-policy mapping plus ThreeCo anti-capture analysisNIST、ISO与客户政策映射，以及三同反俘获分析
- Executive briefing, full report, remediation workshop and one 90-day retest管理层简报、完整报告、整改工作坊与一次90天复测
It does not include不包含
- Source-code audit, penetration testing, full model red-teaming or privacy legal opinions源代码审计、渗透测试、完整模型红队或隐私法律意见
- Statutory certification, unqualified assurance, guaranteed regulatory approval or financial audit法定认证、无保留鉴证、监管批准保证或财务审计
- Psychological diagnosis, performance ranking or disciplinary recommendations for any employee对员工的心理诊断、绩效排名或纪律建议
Reading room客户导读
Six short reads on how the Center works读懂中心的六篇短文
Written for board members, general counsel, risk owners and investment committees who want to understand what an assessment will and will not tell them, before commissioning one.写给董事会成员、总法律顾问、风险负责人与投资委员会，在委托评估之前，先了解评估能告诉您什么，不能告诉您什么
Independence is a control system独立性是一套制度
Credibility comes from what we refuse to promise公信力来自拒绝承诺
We do not guarantee a favorable rating, regulatory approval or a commercial outcome. “Independent” describes governed behavior: conflict disclosure, evidence boundaries, separated review and correction, all written into contract and process.我们不保证低风险结论、监管批准或商业成功，独立指的是受制度约束的行为：利益冲突披露、证据边界、隔离复核与纠错，全部写进合同与流程
No outcome-contingent fees不按结论收费
Fees never depend on a low-risk result, an approval or a financing outcome.费用从不与低风险结论、审批或融资结果挂钩
Facts can be corrected, conclusions cannot be bought可纠正事实，不可购买结论
Clients may correct factual errors and attach a management response; they cannot remove adverse findings.客户可以纠正事实错误并附上管理层回应，但不能删除不利发现
Separated review隔离复核
Every material finding is reviewed by a qualified person who did not gather the primary evidence.每项重大结论都由未承担主要取证工作的合格人员复核
Limitations travel with the report限制随报告一同披露
Scope, untested controls, evidence quality, prior advisory work and unresolved dissent are stated in the document itself.范围、未测试的控制、证据质量、既往咨询关系与未解决的分歧都写在报告之中
External methods committee外部方法委员会
A part-time committee reviews method changes, major disputes and conflicts each quarter, paid by fixed fee and never by client outcome.兼职委员会每季度审阅方法变更、重大争议与利益冲突，按固定会议费支付，从不与客户结论挂钩
Respect for professional boundaries尊重专业边界
Legal opinions, statutory assurance, certification and specialized technical testing remain with qualified partners.法律意见、法定鉴证、认证与专业技术测试由有资质的合作机构承担
Ways to engage合作方式
Start with the smallest engagement that yields decision-grade evidence三种合作方式
Most clients begin with a two-week diagnostic of one workflow. Deeper work follows only when the diagnostic shows it is warranted.多数客户从一条流程的两周诊断开始，只有当诊断显示确有必要时，才进入更深入的评估
Rapid Diagnostic快速诊断
Determine whether a defined AI-assisted workflow warrants deeper review.判断一条明确的AI辅助流程是否值得开展深入评估
- Risk heatmap风险热图
- Evidence gaps证据缺口清单
- 90-day action list90天行动清单
Assurance Review保障评估
A four-layer review of one consequential human–AI decision process.针对一条重大的人机联合决策流程开展四层评估
- Behavioral stress test行为压力测试
- Standards and ThreeCo mapping标准与三同映射
- Prioritized remediation and 90-day retest优先级整改方案与90天复测
Monitoring & Retest持续监测与复测
Quarterly control checks, incident review and annual retesting for systems already assessed.面向已评估系统开展季度控制检查、事件复盘与年度复测
- Operating evidence of controls控制运行证据
- Remediation tracking整改跟踪
- Residual risk review残余风险复核
Bring us one decision that matters从一个重要决策开始
A useful first conversation identifies the decision, the role AI plays in it, the plausible harm, the evidence available, and who has the authority to act on what we find.第一次有效沟通应当明确：要评估什么决策，AI扮演什么角色，可能造成什么损害，手上有哪些证据，以及谁有权根据发现采取行动
Phoenix Institute · Seattle, United States菲尼克斯智库 · 美国西雅图
phoenix-institute.us ↗
Helpful to have ready:沟通前可先准备：
- The decision and what AI does in it要评估的决策，以及AI在其中做什么
- Who can override it today今天谁可以否决它
- What a wrong outcome would cost一次错误结果的代价