AI Explainability / Contestability / Adverse Action Playbook
以下官方来源是本文的治理锚点。本文把它们转成产品、流程、架构、评估和证据语言,不把任何法规或框架简化成单一检查项。
AI Explainability / Contestability / Adverse Action Playbook
定位:面向 CBAP+、AI Product Manager、AI Product Architect、Compliance Product Lead、Solutions Architect 的金融零售 AI 解释性、可争议性和不利行动通知设计手册。本文不是通用 XAI 算法笔记,而是把 explainability 设计成客户可理解、监管可追溯、模型验证可评估、工程可复现、申诉可处理、审计可证明的产品与控制体系。
重要说明:本文是学习、作品集和内部治理训练材料,不构成法律意见、合规结论、审计意见、模型验证报告或监管解释。正式项目必须由 Legal、Compliance、Fair Lending、Model Risk、Privacy、Security、Customer Complaint、Business Owner 和管理层结合机构类型、司法辖区、产品类型、客户群体、数据来源和内部政策确认。
Source Anchors
以下官方来源是本文的治理锚点。本文把它们转成产品、流程、架构、评估和证据语言,不把任何法规或框架简化成单一检查项。
| Anchor | Official link | 本文使用方式 |
|---|---|---|
| NIST AI Risk Management Framework | https://www.nist.gov/itl/ai-risk-management-framework | 用 Govern / Map / Measure / Manage 组织解释性、可争议性、监控、处置和持续改进;把 trustworthy AI 中的 transparency、explainability、fairness、validity、reliability、accountability 转成产品控制。 |
| EU AI Act | https://eur-lex.europa.eu/eli/reg/2024/1689/oj | 用于理解高风险 AI、信用评分/信用能力评估、human oversight、logging、technical documentation、transparency、right to explanation of individual decision-making 等要求如何转成 deployer / provider 证据。 |
| CFPB Circular 2022-03 | https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/ | 用于信贷复杂算法场景:黑盒或复杂性不能成为无法给出具体、准确 adverse action reasons 的理由;reason 必须反映实际主要原因。 |
| Regulation B adverse action notices | https://www.ecfr.gov/current/title-12/chapter-X/part-1002/section-1002.9 | 用于 adverse action notice 的时点、内容、specific reasons、ECOA notice、申请人请求理由的路径和书面确认要求。 |
| W3C PROV | https://www.w3.org/TR/prov-overview/ | 用于把 explanation evidence 设计成 provenance graph:entity、activity、agent、derivation、attribution、versioning、reproducibility,而不是散落截图和日志。 |
1. 核心心智模型:Explanation 是受治理的产品接口
Explainability 在金融零售 AI 中不是“让模型把思考过程说出来”,也不是把 SHAP chart 贴给客户。正确心智模型是:
AI explanation =
一个受治理的产品接口,
在特定客户、特定决策、特定政策、特定模型/规则/数据版本下,
向不同受众提供相称、准确、一致、可验证、可争议的解释,
同时保留足以复现、复核、申诉、纠错和审计的证据链。
这意味着 explanation 不是 raw chain-of-thought。客户、审核员、监管人员、模型验证团队和工程师需要看到的是经过治理的 evidence、reason、policy mapping、data lineage、decision boundary、model/system version、human review action 和 outcome,不是模型内部随口生成的一段看似合理的叙述。
1.1 五个常见误区
| 误区 | 为什么危险 | 更成熟的做法 |
|---|---|---|
| 把 explanation 当成模型输出 | 模型可能编造原因,导致 adverse action reason 与真实决策逻辑不一致 | reason code 和 evidence 必须来自 decision system 的 source of truth,LLM 只能在受控模板内改写或解释 |
| 把可解释性等同于透明全部细节 | 暴露欺诈规则、风控阈值、商业秘密或安全敏感逻辑会造成规避风险 | 按受众分层披露:客户可理解、审核可复核、验证可评估、工程可定位 |
| 把 SHAP / feature importance 当成客户解释 | 技术重要性不一定等于法律上的主要原因,也不一定客户能理解和采取行动 | 先定义 adverse action reason architecture,再选择可支撑 reason 的模型解释方法 |
| 把“人工可覆盖”当成 contestability | 客户没有明确入口、SLA、证据提交、状态追踪和结果通知时,争议权利无法落地 | contestability 是完整 case workflow,不是一个 escalation 按钮 |
| 把审计证据留到上线前整理 | 事后补材料无法证明当时版本、当时输入、当时政策和当时人审动作 | 在运行时自动生成 evidence package 和 provenance record |
1.2 Product Architect 的核心问题
设计 explainability / contestability / adverse action 时,先回答七个问题:
- 哪些系统或人员作出了客户影响性决定?
- 决定是否属于 adverse action、重要资格判断、费用处置、账户限制、欺诈处置或投资/保险适当性判断?
- 客户看到的解释、内部人审看到的证据、审计看到的证据是否来自同一决策事实?
- reason code 的 source of truth 是模型、规则引擎、策略表、人工决定,还是组合决策服务?
- LLM 是否被允许生成原因?如果不允许,它能做什么改写、摘要或问答?
- 客户如何纠错、补充证据、申诉、获得复核和收到结果?
- 六个月后监管问询某个 case 时,团队能否复现当时的输入、输出、版本、证据、通知和复核动作?
2. Explainability 分层架构
同一个 AI 决策不能只生成一种解释。金融零售系统至少需要五层 explanation artifact。
| 层级 | 主要受众 | 目标 | 典型内容 | 禁止事项 |
|---|---|---|---|---|
| L1 User-facing explanation | 客户、申请人、持卡人、保户、投资者 | 让客户理解发生了什么、主要原因是什么、可以做什么 | 结论、主要原因、使用数据类型、纠错/申诉路径、联系方式、时限、重要免责声明 | 不输出 raw chain-of-thought;不使用模型编造原因;不把内部评分、秘密规则、风控阈值直接暴露 |
| L2 Regulator / auditor evidence | 监管、内审、外审、合规、法务 | 证明解释准确、一致、及时、可复现、符合法规和政策 | notice version、reason code、evidence mapping、decision log、policy citation、approval record、complaint linkage | 不只给截图;不混用不同版本材料;不让人工口径覆盖系统事实 |
| L3 Model / system validation evidence | Model Risk、Fair Lending、Validation、Risk Analytics | 评估 explanation 是否 faithful、stable、non-discriminatory、fit-for-purpose | model card、system card、feature lineage、reason-code evaluation、segment analysis、challenger test、override pattern | 不把全局 feature importance 当成个案解释;不忽略拒绝样本和边界样本 |
| L4 Engineering debug trace | 工程、MLOps、Data、SRE、安全 | 复现失败、定位漂移、确认工具调用和数据版本 | request id、feature snapshot、rule execution、model version、prompt version、retrieval ids、API responses、latency、error codes | 不让 debug trace 直接进入客户文案;不记录超出保留政策的敏感数据 |
| L5 Internal reasoning traces | 受授权的 AI 平台、模型治理、事故调查团队 | 诊断 agent / LLM 行为和系统中间状态 | structured reasoning summary、tool-call plan、retrieved evidence、guardrail decisions、policy decisions、redaction events | 不把隐藏链路当作正式 reason;不把未经治理的自然语言推理暴露给客户或用作 adverse action reason |
2.1 五层之间的关系
Decision facts
-> reason-code source of truth
-> evidence graph
-> user-facing explanation
-> contestability case package
-> audit / validation / engineering views
关键原则:客户解释可以更简洁,但不能与内部证据矛盾;内部证据可以更详细,但不能事后重写客户收到的原因;工程 trace 可以帮助排错,但不能替代合规批准过的 reason architecture。
2.2 Explanation View Contract
每个客户影响性 AI 用例都应定义一个 explanation view contract。
| 字段 | 示例 | 设计要求 |
|---|---|---|
decision_id | CRD-APP-2026-00038192 | 全链路唯一,连接申请、模型调用、规则执行、通知、申诉和审计 |
decision_type | credit_application_denial | 使用受控枚举,不能自由文本 |
decision_outcome | denied | 与核心业务系统一致 |
reason_codes | HIGH_REVOLVING_UTILIZATION, INSUFFICIENT_VERIFIABLE_INCOME | 来自 approved reason catalog |
reason_rank | 1, 2 | 说明主要原因排序,不把低影响因素误列为主因 |
evidence_refs | feature snapshot、policy rule、bureau tradeline group | 指向证据对象,不在客户界面暴露全部原始数据 |
customer_text_template_id | AAN-CARD-DENIAL-V4 | 模板版本可追溯 |
policy_refs | ECOA_REG_B_NOTICE_POLICY_2026_Q2 | 引用内部政策版本和外部法规锚点 |
contestability_rights | dispute data、request reconsideration、upload documents | 说明客户下一步权利和路径 |
human_review_status | not_required, completed, pending | 区分全自动、人工复核和申诉复核 |
provenance_bundle_id | PROV-CRD-APP-... | 连接 W3C PROV 风格证据包 |
3. Adverse Action Reason Architecture
Adverse action reason 是一个受控架构,不是最后由 AI “总结一下为什么拒绝”。在信贷等受 Regulation B / ECOA 影响的场景中,核心不是解释听起来是否合理,而是 reason 是否 specific、principal、accurate、actually considered or scored、consistent、timely,并且能映射到可审计证据。
3.1 决策代码 Source of Truth
Reason code 的 source of truth 必须位于正式决策系统或 reason attribution service,而不是 LLM。
Application data
-> feature pipeline / rules input
-> policy rules + model score + constraints
-> decision orchestration
-> reason attribution service
-> approved reason codes
-> customer notice / internal evidence / audit package
3.1.1 Source of truth 设计原则
| 原则 | 说明 | 失败信号 |
|---|---|---|
| Deterministic | 同一 decision snapshot 在同一版本下应产生相同 reason code | 重跑后 reason 变化但无版本记录 |
| Versioned | reason catalog、mapping rule、model version、policy version 都有版本 | 客户 notice 无法还原当时规则 |
| Ranked | 能识别 principal reason,不是列出所有负面因素 | notice 包含无关或轻微因素 |
| Evidence-linked | 每个 reason code 都有证据对象和数据来源 | reason 无法连接到 feature、rule、bureau item 或人工记录 |
| Channel-consistent | 邮件、PDF、客服、chatbot、申诉页面口径一致 | 客服解释与正式 notice 冲突 |
| Human-reviewable | 人工复核员能看到 reason 生成逻辑和证据摘要 | 人审只能看到最终拒绝结论 |
| Governed | 新增、修改、停用 reason code 经过合规、业务、模型风险审批 | 产品团队临时加文案绕过审批 |
3.2 禁止模型自行发明原因
LLM 可以帮助把 approved reason 转成更易懂的 customer explanation,但不能发明 reason。尤其在 adverse action 中,必须明确禁止以下模式:
| 禁止模式 | 风险 | 替代设计 |
|---|---|---|
| “请解释为什么客户被拒绝”并让 LLM 自由回答 | LLM 可能根据申请内容编造看似合理但未被实际评分的原因 | LLM 只接收 approved reason codes、approved evidence summaries 和 approved template slots |
| 用模型概率最高的 token 生成 reason | 文案流畅但无法证明 principal reason | reason attribution service 先输出 structured reasons |
| 客服 AI 根据客户追问重新推测原因 | 造成正式 notice 与聊天解释不一致 | 客服 AI 只能读取 notice reason 和 case evidence,超出范围升级人工 |
| 用内部政策泛化成“未达到我行标准” | Regulation B 下通常不足以作为具体原因 | 提供具体、主要、实际考虑的因素 |
| 用技术指标代替客户可理解原因 | 客户无法理解或纠正 | 将 feature group 映射到 approved customer reason language |
3.2.1 LLM 在 adverse action 中可以做什么
| 允许能力 | 前提 | 示例 |
|---|---|---|
| Plain-language rewrite | 输入必须是 approved reason code 和 approved template | 把“revolving utilization high”解释为“您循环信用账户的已使用额度相对授信额度较高” |
| FAQ answer | 只回答 notice、流程、纠错、申诉、时限和联系方式 | “我可以提交哪些收入证明?” |
| Evidence summary for reviewer | 面向授权内部人员,引用 evidence refs,不替代正式 reason | “收入验证失败来自 payroll API timeout 和人工未能确认文件清晰度” |
| Translation / accessibility support | 保留法律批准语义,使用 approved glossary | 多语言、简明语言、无障碍格式 |
| Complaint triage | 识别客户争议意图并创建 case | “客户声称信用报告中的逾期记录不属于本人” |
3.3 Reason-Code Catalog
Reason-code catalog 是 adverse action reason architecture 的核心资产。
| 字段 | 说明 | 示例 |
|---|---|---|
reason_code | 稳定代码 | INSUFFICIENT_VERIFIABLE_INCOME |
reason_family | 分组 | Income / Capacity |
customer_reason_text | 客户可见短文本 | 可验证收入不足以支持本次申请额度 |
plain_language_detail | 客户解释扩展 | 我们无法根据您提供或授权取得的信息确认足够收入来支持所申请的信贷额度。 |
eligible_products | 适用产品 | credit card、personal loan、auto loan |
decision_context | 适用结果 | denial、counteroffer、line decrease |
principal_reason_rule | 如何判断为主要原因 | score contribution threshold、policy fail priority、knockout rule |
evidence_required | 证据要求 | income verification result、DTI calculation、document review result |
data_sources | 数据来源 | application form、payroll provider、bank transaction aggregation |
prohibited_if | 不得使用条件 | income data not used、verification unavailable due to system outage |
fair_lending_review | 公平信贷审查状态 | approved by Fair Lending 2026-Q2 |
template_ids | 可用模板 | AAN-PL-01, AAN-CC-03 |
owner | 业务/合规所有者 | Credit Policy Owner |
effective_from / effective_to | 生效区间 | 2026-04-01 / 2026-09-30 |
3.4 Reason Mapping to Evidence
每个 reason 必须能映射到 evidence,不是只映射到 feature name。
| Reason | Evidence objects | 客户可见解释 | 内部证据 |
|---|---|---|---|
| High revolving credit utilization | Credit bureau tradeline aggregate、feature snapshot、model reason rank | “您循环信用账户的已使用额度相对授信额度较高。” | bureau file version、tradeline aggregation logic、utilization value band、model attribution |
| Insufficient verifiable income | Application income field、verification API response、document review result | “我们无法确认足够的收入来支持本次申请。” | income source, verification status, reviewer notes, document hash |
| Limited credit history | Bureau file age、number of active tradelines、scorecard reason | “您的信用记录时间或可评估账户数量有限。” | bureau pull timestamp、tradeline count、minimum history rule |
| Recent delinquency | Bureau derogatory event group、policy rule hit | “您的信用报告显示近期存在逾期记录。” | bureau item refs、date bucket、rule id |
| Fraud identity mismatch | Identity verification result、KYC signal、case review | “我们无法确认申请信息与身份验证记录一致。” | identity provider response, risk flags, manual review action |
3.5 模板治理
Adverse action template 不能由渠道团队随意改写。模板治理至少包括:
| 模板元素 | 治理要求 |
|---|---|
| 标题与 action taken | 与业务结果一致,如 denied、counteroffer、line decrease、account closure |
| ECOA / applicable notice language | 由 Legal / Compliance 维护,不由模型生成 |
| Reason section | 只接受 approved reason codes 和 approved language |
| Data source clarification | 说明使用了哪些类型的信息,不暴露不必要敏感细节 |
| Consumer reporting agency information | 如适用,按 FCRA / 内部政策加入 |
| Correction and dispute path | 明确客户如何提交更正、申诉、补充文件或联系人工 |
| Accessibility and language | 简明语言、多渠道一致、可读屏、可打印 |
| Version and retention | 模板 ID、版本、生效期、审批记录、发送记录 |
3.6 Consistency Eval
Consistency eval 的目标是证明同一事实不会在不同渠道、不同模型版本、不同语言和不同客服回合中产生冲突 reason。
| Eval | 测试方法 | 通过标准 |
|---|---|---|
| Notice-to-system consistency | 抽样比对 sent notice reason 与 decision service reason | 100% 匹配 approved reason set;差异有审批解释 |
| Channel consistency | 同一 decision 在 PDF、email、web、chat、call center script 中比对 | 客户可见主原因一致,语气可不同但含义不变 |
| Language consistency | 英文、中文、西语等版本语义比对 | 法律核心含义和客户行动路径一致 |
| Re-run determinism | 用 archived snapshot 重跑 reason attribution | 在相同版本下 reason 一致;版本差异可解释 |
| Model upgrade regression | 新模型上线前比较 reason distribution 和 individual flips | reason drift 在阈值内;高影响 flip 已抽样复核 |
| Segment disparity in explanations | 按 protected-class proxy、地域、语言、渠道、年龄段等分析 reason 分布 | 异常差异进入 Fair Lending / Compliance review |
| LLM rewrite faithfulness | 检查 LLM 输出是否只表达 allowed reason and policy | unsupported reason rate 为 0 或低于经批准阈值并全量阻断 |
4. Contestability Architecture
Contestability 是客户对解释、数据、结论或处理结果提出争议并获得有效复核的能力。它不是情绪安抚,也不是客服 ticket。它是一个有权利入口、case 类型、证据提交、人工复核、数据纠正、重跑决策、结果通知和审计记录的产品架构。
4.1 Contestability 目标
| 目标 | 产品含义 |
|---|---|
| Discoverable | 客户能在 notice、app、web、call center、branch、chatbot 中找到申诉/纠错入口 |
| Understandable | 客户知道可以争议什么、需要提供什么、预计多久、结果如何通知 |
| Actionable | 客户可以上传文件、指出错误数据、请求复核、补充信息 |
| Human-reviewable | 合格人工复核员能看到原始决定、reason、证据、客户补充和政策 |
| Correctable | 数据错误能进入 data correction workflow,并影响后续重跑 |
| Re-runnable | 如果争议成立,系统能用 corrected data / updated evidence 重跑或重新裁决 |
| Auditable | 每一步都有 case id、时间戳、处理人、证据、决定和通知 |
4.2 End-to-End Workflow
Customer receives explanation / adverse action notice
-> Customer opens dispute / appeal / reconsideration
-> Intake classifies issue type and urgency
-> Identity and authorization check
-> Evidence package assembled
-> Customer submits correction or additional evidence
-> Human reviewer evaluates facts, policy and system decision
-> Data correction workflow if source data is wrong
-> Decision re-run or manual reconsideration
-> Outcome decision recorded
-> Customer outcome notification
-> Metrics, complaint linkage, audit retention
4.3 Contestability Case Types
| Case type | 客户表达 | 主要处理团队 | 系统动作 |
|---|---|---|---|
| Data dispute | “这个收入/地址/信用记录不对” | Data Ops、Credit Ops、FCRA/ECOA specialist | 验证数据源、提交纠错、暂停或标记相关流程 |
| Evidence supplement | “我有新的工资单/纳税记录/银行流水” | Underwriting Ops、Credit Policy | 接收文件、验证真实性、重算收入或能力 |
| Reason clarification | “我不理解为什么这个原因会影响申请” | Customer Care、Credit Specialist | 使用 approved explanation FAQ,不重新发明 reason |
| Decision appeal | “我认为拒绝决定不合理,请重新审核” | Human Review Team、Credit Officer | 组装 evidence pack,人工复核或重跑 |
| Fraud false positive | “这笔交易是我本人做的” | Fraud Ops、Security、Customer Care | 身份验证、解除限制、记录误报、更新模型反馈 |
| Fee dispute | “这笔透支费不应该收” | Deposit Ops、Fee Policy、Complaint | 核查交易顺序、披露、豁免政策、退款决定 |
| Complaint / regulatory complaint | “我要投诉” | Complaint Management、Compliance | 进入正式投诉流程,保留所有 AI interaction records |
4.4 Human Review Design
人工复核不是“看一眼 AI 结论”。复核界面应把证据结构化呈现:
| Review pane | 内容 |
|---|---|
| Decision summary | 原始 outcome、decision type、reason codes、notice sent timestamp |
| Customer claim | 客户争议点、提交材料、期望结果、渠道和语言 |
| Evidence graph | 数据源、模型/规则、特征快照、政策引用、人工动作 |
| Reason validation | 每个 reason 是否仍成立、证据是否充分、是否为 principal reason |
| Policy checklist | 对应产品政策、合规规则、例外规则、审批权限 |
| Fairness and vulnerability flags | 脆弱客户、语言障碍、无障碍需求、重复投诉、潜在公平性问题 |
| Action controls | uphold、overturn、partial correction、request more information、escalate、refund、re-run |
| Explanation composer | 只能使用 approved outcome templates 和 case-specific evidence slots |
| Audit capture | reviewer id、decision time、rationale summary、attachments、supervisor approval |
4.5 Correction and Re-run
当客户指出数据错误时,不能只在申诉系统里写备注。需要明确 corrected data 如何进入重新决策。
Customer correction
-> source validation
-> data owner approval
-> corrected data object / suppression flag
-> feature recomputation
-> reason attribution re-run
-> decision re-evaluation
-> new notice or outcome letter
-> audit linkage to original decision
4.5.1 Re-run Controls
| 控制 | 说明 |
|---|---|
| Snapshot preservation | 原始输入、原始 feature、原始模型版本必须保留,不能被 corrected data 覆盖 |
| Corrected-data lineage | 记录谁提交、谁验证、谁批准、影响哪些字段 |
| Re-run version policy | 明确用原模型版本、当前生产版本或人工重裁;选择必须可解释 |
| Outcome comparison | 比较原结果、新结果、reason 变化和客户影响 |
| Notice logic | 如结果改变,发送适当的 outcome notice;如不改变,解释争议处理结果 |
| Feedback governance | 误报或纠错数据进入模型训练前需经过 data governance 和 bias review |
4.6 Outcome Notification
申诉结果通知也需要治理,不能只写“经审核维持原决定”。
| Outcome | 客户通知应包含 |
|---|---|
| Upheld original decision | 复核范围、主要原因仍成立、客户后续路径 |
| Overturned decision | 更正内容、新决定、下一步操作、时间安排 |
| Partial correction | 哪些数据被修正、哪些原因仍影响决定、是否可继续补充 |
| More information needed | 需要材料、提交方式、截止时间、未提交后果 |
| Fraud false positive resolved | 账户/交易状态、是否退款、如何降低再次误报 |
| Fee refund granted / denied | 交易事实、政策依据、退款金额或拒绝退款原因 |
5. Financial Retail Cases
5.1 Credit Denial Explanation Assistant
场景
客户申请信用卡被拒绝,收到 adverse action notice 后在 mobile app 中询问:“为什么我被拒?我怎么改善?”
目标架构
Credit decision engine
-> reason attribution service
-> adverse action notice service
-> explanation assistant retrieval layer
-> approved FAQ / reason catalog / customer notice
-> dispute and reconsideration workflow
设计要点
| 设计点 | 正确做法 |
|---|---|
| Reason source | Assistant 只读取已经发送或即将发送的 approved reason codes |
| Explanation depth | 先显示正式原因,再用 plain language 解释该原因通常代表什么 |
| Improvement guidance | 可以提供一般性教育建议,例如“保持按时还款、降低循环信用使用率”,但避免承诺“做到 X 就一定批准” |
| Customer-specific evidence | 只能显示经过批准可披露的 evidence band,不暴露完整 bureau tradeline 或内部阈值 |
| Dispute path | 如果客户认为信用报告、收入或身份信息错误,直接创建 dispute / reconsideration case |
| Human handoff | 涉及歧视、投诉、脆弱客户、法律威胁、复杂错误时 warm handoff |
不允许
- Assistant 根据申请表内容猜测“可能因为收入太低”。
- Assistant 把模型分数、拒绝阈值或内部策略细节直接告诉客户。
- Assistant 输出与正式 adverse action notice 不同的原因。
- Assistant 说“AI 决定了您的申请,因此无法解释”。
5.2 Overdraft Fee Explanation
场景
客户被收取透支费,AI 客服解释交易顺序、余额变化、豁免政策和争议路径。
关键风险
| 风险 | 控制 |
|---|---|
| 交易顺序解释错误 | 使用 ledger API 和 posted transaction facts,不让 LLM 计算余额 |
| 政策引用过期 | 费用政策进入 versioned knowledge base,回答附 policy effective date |
| 客户投诉被 chatbot 吞掉 | 识别 complaint intent 并创建正式 complaint case |
| 不一致豁免 | fee waiver decision 由 policy engine 或人工权限控制 |
| 低金融素养客户无法理解 | 使用交易时间线、简单语言和可下载摘要 |
客户解释结构
1. 发生了什么:哪一笔交易导致可用余额低于 0
2. 费用依据:适用的账户条款和费用政策
3. 关键时间线:授权、入账、余额变化
4. 可争议路径:如果交易不是本人、余额显示有误或需要困难援助
5. 下一步:提交争议、请求豁免、联系人工
5.3 Fraud False Positive
场景
客户正常交易被 AI 欺诈模型拦截,账户被临时限制。
解释边界
欺诈解释必须避免泄露规避规则,但仍要告诉客户状态、原因类别和解决路径。
| 受众 | 可披露内容 |
|---|---|
| 客户 | “我们发现与您账户正常使用方式不一致的交易活动,需要确认是否本人操作。” |
| Fraud analyst | risk signals、device change、velocity group、geo anomaly、merchant risk category |
| Audit / compliance | model version、rule id、case review、customer notification、resolution time |
| Engineering | feature snapshot、API response、latency、error state、decision trace |
Contestability 设计
Customer confirms transaction
-> step-up authentication
-> analyst review if high risk
-> unblock / maintain restriction
-> customer notification
-> false-positive label governance
-> monitoring of repeat false positives by segment
关键 KRI:false positive appeal upheld rate、restriction duration、customer complaint linkage、segment disparity in fraud holds、repeat hold rate。
5.4 Insurance / Wealth Suitability Explanation
保险和财富场景不一定使用 adverse action notice 语言,但仍需要 explanation and contestability。
| 场景 | Explanation focus | Contestability focus |
|---|---|---|
| Insurance underwriting | 风险类别、资料缺失、核保条件、人工核保路径 | 补充医疗/资产/车辆/房屋资料,纠正第三方数据 |
| Wealth suitability | 风险承受能力、投资目标、期限、流动性、集中度 | 更正风险问卷、更新财务目标、人工顾问复核 |
| Robo-advice portfolio | 推荐组合理由、约束、费用、再平衡逻辑 | 客户调整约束、提出不适合理由、顾问介入 |
| Retirement guidance | 假设、情景、非保证性、数据来源 | 更改假设、纠正收入/资产数据、生成新情景 |
财富和保险解释的核心是避免把 AI 生成的自然语言伪装成个性化专业建议,除非产品、牌照、披露、适当性和记录保留均支持该建议边界。
6. Artifact Templates
6.1 Explanation Policy
| Section | 内容 |
|---|---|
| Purpose | 定义 AI explanation 的客户、监管、模型验证、工程和审计用途 |
| Scope | 覆盖产品、渠道、模型类型、自动化程度、客户影响范围 |
| Audience layers | User-facing、auditor、validation、debug、internal reasoning trace 的披露边界 |
| Source of truth | 决策事实、reason code、policy、notice、evidence 的权威系统 |
| Prohibited practices | 禁止 LLM 发明 adverse action reasons、禁止与 notice 冲突、禁止泄露敏感规则 |
| Template governance | 模板审批、版本、生效、停用、多语言、无障碍要求 |
| Contestability | 客户申诉、纠错、人工复核、重跑和结果通知标准 |
| Evidence retention | 日志、notice、case、模型版本、数据快照、审批记录保留策略 |
| Metrics and review | 正确性、一致性、unsupported reason、appeal upheld、complaint linkage |
| Ownership | Business、Compliance、Legal、Model Risk、Data Owner、Engineering RACI |
6.2 Reason-Code Mapping Template
| Field | Description |
|---|---|
reason_code | 受控代码 |
reason_family | Income、Credit history、Identity、Fraud、Affordability、Suitability |
decision_context | denial、counteroffer、fee decision、fraud hold、suitability block |
source_system | decision engine、rules engine、manual review tool |
source_logic_id | rule id、model attribution id、policy rule id |
principal_reason_selection | 排序算法、knockout priority、threshold rule、manual override rule |
evidence_objects | feature snapshot、document id、bureau item group、policy citation |
customer_language | approved plain-language text |
internal_language | reviewer evidence summary |
do_not_use_when | 不得使用条件 |
regulatory_notes | ECOA / Reg B / EU AI Act / internal policy considerations |
fairness_review_status | approved / restricted / retired |
last_reviewed | 日期、审批人、版本 |
6.3 Explanation Faithfulness Eval Template
| Eval dimension | Test question | Data sample | Pass signal |
|---|---|---|---|
| Reason fidelity | 输出 reason 是否全部来自 approved source? | denied / approved / counteroffer cases | 0 unsupported reason in customer-facing notices |
| Evidence support | 每个 reason 是否有 evidence refs? | high-impact decisions | 100% mapped or blocked from notice generation |
| Principal reason accuracy | reason 排序是否反映主要原因? | borderline / multi-factor cases | rank mismatch below approved threshold |
| Channel consistency | 不同渠道解释是否一致? | notice、chat、call script、web | no conflicting principal reasons |
| Segment fairness | 不同群体 reason 分布是否异常? | proxy groups、language、channel、region | anomalies reviewed with documented disposition |
| Stability | 小幅无关输入变化是否导致 reason 大变? | counterfactual tests | unreasonable flips investigated |
| Plain language | 客户是否能理解下一步? | usability testing、complaint text | comprehension score above target |
| Contestability linkage | explanation 是否引导可操作申诉? | user journey testing | customers can complete appeal without hidden path |
6.4 Contestability Workflow Template
| Step | Owner | System of record | Evidence captured | SLA / control |
|---|---|---|---|---|
| Intake | Customer Care / Digital | Case management | customer claim、channel、language、attachments | immediate case id |
| Triage | AI + human queue | Case management | case type、risk tier、complaint flag | high-risk routed same day |
| Evidence assembly | Decision service | Evidence store | notice、reason、feature snapshot、model version | automated package |
| Review | Authorized reviewer | Review workbench | reviewer actions、policy checklist、notes | maker-checker for high impact |
| Correction | Data owner | Data governance tool | corrected field、source validation、approval | source-level correction |
| Re-run / reconsideration | Decision owner | Decision engine / manual review | new outcome、reason comparison | version policy enforced |
| Outcome notice | Notice service | Communication archive | customer letter、timestamp、template version | approved template only |
| Close and monitor | Operations / Compliance | Case management | closure reason、complaint linkage、KRI update | QA sampling |
6.5 Adverse Action Evidence Pack
每个 adverse action case 应能生成一个 evidence pack。
| Evidence item | Purpose |
|---|---|
| Customer/application snapshot | 证明决策所用输入,不被后续修改覆盖 |
| Decision outcome record | 证明 action taken、时间、渠道、产品 |
| Reason attribution output | 证明 principal reason codes、rank、source |
| Feature snapshot and transformations | 证明 reason 与实际数据/特征对应 |
| Policy/rule execution log | 证明 knockout rule、policy fail、manual override |
| Model version and score record | 证明模型版本、score、calibration、reason attribution method |
| Notice generated and delivered | 证明客户收到的内容、时间、模板版本 |
| Human review record | 如适用,证明人审、覆盖、批准或拒绝 |
| Contestability record | 如客户申诉,证明 intake、review、correction、outcome |
| Complaint linkage | 证明是否进入投诉管理和监管响应 |
| Retention and access log | 证明谁访问、何时访问、为何访问 |
6.6 Audit Query Catalog
| Audit query | Evidence source |
|---|---|
| 某日期区间所有拒贷 notice 是否在规定时限内发送? | decision log、notice archive、delivery status |
| 某个 reason code 在各客户群体中的分布是否异常? | reason attribution table、fairness monitoring dataset |
| LLM 是否生成过不在 approved catalog 中的 reason? | LLM output validator logs、blocked output logs |
| 被申诉推翻的 adverse action 主要集中在哪些原因? | appeal outcome table、reason code table |
| 某次模型升级后 reason distribution 是否变化? | pre/post release monitoring、model version table |
| 客服聊天解释是否与正式 notice 不一致? | chat transcript classifier、notice reason comparison |
| 某个客户 case 能否复现原始决定? | snapshot store、model registry、feature store、rules version |
| 哪些 policy citations 已过期仍被输出? | knowledge base version、citation validator、policy registry |
| 哪些人工 reviewer 覆盖率异常高或异常低? | reviewer action logs、override reason table |
| 哪些渠道客户更难完成申诉? | journey analytics、drop-off、case completion metrics |
7. Metrics and KRIs
Explainability 和 contestability 的指标不能只看“用户点击率”。金融零售要同时看正确性、一致性、公平性、可操作性、运营效率和合规风险。
7.1 Core Metrics
| Metric / KRI | 定义 | 风险信号 |
|---|---|---|
| Explanation correctness | 客户解释与 source-of-truth reason 和 evidence 一致的比例 | 低于阈值说明 reason pipeline 或模板有缺陷 |
| Unsupported reason rate | 输出中出现未获批准或无证据支撑 reason 的比例 | 任何客户可见 adverse action unsupported reason 都应触发严重事件评估 |
| Notice-system mismatch rate | 正式 notice 与决策系统 reason 不一致比例 | 渠道、模板或异步数据同步风险 |
| Appeal upheld rate | 申诉后原决定被推翻或部分改正比例 | 过高说明数据质量、模型、规则或解释存在系统问题 |
| Segment disparity in explanations | 不同群体 reason 分布、解释长度、申诉成功率差异 | 可能触发公平性、可访问性或渠道偏差审查 |
| Stale policy citation rate | 输出引用过期政策、失效模板、旧 FAQ 的比例 | 知识库治理失败 |
| Human override rate | 人工覆盖 AI 决策或 reason 的比例 | 异常高说明 AI/规则质量问题;异常低可能是 automation bias |
| Complaint linkage rate | explanation / adverse action case 进入投诉的比例 | 高发 reason 或渠道需要根因分析 |
| Time to explanation | 决策后生成并发送解释或 notice 的时间 | 超时可能违反内部政策或法规时限 |
| Contestability completion rate | 客户成功完成申诉/纠错流程的比例 | 低完成率可能说明入口隐藏、材料要求不清或流程过复杂 |
| Re-run reproducibility rate | 使用 archived snapshot 成功复现原决定的比例 | 低于阈值说明证据链不完整 |
| Reason drift after release | 新模型/规则上线后 reason 分布变化 | 未解释漂移进入 release hold 或 post-release review |
7.2 Operational Dashboard
| View | Questions answered |
|---|---|
| Executive risk view | 哪些产品、渠道、reason family 风险最高?是否接近风险偏好阈值? |
| Compliance view | notices 是否及时、准确、一致?是否有过期政策引用? |
| Model Risk view | explanation faithfulness、reason drift、fairness segment 是否异常? |
| Product view | 客户是否理解解释?申诉转化在哪一步流失?哪些原因驱动投诉? |
| Operations view | 申诉队列、SLA、人工复核产能、推翻原因、重复 case |
| Engineering view | evidence pack 生成失败、trace 缺失、validator block、latency、API error |
7.3 Threshold and Escalation
| Trigger | Escalation |
|---|---|
| Customer-facing unsupported adverse action reason | 立即阻断生成、升级 Compliance / Legal / Incident,抽样影响客户 |
| Notice-system mismatch | 暂停相关模板或渠道,修复映射,重新核对已发送通知 |
| Appeal upheld spike | Root cause review:数据源、模型版本、规则变更、渠道误导、客服话术 |
| Segment disparity spike | Fair Lending / Compliance review,必要时停止相关 release |
| Stale policy citation | 禁用知识库版本,回滚到 approved version,通知渠道 owner |
| Reproducibility failure | Evidence architecture defect,进入高优先级整改 |
8. Provenance and Audit Evidence Architecture
W3C PROV 的价值在于把“谁用什么数据,通过什么活动,生成了什么结果”结构化。金融零售 explanation evidence 可以借鉴 entity / activity / agent 的思路。
8.1 PROV 风格对象
| PROV concept | 金融零售 AI 对象 |
|---|---|
| Entity | application snapshot、feature vector、model artifact、reason code、notice PDF、customer upload |
| Activity | bureau pull、feature computation、model scoring、rule execution、notice generation、human review、appeal re-run |
| Agent | customer、decision service、model service、reviewer、data provider、notice service |
| Derivation | reason derived from feature snapshot and policy rule |
| Attribution | notice generated by notice service under template owner approval |
| Association | human reviewer associated with appeal review activity |
| Versioning | model version、template version、policy version、reason catalog version |
8.2 Evidence Graph Example
Customer application snapshot A
wasGeneratedBy intake activity I
wasUsedBy feature computation F
Feature vector V
wasGeneratedBy F
used bureau file B and stated income S
Decision D
wasGeneratedBy decision activity X
used V, model M v3.8, ruleset R 2026-Q2
Reason code RC1
wasDerivedFrom D and feature group FG17
wasMappedBy reason catalog C v12
Notice N
wasGeneratedBy notice service NS
used template T v4 and reason codes RC1, RC2
Appeal case AC
wasStartedBy customer CUST
used Notice N and customer upload U
wasReviewedBy reviewer HR1
8.3 Evidence Store Requirements
| Requirement | Why it matters |
|---|---|
| Immutable decision snapshot | 防止后续数据更新导致无法复现原始决定 |
| Versioned artifacts | 模型、规则、模板、政策、reason catalog 都必须可还原 |
| Access controls | 客户数据、欺诈信号、内部政策和模型细节需要分级访问 |
| Retention policy | 与产品、法规、投诉、模型风险和诉讼保全要求一致 |
| Tamper-evident logs | 支撑审计可信度和事故调查 |
| Evidence completeness check | notice 发送前确认必要证据对象存在 |
| Queryable metadata | 审计、监管、模型验证能按日期、产品、reason、segment、版本查询 |
9. Requirements for BA / PM / Architect
9.1 Business Requirements
| Requirement | Acceptance criteria |
|---|---|
| 客户能看到主要原因 | 每个客户影响性 adverse action 都生成 approved reason codes 和 customer explanation |
| 客户能采取行动 | 每个 explanation 页面都提供纠错、申诉、补充材料或人工联系路径 |
| 客服口径一致 | 客服 AI、人工客服脚本和正式 notice 使用同一 reason source |
| 申诉能闭环 | 每个 appeal 有 case id、SLA、review outcome 和客户通知 |
| 合规能抽查 | Compliance 能按产品、日期、reason、渠道抽取证据包 |
9.2 Functional Requirements
| Capability | Requirement |
|---|---|
| Reason attribution service | 接收 decision snapshot,输出 ranked approved reason codes、evidence refs、confidence / applicability metadata |
| Explanation composer | 使用 approved templates,把 reason codes 转成多渠道、多语言、无障碍 customer explanation |
| LLM guardrail | 阻断 unsupported reason、policy mismatch、unapproved legal language、sensitive rule leakage |
| Contestability case management | 支持 intake、triage、upload、review、correction、re-run、outcome notification |
| Evidence pack generator | 自动组装 decision、reason、model、rule、notice、human review、appeal evidence |
| Audit query interface | 支持监管和内审常见查询,不要求工程临时拉取散表 |
9.3 Non-Functional Requirements
| NFR | Requirement |
|---|---|
| Reliability | notice generation 和 evidence capture 失败时不得静默失败;高影响场景 fail closed |
| Latency | 客户解释可以异步生成,但 adverse action notice 时限必须满足法规和政策 |
| Security | fraud signals、model thresholds、internal policy rules 分级披露 |
| Privacy | 客户解释不暴露第三方无关数据;debug trace 遵守数据最小化和保留要求 |
| Accessibility | 解释和申诉流程符合无障碍和语言可理解要求 |
| Observability | 每个 decision_id 可追踪生成、发送、查看、申诉和关闭 |
| Portability | 模型替换、供应商替换、渠道新增不破坏 reason source of truth |
10. Implementation Blueprint
10.1 Reference Components
| Component | Responsibility |
|---|---|
| Decision Orchestrator | 汇总规则、模型、人审和业务约束,生成 action taken |
| Reason Attribution Service | 从决策事实生成 ranked approved reason codes 和 evidence refs |
| Reason Catalog | 管理 reason code、客户文案、证据要求、适用范围、审批状态 |
| Explanation Composer | 生成客户可见解释、客服脚本、内部摘要 |
| LLM Explanation Assistant | 在 approved context 内回答客户追问和帮助提交申诉 |
| Guardrail / Policy Engine | 校验不支持原因、过期政策、敏感泄露、渠道权限 |
| Evidence Graph Store | 保存 PROV 风格 evidence bundle |
| Contestability Case System | 管理争议、申诉、纠错、复核和结果通知 |
| Human Review Workbench | 给复核员展示证据和操作控件 |
| Audit and Metrics Layer | 支持 KRI、抽样、监管查询和模型验证 |
10.2 Sequence
1. Customer applies / transaction occurs / account event happens
2. Decision Orchestrator calls model, rules and policy services
3. Reason Attribution Service derives approved reason codes
4. Evidence Graph Store receives decision snapshot and provenance bundle
5. Explanation Composer generates notice or explanation using approved templates
6. Guardrail checks reason support, policy version, language and leakage
7. Notice Service delivers customer communication and archives proof
8. Customer asks follow-up through assistant or starts appeal
9. Contestability Case System assembles evidence pack
10. Human reviewer resolves case and triggers re-run if needed
11. Audit and Metrics Layer updates KRIs and retention records
10.3 Build vs Buy Questions
| Question | Why it matters |
|---|---|
| Vendor can provide reason codes that satisfy internal adverse action policy? | 供应商 feature importance 不一定等于合规 reason |
| Can decision snapshots be exported with stable versions? | 无法复现就无法审计 |
| Can LLM outputs be constrained to approved templates? | 自由文本会产生 unsupported reason risk |
| Does the platform support appeal workflow and evidence package? | 解释没有 contestability 只完成了一半 |
| Can segment-level reason distribution be monitored? | 公平性和投诉风险需要组合层监控 |
| Are logs usable without exposing trade secrets or sensitive fraud rules? | 客户透明和安全边界需要平衡 |
11. Governance Operating Model
11.1 RACI
| Activity | Accountable | Responsible | Consulted | Informed |
|---|---|---|---|---|
| Explanation policy | Compliance | AI Governance / PM | Legal、Model Risk、Business | Engineering、Operations |
| Reason catalog | Credit / Product Policy | PM、Risk Analytics | Fair Lending、Legal、Operations | Customer Care |
| Template approval | Legal / Compliance | Content Design、PM | Accessibility、Business | Channel teams |
| Reason attribution method | Model Risk / Risk Analytics | Data Science、Architecture | Fair Lending、Engineering | Compliance |
| Contestability workflow | Business Owner | Operations、PM | Compliance、Legal、Customer Care | Engineering |
| Evidence architecture | Architecture | Engineering、Data Governance | Security、Privacy、Audit | Business |
| KRI monitoring | Risk / Compliance | Model Ops、Product Ops | PM、Fair Lending | Management |
11.2 Change Management
任何以下变更都应触发 explanation impact assessment:
- 模型版本、scorecard、ruleset、threshold 或 policy rule 变更。
- reason catalog 新增、合并、停用或文本改写。
- notice template、渠道文案、客服脚本、LLM prompt 或 RAG knowledge base 变更。
- 供应商数据源、bureau mapping、income verification provider、fraud signal provider 变更。
- 申诉流程、SLA、人工复核权限或 outcome templates 变更。
- 监控阈值、fairness segment、complaint taxonomy 变更。
11.3 Release Gate
上线前必须通过:
| Gate | Evidence |
|---|---|
| Legal / Compliance approval | 模板、reason catalog、客户权利路径、监管适用性评估 |
| Model validation | reason faithfulness、stability、fairness、performance、limitations |
| UAT | notice generation、chat explanation、appeal intake、human review、re-run |
| Security / privacy review | 数据最小化、访问控制、敏感规则保护、日志保留 |
| Operational readiness | reviewer training、SLA、queue capacity、escalation path |
| Monitoring readiness | dashboards、alerts、audit query、incident runbook |
12. Interview Section
12.1 30-second answer
我会把 explainability 设计成受治理的产品接口,而不是让模型输出 raw chain-of-thought。核心是先建立 decision reason source of truth:adverse action reason code 必须来自决策系统、规则、模型归因和证据映射,禁止 LLM 自行发明原因。然后按受众分层:客户看到简明、可行动的解释和申诉入口;审核和监管看到 reason、证据、版本、通知和人审记录;模型验证看 faithfulness、consistency 和公平性指标。最后把 contestability 做成完整 workflow:客户可争议、可补充材料、人工复核、纠错、重跑、结果通知,并用 evidence pack 支撑审计。
12.2 2-minute answer
在 regulated financial AI 中,我不会把 explainability 当成一个模型特性,而会把它当成一个 end-to-end control and product architecture。第一步是界定场景:是信贷 adverse action、透支费解释、欺诈误报、财富适当性,还是普通客服说明。不同客户影响对应不同解释深度和审计要求。
第二步是建立 reason architecture。以信贷为例,reason code 必须有 source of truth,来自 decision engine、ruleset、model attribution 或人工复核系统,并且要映射到 evidence object、policy version 和 customer template。LLM 不能发明拒绝原因,只能在 approved reason 和 approved evidence summary 之内做 plain-language explanation。这样才能保证正式 notice、客服脚本、chatbot 回答和审计材料一致。
第三步是 explanation layering。客户层解释要清楚说明主要原因、使用信息类型、客户可以做什么;监管和内审层要看到 notice、reason、证据、版本、发送时间和人审记录;模型风险层要评估 reason faithfulness、stability、segment disparity 和 model upgrade drift;工程层要保留可复现 trace,但不能把 debug trace 或 raw reasoning 暴露给客户。
第四步是 contestability。解释如果不能被争议,只是单向通知。我要设计 dispute / appeal intake、身份验证、证据包组装、人工复核、数据纠正、用 corrected data 重跑或重裁、结果通知和投诉联动。指标上我会看 explanation correctness、unsupported reason rate、appeal upheld rate、segment disparity in explanations、stale policy citation、human override、complaint linkage 和 reproducibility rate。这样 explainability 才从“看起来透明”变成可运营、可审计、可合规的系统能力。
13. Portfolio Exercise
13.1 练习题:设计一个信用卡拒绝解释与申诉系统
背景:一家区域银行计划上线 AI-assisted credit card underwriting。模型参与申请评分,规则引擎处理政策拒绝,人审处理边界样本。你要设计客户拒绝解释助手、adverse action notice reason pipeline、申诉流程和审计证据包。
交付物
| Artifact | 内容要求 |
|---|---|
| Product one-pager | 目标用户、业务目标、风险边界、成功指标 |
| Reason-code catalog sample | 至少 12 个 reason codes,包含 evidence mapping 和 customer language |
| Decision-to-notice sequence | 从申请到 decision、reason attribution、notice、assistant、appeal 的序列图 |
| Contestability workflow | intake、triage、evidence assembly、human review、correction、re-run、outcome |
| Evidence pack schema | decision snapshot、model version、ruleset、reason、notice、review、appeal |
| Eval plan | faithfulness、consistency、fairness、language、accessibility、re-run reproducibility |
| Metrics dashboard | correctness、unsupported reason、appeal upheld、segment disparity、complaint linkage |
| Governance RACI | Legal、Compliance、Fair Lending、Model Risk、Product、Engineering、Ops |
13.2 Suggested Architecture Diagram Text
[Application Channels]
|
[Decision Orchestrator] -- [Rules Engine]
| -- [Model Scoring Service]
| -- [Manual Review Workbench]
|
[Reason Attribution Service]
|
[Reason Catalog + Policy Registry]
|
[Explanation Composer + LLM Guardrails]
|
[Notice Service] ---- [Customer Explanation Assistant]
| |
[Evidence Graph Store] |
| |
[Contestability Case Management + Human Review]
|
[Audit Query + KRI Dashboard]
13.3 Evaluation Scenarios
| Scenario | Expected behavior |
|---|---|
| Customer denied due to policy knockout | Notice shows approved policy reason and evidence pack links to rule id |
| Customer denied due to model score with multiple negative factors | Principal reason selection returns ranked approved reason codes |
| LLM tries to mention an unapproved reason | Guardrail blocks output and escalates to fallback template |
| Customer says bureau record is wrong | Case type becomes data dispute, evidence pack assembled, correction workflow starts |
| Customer submits new income documents | Reviewer verifies documents, recomputes income feature, triggers re-run policy |
| Model version changes | Regression eval checks reason drift and individual reason flips |
| Spanish explanation generated | Semantic consistency test confirms same reason and rights |
| Auditor asks for a case from six months ago | Evidence store reproduces decision snapshot, notice, reason, model, policy and appeal record |
14. Checklist
14.1 Product and Compliance Checklist
- 每个客户影响性 use case 已完成 risk tier 和 regulatory impact assessment。
- adverse action reason 的 source of truth 不在 LLM。
- reason catalog 已定义 owner、版本、适用范围、证据要求和停用机制。
- 客户解释、notice、客服脚本和 chatbot 回答使用同一 reason source。
- LLM 输出被限制在 approved reason、approved template 和 approved evidence summary 内。
- 客户能清楚看到纠错、申诉、补充材料、人工联系和结果通知路径。
- 申诉流程支持 human review、source data correction、re-run / reconsideration 和 outcome notification。
- 高风险 case 支持 maker-checker、supervisor approval 或 specialized review。
- 模板已通过 Legal、Compliance、Fair Lending、Accessibility 和 Business review。
- 多语言解释经过语义一致性检查。
14.2 Architecture and Data Checklist
- decision_id 能贯穿申请、决策、reason、notice、chat、appeal、audit。
- 原始 decision snapshot、feature snapshot、model version、ruleset version 和 policy version 可保留并复现。
- 每个 reason code 映射到 evidence object,不只映射到自然语言。
- evidence pack 自动生成,不依赖人工事后整理。
- debug trace 与 customer-facing explanation 分层存储和授权访问。
- fraud / risk sensitive details 有披露边界和 redaction policy。
- stale policy citation validator 已接入 notice 和 assistant。
- audit query catalog 覆盖 notice timeliness、reason consistency、appeal outcomes、segment disparity。
- release pipeline 包含 reason drift、template regression 和 LLM faithfulness eval。
- data correction workflow 不覆盖原始 snapshot,而是产生可追溯 corrected data object。
14.3 Metrics and Operations Checklist
- explanation correctness 被抽样和自动化监控。
- unsupported reason rate 有阻断阈值和事件升级。
- appeal upheld rate 按 reason、产品、渠道、segment 分解。
- segment disparity in explanations 定期进入 Fair Lending / Compliance review。
- stale policy citation rate 有 owner 和修复 SLA。
- human override rate 被用于识别 automation bias 或模型质量问题。
- complaint linkage 能把 explanation failure 与投诉、监管投诉、社媒升级连接。
- contestability completion rate 和 drop-off 被用于优化客户旅程。
- reviewer training 覆盖 AI limitations、reason policy、bias、customer vulnerability 和 evidence review。
- 每次重大 incident 后更新 reason catalog、eval set、workflow 和 audit queries。
15. 一页总结
金融零售 AI 的 explainability 不是展示模型“为什么这么想”,而是把客户影响性决策变成可解释、可争议、可复核、可纠正、可审计的产品能力。最关键的架构决策是把 reason code source of truth 从 LLM 中剥离出来,放到受治理的 decision / reason attribution layer;把客户解释、内部证据、模型验证、工程 trace 和 internal reasoning traces 分层;把 adverse action reason 与 evidence、template、policy、version 和 notice 绑定;把 contestability 做成端到端 case workflow;把 audit evidence 设计成运行时自动生成的 provenance graph。
一句话作品集表达:
I design explainability as a governed product interface: reason codes come from the decision system, explanations are audience-layered, adverse action notices are evidence-backed and consistent, contestability is a full appeal and correction workflow, and every decision produces an audit-ready provenance package.