AI 影子模式架构：Shadow Mode / Counterfactual Evaluation / Silent Launch

Date: 2026-06-30
Status: evergreen
Audience: experienced CBAP / financial retail PM / product architect / solution architect / AI governance lead
Output: advanced architecture note, decision model, evidence pattern, interview asset

---

Why Shadow Mode Matters For AI Product/Architecture

Shadow mode 是把 AI 系统接入真实业务上下文, 但不让它影响客户、员工操作、系统状态或监管承诺。它回答的不是“模型在测试集上分数高不高”, 而是更接近架构和产品决策的问题:

这份笔记刻意不重复 online experimentation、UAT regression certification、release governance 或 adoption analytics。它定位在更早一层: pre-production and pre-decision architecture。在金融零售里, 这层能力决定 AI 能不能从 demo 进入真实决策流程。

一句话:

Shadow mode turns production-like traffic into decision evidence before the AI is allowed to change customer outcomes.

---

Concept Diagram

flowchart LR
  A[Live business event] --> B[Current champion path]
  A --> C[Shadow decisioning path]

  B --> B1[Human / rules / existing model decision]
  B1 --> B2[Actual action taken]
  B2 --> B3[Delayed outcome / label]

  C --> C1[Feature snapshot]
  C1 --> C2[AI challenger decision]
  C2 --> C3[Counterfactual event log]

  B1 --> D[Decision comparison]
  C2 --> D
  B3 --> E[Outcome join]
  C3 --> E
  D --> F[Segment / fairness / control analysis]
  E --> F
  F --> G[Gate memo]
  G --> H{Rollout decision}
  H -->|No-go| I[Redesign / retrain / control update]
  H -->|Limited go| J[Assisted mode with HITL]
  H -->|Go| K[Controlled production rollout]

架构边界:

• Champion path 仍然拥有业务决定权。
• Challenger path 只能观察、推理、记录、解释, 不能写业务状态。
• Counterfactual log 必须能在 outcome 延迟后回放和归因。
• Gate decision 必须同时看效果、风险、可解释性、操作准备和证据质量。

---

Core Architecture Model

1. 核心组件

2. Shadow Mode 的四个成熟度级别

关键原则:

• 不允许 shadow 输出触发客户通知、价格变更、额度变更、冻结、拒绝、催收动作或 SAR / regulatory filing。
• 如果需要员工看到 AI 建议, 必须明确是否会改变人工判断。否则会把 silent launch 变成事实上的 pilot。
• 所有分析必须区分 “AI would have recommended” 和 “business actually did”。

3. Decision Object

Shadow mode 不是只记录一个 score。每次 AI 决策至少应保存:

---

Counterfactual Logging And Evaluation Lifecycle

1. Lifecycle

use case intake
  -> decision boundary and prohibited actions
  -> feature and policy snapshot design
  -> shadow router implementation
  -> counterfactual event logging
  -> daily quality and control checks
  -> delayed label / outcome join
  -> human comparison and disagreement review
  -> segment / fairness scorecard
  -> gate memo and rollout recommendation
  -> evidence archive

2. Logging Model

Counterfactual logging 要能回答三个问题:

3. Outcome Delay Handling

金融零售很多 outcome 不会当天成熟:

Outcome delay 不是纯数据问题, 是产品决策问题。过早 gate 会高估速度和低估损害; 过晚 gate 会拖慢学习。成熟做法是把 gate 分成:

• Readiness gate: 输入完整性、日志质量、控制完整性、初步一致性。
• Risk gate: 高风险 disagreement、fairness slice、prohibited action、leakage、complaint-sensitive cases。
• Outcome gate: 标签成熟后的 lift / harm / false positive / false negative / cost / operational burden。

4. Leakage Control

Shadow mode 最常见的假进步来自 leakage。

---

Gate Criteria And Rollout Decision Model

1. Gate Layers

2. Decision Model

3. Rollback Triggers Before Full Rollout

Rollback 不只适用于 production release。Shadow / silent launch 阶段也需要 stop rules:

• Trace completeness below agreed threshold for two consecutive business days.
• Challenger produces prohibited actions, unauthorized tool intent, or customer-impacting recommendation outside approved scope.
• Material leakage discovered in feature, label, prompt context, or human comparison protocol.
• High-severity disagreement rate exceeds threshold in credit, fraud, AML, KYC or vulnerable customer slices.
• Fairness scorecard shows unexplained false positive / false negative disparity in protected or proxy segments.
• Reviewer queue load exceeds planned capacity and delays existing controls.
• Evidence binder cannot reconstruct model, prompt, data, policy and decision versions.

---

Financial Retail Scenarios

1. Credit Line Management

Shadow question: if AI recommended line increase / decrease / hold, would it improve risk-adjusted growth without unfair treatment or adverse action inconsistency?

2. AML Alert Triage

Shadow question: can AI triage alerts, summarize rationale, and recommend priority without missing suspicious activity or creating analyst automation bias?

3. KYC Onboarding

Shadow question: can AI reduce manual review and detect synthetic identity risk without discouraging legitimate applicants?

4. Payment Fraud Intervention

Shadow question: would AI intervene on risky payments with lower fraud loss and fewer false declines?

5. Collections Contact Strategy

Shadow question: would AI choose a better contact channel, timing and treatment while respecting hardship, consent and conduct risk?

6. Contact Center Agent Assist

Shadow question: can a copilot suggest policy-grounded answers and next best actions without misleading agents or changing regulated communications?

---

Metrics/Control/Evidence Model

1. Metrics

2. Control Model

3. Evidence Packet

一份可进入 governance review 的 packet 应包含:

1. Use case and decision boundary.
2. Customer and employee impact statement.
3. Champion/challenger architecture diagram.
4. Data, feature, prompt, RAG, model and policy version lineage.
5. Counterfactual schema and sample trace.
6. Leakage assessment and remediation record.
7. Human comparison protocol and calibration results.
8. Outcome label plan and maturity analysis.
9. Segment/fairness scorecard.
10. Gate recommendation with residual risks and controls.
11. Rollout limits, rollback triggers and owner map.
12. Audit-ready evidence index.

---

Anti-Patterns And Failure Modes

---

Architecture Mapping To RAG / Agent / Copilot / Eval / Governance

---

ADR Draft

---

Interview Answer

30秒版本

Shadow mode 是在不影响客户和业务状态的前提下, 让 AI 读取真实业务事件并记录“如果由 AI 决策会怎么做”。我会把 champion 决策、AI challenger 输出、决策时点特征、延迟 outcome、人工复核差异、segment fairness 和审计证据全部记录下来, 再用 gate 判断是否进入 assisted mode 或受控 rollout。重点不是跑一个模型分数, 而是在客户无影响阶段证明 AI 决策边界、风险和运营准备。

2分钟版本

我会先定义决策边界: AI 是建议额度、排序 AML 告警、建议 KYC 处理、拦截支付、推荐催收策略, 还是辅助客服回答。然后设计 read-only shadow path, 让真实业务事件同时进入现有 champion 流程和 AI challenger。Champion 仍然做实际决定; challenger 只能输出建议、置信度、原因、引用和是否 abstain。

核心是 counterfactual event log。它要保存决策时点的 feature snapshot、policy version、RAG corpus、model/prompt/tool 版本, 还要保存实际人工或规则决定。等 outcome 成熟后, 比如欺诈确认、逾期、SAR 结果、投诉、QA 分数, 再回连分析 AI 原本会带来什么收益或伤害。

Gate 不能只看平均准确率。我会看 high-severity disagreement、false positive/negative、fairness slice、leakage、人工复核一致性、trace 完整性、操作队列负荷和 rollback trigger。如果证据不足, 继续 shadow; 如果低风险场景稳定, 先进入 human-approved assisted mode; 如果出现 segment harm 或泄漏, no-go 并回到数据、模型或流程修复。

CTO版本

我会把 shadow mode 当成一个 pre-decision control plane, 而不是一次测试活动。架构上, production event 进入现有 champion path 的同时, 复制最小必要上下文到只读 challenger path。Challenger 的任何 tool intent 都在 sandbox 中记录但不执行; 所有输出都带 trace id、model/prompt/RAG/policy/tool 版本和 decision-time feature snapshot。

数据层需要 append-only counterfactual store、outcome joiner、segment analyzer 和 evidence binder。控制层需要 leakage registry、population sampling plan、blind human comparison、fairness hard gates、operational readiness 和 rollback triggers。这样我们可以在不改变客户结果的前提下回答: AI 会在哪些场景改善决策, 会在哪些场景制造客户伤害, outcome 延迟是否改变结论, 一线团队能否承接, 审计能否重建整个判断链。

我不会让高影响金融 AI 直接从 offline eval 跳到 customer-impacting release。合理路径是 replay -> silent shadow -> human comparison -> assisted mode -> narrow rollout。每一层都要有 gate memo 和 residual risk decision。这样 CTO 可以向风险、审计和业务解释: 我们不是凭 demo 上线, 而是用生产相似证据逐步放权。

---

7-Day Practice Plan

---

Source Anchors