AI Closed-Loop Learning / Corrective Action Architecture 解读

面向对象: AI PM / AI BA / AI Architect / Model Risk Partner / Customer Operations Lead / 金融零售 AI 治理负责人。 核心问题: 生产反馈、人工覆盖、投诉、专家审核、eval 失败、漂移信号、事件和审计发现如果只进入 dashboard 或 backlog, AI 系统不会真正变好。 学习目标: 区分 active learning 与 CAPA-like corrective action loop, 掌握 issue taxonomy、root cause、change linkage、更新治理和 effectiveness verification。

---

Source Anchors

Closed-loop learning 是把 AI 失败和弱信号从“可观察”推进到“已纠正、已验证、可审计、可防复发”。

---

1. Thesis

Active learning 问的是:

Which samples should humans label next to improve the model?

Closed-loop corrective action 问的是:

Which production issue harmed quality, customers, controls or risk assumptions?
What was the root cause?
Which data, model, prompt, RAG, tool, workflow or policy changed?
What evidence proves the fix worked and did not create new risk?

两者可以连接, 但不能混同。Active learning 是学习样本选择机制; corrective action architecture 是治理过的生产改进机制。它覆盖:

---

2. Why It Matters

金融零售 AI 的生产失败往往不是单一模型错误:

• 客服 RAG 引用旧费用政策, 客户错过申诉窗口。
• 欺诈模型误拦截上升, 员工大量 override, 但没有结构化原因。
• 信贷解释生成器使用不完整 reason code, 客户投诉“原因不对”。
• 审计发现模型更新后没有证明旧缺陷已被关闭。
• 漂移 dashboard 报警, 但没有 owner、SLA、修复证据和复发监控。 高成熟度组织会把这些信号汇聚成 corrective action ledger:

signal -> issue classification -> severity and customer impact
  -> root cause analysis -> corrective and preventive action
  -> linked change artifacts -> release gate
  -> effectiveness verification -> recurrence monitoring -> closure evidence

---

3. Core Concepts

3.1 Feedback Is Not Automatically Truth

3.2 CAPA-like Loop

3.3 Change Linkage

issue_id -> root_cause_id -> corrective_action_id -> change_request_id
  -> dataset_version / prompt_version / model_version / retriever_version / tool_policy_version
  -> eval_run_id -> approval_record -> post_fix_monitoring_window

没有 change linkage 的“已修复”通常只是口头关闭。

4. Architecture Diagram

Feedback and risk signals
  complaints | appeals | overrides | expert QA | eval failures | drift alerts | incidents | audit findings
        |
        v
Signal normalization and AI contribution matching
  system id | model/prompt/RAG/tool version | journey step | customer impact | segment
        |
        v
Issue taxonomy and severity engine
  model | data | policy | prompt | retrieval | tool | workflow | governance | vendor
        |
        v
CAPA ledger
  containment | RCA | corrective action | preventive action | owner | SLA | evidence status
        |
        v
Change linkage layer
  dataset registry | eval registry | prompt repo | model registry | vector index | tool policy | SOP
        |
        v
Release and approval gate
  risk review | model validation | compliance | operations readiness | rollback plan
        |
        v
Effectiveness verification and closure
  regression eval | replay | segment metrics | complaint trend | override rate | recurrence

---

5. Financial Retail Case

Case: Credit card servicing RAG gives stale fee-waiver guidance

Signals: complaint says chatbot gave wrong fee waiver condition; internal “fee explanation wrong” complaints rise; agent override rate increases; RAG eval shows policy effective-date failures; drift monitor finds a new annual-fee cluster after campaign launch.

• Deactivate superseded source and rebuild vector index.
• Add effective-date filter and source authority ranking.
• Update prompt to require current policy citation for fee claims.
• Add eval cases for fee waiver, annual fee, refund and dispute intents.
• Route conflicting citation answers to human review. Effectiveness evidence:
• Unsupported fee-policy claim rate drops from 14% to 3% on locked eval set.
• Human QA confirms 50 sampled production answers cite current policy.
• Fee explanation complaints return to baseline for two consecutive weeks.
• Agent override rate for fee answers drops below pre-incident average.
• No repeat H2+ harm cases in 30-day post-fix window.

---

6. PM / BA / Architect Checklist

---

7. Code-lite Experiment

Goal: show that closed-loop learning needs traceability, not just a better model score.

issues = [
    {"id": "I-101", "signal": "complaint", "theme": "wrong_fee_policy", "severity": "H2"},
    {"id": "I-102", "signal": "agent_override", "theme": "wrong_fee_policy", "severity": "H1"},
]
changes = [
    {"id": "CR-77", "artifact": "retriever_config", "version": "v5.3", "root_cause": "superseded_source_ranked_first"},
    {"id": "CR-78", "artifact": "rag_eval_set", "version": "fee_policy_gold_v2", "root_cause": "missing_effective_date_filter"},
]
verification = [
    {"change": "CR-77", "metric": "unsupported_claim_rate", "before": 0.14, "after": 0.03, "pass": True},
    {"change": "CR-78", "metric": "fee_policy_eval_pass_rate", "before": 0.82, "after": 0.97, "pass": True},
]
closed = all(v["pass"] for v in verification)

讨论: 如果 issues 没有 link 到 changes, 就无法证明问题被修; 如果只有 eval 变好, 但投诉和 override 没下降, 不能关闭 CAPA; 如果 change 只改 retriever, 但 root cause 包含 governance, 还需要 source owner SOP。

8. Interview Questions

Q1: Closed-loop learning 和 active learning 的区别?

30 秒回答: Active learning 关注选择哪些样本让专家标注来提升模型。Closed-loop learning 关注生产问题如何被分类、找根因、链接到具体变更并验证有效。它更像 AI 版 CAPA, 覆盖数据、模型、prompt、RAG、tool、流程、政策和控制。

Q2: 如何证明一个 AI 修复真的有效?

30 秒回答: 我会要求四类证据: locked regression eval 通过, 生产 replay 或 shadow 结果改善, 相关投诉 / override / QA defect / drift KRI 回到阈值内, 且在定义好的 post-fix monitoring window 没有复发。只说“prompt 已更新”不够。

Q3: SR 26-2 对闭环学习有什么启发?

30 秒回答: SR 26-2 是 2026 年银行模型风险管理的重要锚点, 强调风险导向、生命周期治理、独立挑战和问题整改。Nuance 是它的正式 scope 是模型风险管理, 生成式和 agentic AI 的闭环学习还要叠加 AI 管理体系、消费者保护、隐私、安全、第三方和运营控制。

9. Pitfalls

---

10. Portfolio Task

做一个 “AI Corrective Action Control Pack”: