AI 底层逻辑 / 经典论文

AI Executive Investment Narrative：商业案例与董事会决策架构

436 行ai-foundations/papers/152-ai-executive-investment-narrative-business-case-board-decision-architecture.md

AI Executive Investment Narrative / Business Case / Board Decision Architecture 解读

配对阅读：本篇的操作手册版（模板/RACI/门禁/runbook）是 docs/AI_EXECUTIVE_INVESTMENT_NARRATIVE_BUSINESS_CASE_PLAYBOOK.md。第一遍读本篇建立原理与架构判断；第二遍做案例时再用 playbook 查表落地，两者不需要重复精读。

核心导读

AI 高管投资叙事不是把模型能力包装成 ROI, 而是把不确定的 AI 机会转成可治理的管理选择权。成熟的 board decision package 要同时回答 outcome thesis、option architecture、causal value logic、cost-to-learn、risk appetite、architecture dependency、benefits realization、evidence binder 和 stop / scale / pivot gate。对有金融零售软件经验的人来说, 重点不是解释 AI 为什么有潜力, 而是证明管理层为什么可以在特定证据、特定风险偏好和特定架构条件下投入下一阶段资金。

重要说明: 本文是学习、作品集和内部架构训练材料, 不构成法律意见、监管解释、合规结论、审计意见、财务投资建议、会计确认、估值建议、董事会治理意见或生产上线批准。正式项目必须由机构授权角色结合司法辖区、牌照、客户群、产品、风险偏好、财务政策、模型风险、信息安全、隐私、供应商合同、内部审计和监管关系确认。访问日期沿用 2026-06-30。

0. Source Anchors and Reading Boundary

Source	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 组织风险识别、度量、治理证据和持续改进。
ISO/IEC 42001:2023 AI management systems	https://www.iso.org/standard/81230.html	用 AI management system、risk and opportunity、operation、performance evaluation 和 management review 组织投资治理。
FFIEC IT Examination Handbook InfoBase	https://ithandbook.ffiec.gov/	用 board oversight、IT investment planning、project business case、architecture 和 board reporting 校准金融机构场景。
ISO/IEC/IEEE 42010:2022 Architecture Description	https://www.iso.org/standard/74393.html	用 stakeholder concerns、viewpoints、model kinds 和 rationale 组织 board decision architecture。

本文默认读者已经熟悉银行、支付、零售运营、风险控制、监管报送和企业架构治理。因此主线放在如何把 AI use case 写成高管可以挑战、限制、分阶段批准或停止的投资决策。

1. 核心问题: AI 投资为什么经常无法进入可靠决策

许多 AI 项目在业务评审中失败, 不是因为场景没有价值, 而是因为叙事停留在功能层。材料会说模型能总结投诉、提升客服效率、帮助 AML 分析师、自动生成监管解释, 却没有说明价值如何兑现、风险如何留在 appetite 内、架构依赖是否可复用、哪些证据会改变管理层判断。

成熟的 AI investment narrative 要回答八个管理问题:

Which business outcome is worth funding?
Which non-AI and AI options are being compared?
How exactly does AI behavior translate into realized value?
What is the cheapest credible evidence we can buy next?
Which risks cannot be traded away for efficiency?
Which architecture and operating capabilities must exist before scale?
Who recognizes benefits, accepts residual risk and owns management actions?
When do we stop, pivot, restrict or scale?

低成熟度材料通常是线性的:

Use GenAI to automate customer service.
Expected benefit: reduce handle time by 30%.
Investment: model integration + RAG.
Risk: manageable with human review.
Decision: approve funding.

更可靠的叙事是分阶段的:

Fund an evidence stage for regulated service intelligence in two complaint and policy-answer workflows.
The thesis is that grounded assistance can reduce agent research load and improve policy citation quality
without increasing complaint, reopen or customer harm rates.
Scale funding is not requested now.
The board is asked to approve a capped discovery-to-pilot envelope,
with kill criteria, architecture conditions and risk appetite thresholds before expansion.

这组问题把 AI 投资从技术提案转成 decision architecture。高管不是为模型本身付费, 而是为一组受控选择权付费: 先用有限资金买证据, 再根据证据决定是否继续。

2. 方法框架: 投资叙事是一套架构化证据链

一份可被董事会、高管、CFO、COO、CIO/CTO、风险、合规和内审共同挑战的 AI decision package, 应沿着以下链条组织:

business problem and strategic context
  -> outcome thesis
  -> option architecture
  -> causal value logic
  -> evidence and confidence level
  -> architecture dependency and control readiness
  -> cost-to-learn and funding envelope
  -> risk appetite and residual risk
  -> benefits realization method
  -> stop / scale / pivot gates
  -> board decision and management action

这条链把不同专业域的关切放到同一张决策图上。业务关注 outcome, 财务关注 recognition, 风险关注 appetite, 技术关注 dependency, 运营关注 adoption, 审计关注 evidence。如果这些关切被拆成不同附件, 决策会变成口头信任; 如果它们被组织成一条证据链, 决策才能被复核和追踪。

2.1 Outcome Thesis

AI outcome thesis 要把 AI 能力、目标人群、业务结果、因果逻辑、人类责任边界和学习计划写清楚。

We believe [AI capability] can improve [business outcome]
for [specific population / workflow]
because [causal value logic].

The investment is attractive if:
- value indicator improves versus baseline,
- risk guardrail remains inside appetite,
- adoption proves workflow behavior change,
- unit economics remains viable,
- architecture dependency is reusable or controlled.

We learn this through [evidence stage] at cost [cost-to-learn],
and stop, pivot or scale based on [decision criteria].

维度	弱 thesis	强 thesis
Outcome	提升效率	把投诉 root-cause 分析周期从月度手工抽样改成每周 evidence-backed insight, 降低 repeat complaint 和 remediation delay。
AI role	使用 LLM 总结	AI 负责聚类、证据检索和 draft insight; regulatory interpretation、customer remediation 和 issue closure 由授权人负责。
Causal logic	AI 节省人工时间	case classification 和 evidence retrieval 变好, QA reviewer 才能更快识别 systemic issue。
Evidence	用户反馈很好	baseline、holdout、QA defect、complaint repeat rate、action closure time、finance-recognized capacity release。
Risk	human in the loop	高风险投诉结论必须引用 approved source, reviewer sign-off, customer harm threshold, stale source stop rule。
Funding ask	批准项目预算	批准 8 周 cost-to-learn envelope, scale 资金另走 gate。

2.2 Option Architecture

AI board packet 不应只呈现一个推荐方案。真正的管理问题通常是如何在 no-action、流程改造、规则自动化、AI assistant、AI automation、shared platform、vendor product 和 hybrid staged option 之间取舍。

Option	适合场景	好处	风险 / tradeoff	Evidence needed
Do nothing / monitor	问题不严重或证据不足	不引入新成本和模型风险	机会成本、旧问题继续累积	baseline trend、risk exposure、customer harm trend
Process redesign only	痛点来自流程、权限、SOP、handoff	成本低、风险低	不解决知识检索或复杂判断负担	process mining、handoff defect、training impact
Rules / workflow automation	规则明确、数据结构化	可解释、可控、低模型风险	覆盖不了非结构化证据和复杂文本	rule coverage、exception rate、maintenance cost
AI assistant / copilot	非结构化文本、摘要、检索、建议	快速增强员工能力	adoption、hallucination、source freshness、review burden	eval、QA sample、workflow adoption、traceability
Shared platform	多 use case 重复能力明显	复用、成本、治理一致性	前期投入较高, 需产品化运营	reuse count、platform economics、service catalog
Vendor product	能快速获得成熟能力	time-to-market 快	vendor lock-in、data boundary、evidence export	third-party risk、exit plan、evidence completeness
Hybrid staged option	价值不确定但学习价值高	先买证据, 再决定规模	管理复杂度更高	gate evidence、kill criteria、confidence trend

Real options thinking 的核心是: 早期资金买的是学习权, 不是规模化承诺。 Discovery / pilot 的小额资金是 option premium; value、risk、adoption、architecture readiness 是 exercise condition; 结束日期是 expiry; 早停留下的 eval set、failure taxonomy 和 data quality insight 是 abandonment value。

2.3 Architecture Dependency

Architecture is not a technical appendix. It changes the funding decision because hidden dependencies become hidden cost, hidden risk and hidden delay.

Dependency	Board-relevant question
Model gateway	Can management control model routes、costs、versions、logs and fallbacks?
RAG / knowledge service	Are sources approved、fresh、permission-filtered and citable?
Workflow integration	Does AI output enter the work system, or remain side-channel advice?
Human review queue	Is there enough SME and supervisor capacity to keep controls credible?
Eval platform	Can prompt、model、RAG and tool changes be regression-tested?
Observability	Can value、harm、cost、adoption、latency and incidents be monitored?
Evidence binder	Can audit reconstruct claims、approvals、outputs and management actions?
Vendor contract	Are evidence export、data boundaries、exit and resilience terms sufficient?

3. Business Case Model: 从 ROI 表格到因果价值模型

AI business case 必须避免把 "AI 输出质量" 直接等同于财务收益。中间需要完整 causal chain:

AI capability
  -> behavior quality
  -> workflow adoption
  -> process performance
  -> business outcome
  -> financial / risk benefit
  -> recognized benefit

Layer	问题	Evidence
Baseline	当前业务量、成本、质量、风险、等待时间、客户影响是什么	workflow data、finance cost model、QA、complaints、incidents
Intervention	AI 改变哪个步骤、谁使用、AI 权限到哪一级	target process、AI role、RACI、architecture boundary
Leading indicators	scale 前能快速看见什么信号	eval score、citation correctness、acceptance rate、review time
Lagging benefits	业务结果是否变好	AHT、backlog、STP、loss avoided、complaint repeat、regulatory cycle time
Guardrails	不允许用什么代价换收益	harm、policy breach、fairness、privacy、AML quality、source freshness
Unit economics	每个合格价值事件的全成本是多少	model cost、platform allocation、review、QA、support、training
Attribution	怎么知道改善来自 AI 与流程改变	holdout、before-after with control、cohort comparison、process mining
Recognition	谁能确认收益进入管理账	business owner、finance reviewer、risk owner、operations owner

Benefit type 也要区分。 Productivity capacity 不能直接把理论节省分钟数当现金节省; cost avoidance 不能把一次性峰值避免当长期 saving; risk reduction 不能用 "更多 alerts reviewed" 代替风险降低证据; revenue enablement 不能忽略 fraud / KYC review 增量成本; customer experience 不能只看 sentiment; platform leverage 不能没有 reuse evidence。

Confidence	证据标准	决策含义
Low	SME estimate、vendor benchmark、small interview signal	只能支持 discovery, 不能支持 scale commitment。
Medium	baseline 数据可用, pilot 或 offline eval 有方向性结果	可支持 pilot 或 limited release, 需要 stronger gate。
High	生产 cohort、holdout、QA、finance recognition 和 risk trend 支撑	可支持 scale decision, 仍需 monitor。
Declining	生产分布、成本、adoption 或风险趋势恶化	hold、pivot 或 stop。

Cost-to-learn discipline 要回答五件事:

What do we need to learn?
What is the cheapest credible way to learn it?
Which scarce capacities are consumed?
What evidence will we have by what date?
Which decision will that evidence enable?

Learning question	Cheap credible test	不成熟做法
AI 能否理解投诉原因	历史投诉样本 offline eval + QA reviewer rubric	直接接入生产投诉系统做 live pilot
员工会不会采用 copilot	concierge pilot 或 limited workflow trial	只做 demo 后问满意度
RAG 引用是否可靠	controlled knowledge set + source freshness tests	用全量知识库, 后期再补治理
收益是否可兑现	4 周 cohort + finance baseline review	用 vendor ROI calculator
风险能否受控	shadow mode + red-team + stop-rule drill	写 "human review" 作为单一控制

Risk appetite 不能停留在口号。它必须翻译成投资条件、禁止边界、阈值、forum 和 stop rule。

Risk appetite statement	Investment implication
No appetite for AI making final adverse customer decisions without authorized human decision	Credit、complaint denial、account closure and AML conclusions remain human-owned unless explicitly approved.
Low appetite for unsupported regulated customer communication	Customer-facing GenAI requires source citation、approved language、QA sampling and stop thresholds.
Limited appetite for vendor concentration in material AI systems	Board case must show model / vendor exposure、fallback plan and exit rights.
Low appetite for untraceable AI-assisted records	Investment must fund trace logging、evidence retention and reconstructability.
Appetite for controlled experimentation in internal productivity workflows	Discovery and pilot funding can move faster if data、scope and customer impact are constrained.

4. Board Decision Architecture

董事会材料不是项目状态汇报。它的第一责任是让管理层可以做清楚的 fund、hold、stop、pivot、restrict、scale 或 accept residual risk 决策。

Recommended storyline:

1. Decision requested
2. Strategic context and risk appetite fit
3. Business problem and baseline evidence
4. Outcome thesis and causal value logic
5. Options considered and recommended option
6. Architecture and operating model dependencies
7. Business case model and confidence level
8. Cost, value and risk evidence
9. Governance gates and management information
10. Stop / scale / pivot criteria
11. Conditions, residual risk and management actions

Section	Good content	Weak content
Decision requested	Approve USD X discovery envelope and conditional pilot gate; no scale funding requested.	Approve AI transformation program.
Baseline	quantified current state with source owners	anecdotal pain
Option architecture	options with value/risk/cost/tradeoffs	only preferred solution
Business case	causal chain、confidence、unit economics、benefit owner	single ROI percentage
Architecture	dependency map and platform leverage	model diagram only
Risk appetite	thresholds、unacceptable outcomes、residual risk owner	generic "risk is manageable"
Evidence pack	references to artifacts and owners	screenshots and slide notes

Investment governance should be staged by evidence, not by calendar milestone.

Gate	Decision	Required evidence
Intake gate	enter discovery / park / reject	owner、problem、baseline signal、AI fit hypothesis、initial risk tier
Discovery gate	fund pilot / pivot / stop	workflow map、no-AI option、data readiness、architecture sketch、risk appetite fit
Pilot gate	limited release candidate / continue / stop	eval、QA、SME review、cost、adoption signal、control design、failure taxonomy
Release gate	production limited release / no-go	runbook、monitoring、rollback、risk sign-off、model/prompt/RAG/tool versioning
Scale gate	expand / hold / restrict / stop	realized benefits、unit economics、incident trend、adoption、platform capacity
Portfolio rebalance	fund / merge / retire / platformize	portfolio metrics、capacity、risk concentration、opportunity cost

Gate decision record should include gate id、use case、decision、evidence reviewed、confidence level、conditions、kill criteria、residual risk、management action、owner、due date and closure evidence.

Kill criteria 不是悲观。它说明管理层知道什么证据会改变决策, 也避免 pilot 变成 sunk-cost project。

Scenario	Kill / pivot criteria
AI complaint intelligence	Stop if systemic issue clustering cannot reach agreed QA precision or sensitive complaint categories are misclassified beyond appetite.
Account opening modernization	Stop automation expansion if false missing-document recommendations or unsupported rejection recommendations exceed threshold.
AML triage	Stop if AI summaries omit critical evidence or produce unsupported escalation / closure recommendations.
Regulatory reporting automation	Stop if lineage、calculation reproducibility or maker-checker evidence cannot be reconstructed.
Branch / contact-center copilot	Stop customer-facing expansion if unsupported policy claim、stale source citation or high-risk topic bypass exceeds appetite.

Decision matrix for scale:

Evidence pattern	Decision	Rationale
Value green, risk green, adoption green, unit economics green, architecture reusable	Scale	Evidence supports broader deployment with monitoring.
Value green, risk amber, controls improving	Limited scale / hold	Value exists, but expansion must wait for control maturity.
Value amber, risk green, adoption high	Pivot	Users want help, but value logic or workflow target needs adjustment.
Value green, adoption low	Redesign workflow	AI capability may work, but change management or UX is blocking benefit.
Risk red	Stop / restrict	Customer harm, regulatory, privacy or control failure overrides value case.
Cost red	Optimize / restrict / stop	Unit economics not viable at scale.
Architecture dependency red	Convert to platform or data investment	Use case cannot scale until shared capability is funded.
Confidence declining after release	Hold / roll back	Production evidence invalidates earlier assumptions.

5. Evidence Binder and Management Information

Evidence binder 让 board claim、business case、architecture dependency、risk acceptance 和 management action 能被复核。

Evidence object	Minimum content	Decision use
Baseline fact sheet	volume、cycle time、cost、quality、risk、complaint、manual effort	判断问题是否值得投资
Outcome thesis	target outcome、population、AI role、causal path	判断叙事是否具体
Option comparison	no-AI、process、rules、AI、vendor、platform options	判断是否过早锁定方案
Architecture dependency map	systems、data、RAG、model、tools、workflow、controls、vendors	判断 scale 条件和 hidden cost
Risk appetite mapping	unacceptable outcomes、thresholds、review forum、stop rule	判断风险是否可接受
Eval and QA pack	test set、rubric、failure taxonomy、reviewer evidence	判断 AI 行为质量
Adoption evidence	eligible users、repeat use、acceptance、override reason、manager cadence	判断价值能否进入流程
Unit economics	cost per qualified value event、review cost、support cost	判断 scale 后是否成立
Benefits register	baseline、target、owner、recognition method、confidence	判断收益兑现纪律

Board claim	Required evidence
The use case targets a material problem.	baseline volume、cost、risk and customer impact
AI is the right intervention.	option comparison and no-AI alternative
Value can be realized.	causal logic、adoption、workflow outcome and finance method
Risk is within appetite.	risk tier、controls、eval、incidents、residual risk owner
Architecture can scale.	dependency map、platform capacity、observability、support
Management can stop if needed.	kill criteria、feature flags、rollback、action path

Executive investment decisions need portfolio metrics, not only use case metrics.

Category	Board / executive question	Metric examples
Flow	Is the AI investment funnel healthy?	ideas by stage、WIP、cycle time to evidence、stage aging
Value	Is value proven or estimated?	qualified value events、finance-recognized benefits、forecast vs recognized
Risk	Is residual risk inside appetite?	red/amber appetite breaches、customer harm、incident severity
Evidence	Are decisions evidence-backed?	gates with complete evidence、trace reconstructability、eval coverage
Cost	Are unit economics and platform costs controlled?	cost per value event、review cost、model spend、platform allocation
Adoption	Do users change workflow behavior?	repeat adoption、accepted output、override reason、manager cadence
Architecture	Are investments building reusable capabilities?	platform reuse、duplicate capability count、integration debt、vendor concentration

Management information path:

AI system events
  -> value / risk / adoption / cost metrics
  -> metric contracts and lineage
  -> portfolio dashboard
  -> executive decision pack
  -> action log
  -> benefits realization and risk review

6. 为什么有效, 以及哪里会被误用

这套方法有效, 是因为它改变了 AI 投资的决策单位。 Outcome thesis 把 "AI 能做什么" 改成 "业务为什么值得投资"; option architecture 防止 preferred technology 被当成唯一选择; causal value logic 避免模型指标冒充财务收益; cost-to-learn 用小额资金购买高价值证据; risk appetite translation 把抽象偏好变成边界、阈值和 stop rule; architecture dependency map 把平台、数据、控制、运营和供应商条件前置; evidence binder 让审批、例外、上线和收益兑现可追溯。

在金融零售环境中, 这尤其重要。 AI 不只是优化内部效率; 它可能改变客户资格、资金访问、收费解释、投诉处理、AML 判断、监管报送和员工复核责任。

Anti-pattern	Why it fails	Better practice
Technology-first pitch	Executives fund outcomes, not model enthusiasm	Start with business problem and decision requested
Single-option recommendation	Prevents real tradeoff discussion	Show no-AI、process、rules、AI、vendor and platform options
ROI without causal chain	Savings cannot be trusted	Link AI behavior to workflow、outcome and recognition
Human review as magic control	Review capacity、quality and evidence may fail	Define reviewer coverage、queue、override reason and QA
Pilot success equals scale approval	Pilot scope may not represent production	Separate release and scale gates
Architecture hidden in appendix	Dependencies determine feasibility and cost	Put dependency map in executive narrative
No kill criteria	Pilot becomes sunk-cost project	Write stop / pivot / scale rules before funding
Average metrics only	High-risk segments can be harmed	Segment by product、channel、customer、risk tier and language

常见误用包括把 evidence binder 当成文档仓库、把 risk appetite 写成合规附录、把平台复用当作默认收益却没有 reuse evidence。

7. Financial Retail System Case

7.1 Account Opening Modernization

Decision requested:
Approve a staged investment to modernize AI-assisted account opening evidence review
for mobile retail deposit applications.
Stage 1 funds discovery and shadow-mode pilot.
Scale funding is not requested.

Outcome thesis:
AI-assisted document extraction and checklist generation can reduce manual review cycle time
and application abandonment while preserving KYC controls,
human final decision, audit evidence and customer recourse.

Decision element	Design choice
Baseline	missing-document journeys have high abandonment, inconsistent OCR, manual checklist rework and SLA pressure.
AI role	extract document fields, draft missing-document checklist, retrieve product/KYC policy, summarize case evidence.
Human boundary	final approve / decline, sanctions/fraud escalation and adverse action explanation remain controlled.
Risk appetite	no automated final rejection; no unsupported missing-document demand; strong recourse and evidence retention.
Architecture dependency	OCR abstraction、model gateway、policy RAG、workflow connector、audit record、review queue、metric contract.
Leading indicators	extraction accuracy、checklist correctness、citation quality、reviewer acceptance、queue impact.
Lagging benefits	time-to-open、abandonment、rework、manual review rate、qualified approval cycle time.
Unit economics	model cost + OCR + reviewer time + QA + support compared with qualified completed application.
Stop rule	stop if false missing-document or unsupported rejection recommendation exceeds appetite.
Scale condition	production cohort evidence, trace completeness, finance-reviewed benefit and operations readiness.

7.2 Scenario Variants

Use case	Outcome thesis	AI role	Human boundary	Core evidence	Stop rule
AI complaint intelligence	Reduce systemic complaint detection time and improve corrective action prioritization.	Cluster complaints, retrieve evidence, draft root-cause hypotheses, flag repeat themes.	Regulatory interpretation, remediation decision and customer communication remain authorized human decisions.	repeat complaint rate、action closure time、issue detection lag、QA review quality	stop expansion if high-severity complaint categories are misclassified beyond threshold
AML triage	Reduce low-risk alert handling time and improve narrative completeness.	retrieve context, summarize case, propose checklist, identify evidence gaps.	final closure, escalation and suspicious activity conclusion remain analyst-owned.	AHT、backlog age、QA narrative defect、escalation quality、review capacity	stop if AI omits critical evidence in material samples or causes QA regression
Regulatory reporting automation	Shorten report production cycle and improve lineage reconstructability.	draft variance explanations, map source changes, generate evidence checklist.	attestation, interpretation, filing and material judgment remain authorized roles.	close-cycle time、rework、evidence completeness、issue aging、audit sample pass	stop automation if calculation lineage or reviewer evidence cannot be reconstructed
Branch / contact-center copilot	Improve policy-answer quality and reduce agent research time in bounded service journeys.	retrieve policy, draft response, suggest next-best operational step.	customer-visible communication, fee commitments, complaint handling and advice boundaries remain controlled.	first-contact resolution、AHT、QA fail、reopen、accepted output、override reason	pause expansion if unsupported claim, stale source or high-risk bypass breaches threshold

These cases share the same pattern: AI role bounded, human decision boundary explicit, and scale dependent on value、risk、adoption、unit economics and architecture readiness moving together.

8. 学习验证

学习验证不做角色套话, 只检查能否把一项 AI 投资写成可执行、可挑战、可停止的决策包。

选择 complaint intelligence、account opening modernization、AML triage、regulatory reporting automation 或 branch / contact-center copilot, 产出一个 decision packet。

Artifact	Completion standard
Outcome thesis	写清 population、AI role、causal logic、human boundary 和 learning decision。
Option architecture	至少比较 no-AI、process-only、rules automation、AI-assisted、vendor 或 platform option。
Business case model	包含 baseline、leading indicator、lagging benefit、unit economics、recognition owner 和 confidence level。
Architecture dependency map	覆盖 model、data、RAG、workflow、human review、controls、observability、vendor 和 evidence retention。
Risk appetite mapping	把至少三条 appetite phrase 翻译成 thresholds、unacceptable outcomes、review forum 和 stop rule。
Board narrative	首屏必须说清 decision requested、recommendation、options、conditions、residual risk 和 management action。
Evidence binder index	每个 board claim 都能追到 evidence object、owner、version 和 retention rule。

自检标准: decision clarity、investment logic、benefits discipline、risk discipline、architecture rigor、evidence traceability 和 stop discipline 都必须可见。核心记忆: outcome thesis 说明 why invest; option architecture 说明 choices; causal value logic 说明 how value happens; cost-to-learn 说明 evidence cost; risk appetite 说明 what cannot be traded away; architecture dependency 说明 what must be true to scale; benefits realization 说明 value recognition; gates and evidence traceability 说明 management control。

SOTA 状态标注 (2026-07-01)

本篇属于第二、三遍深读池（参考架构/深读笔记），未列入 12 周主线必读。时效基线为写作时点；引用前请按 CLAUDE.md 全局时效性硬规则复查最新进展。模块级 SOTA 对照见 docs/AI_SYSTEMATIC_LEARNING_ROADMAP_2026.md 各周「2026 SOTA 对照」行与文末「SOTA 检查」。