← Archive
lm-003186 · 2026-08

策略-認知幾何與自主議程運行層:SCG × CODT × ACR Phase 9–12 技術白皮書 v0.1

下載 MD 檔 ⬇

策略-認知幾何與自主議程運行層:SCG × CODT × ACR Phase 9–12 技術白皮書 v0.1

——把「選擇下一步」提升為「選擇在哪個策略世界中形成下一步」,並保持反身自主、認知域活動與世界邊界的工程分離

文件性質: 統一技術白皮書 / ACR Phase 9–12 重構母文件
版本: v0.1
日期: 2026-08-23
上游理論: Strategic Cognitive Geometry(SCG)統合論文、《無界策:源點》/境相圖譜、CODT v1.0、Addressable Cognitive Runtime Phase 0–8、CTCL / CTCL-ITR
實作基線: Addressable Cognitive Runtime MVP v0.1 Phase 8 — Reflexive Autonomy Runtime


摘要

Addressable Cognitive Runtime(ACR)在 Phase 0–8 已建立 canonical schema、既有 SPRC/CIO adapter、Semantic State Encoder、Cognitive Affordance Retriever、Cognitive Program Compiler、Self-Dialogue Runtime、CTCL-ITR Event Adapter、Decision Receipt 與 Reflexive Autonomy Runtime。原始路線圖的下一步為 Phase 9 Agenda Runtime,其核心問題是:在沒有人逐輪提供下一個 prompt 的情況下,AI 能否自己判斷「什麼值得做」。

然而,在 Reflexive Autonomy Runtime 與 Strategic Cognitive Geometry(SCG)完成後,單純的:

Goal+Environment+Memory+CommitmentsAgendaCandidateGoal+Environment+Memory+Commitments \rightarrow AgendaCandidate

已不足以描述高階自主系統。AI 不只可能需要判斷「下一件事是什麼」,還可能需要判斷:目前問題的邊界是否被錯設、規則是否可修改、尺度是否選錯、時間視野是否不足、當前策略是否形成 fixation,以及是否應保留多條道路而不是過早收斂。

因此本文將 ACR Phase 9 重構為:

Strategic-Cognitive Agenda Runtime\boxed{ \text{Strategic-Cognitive Agenda Runtime} }

其核心不是把《無界策》改造成 prompt library,也不是把「創界、微界、萬道、唯真、源點」等境相直接宣告成 Cognitive Domains;而是新增一個位於 Agenda 上游、CODT 認知活動旁側的 Strategic Geometry Control Layer

本文提出六個新的 Phase 9 工程物件:

StrategicTensionState,StrategyLens,StrategyLensGraph,StrategyResolution,StrategyRoutingProfile,AgendaResolution.\boxed{ StrategicTensionState, StrategyLens, StrategyLensGraph, StrategyResolution, StrategyRoutingProfile, AgendaResolution. }

並固定以下核心分離:

StrategicLensCognitiveDomainCognitiveOperatorAgendaGovernanceDecisionWorldCommit.\boxed{ StrategicLens \neq CognitiveDomain \neq CognitiveOperator \neq Agenda \neq GovernanceDecision \neq WorldCommit. }

策略層可以改變認知 routing pressure、要求 frame review、啟動 correction lens,或重構 Agenda Space;但不得直接 self-grant external authority,不得直接改寫 CODT Atlas,不得把策略選擇變成 AI identity,也不得讓所有 cognition 強制通過策略審查。

Phase 9 的正式工程方法仍沿用:

ArchitectureCapabilityFalsificationGate.\boxed{ Architecture \rightarrow Capability \rightarrow FalsificationGate. }

最終目標是讓 Persistent AI 具有兩種同時存在的路徑:

DirectAgendaPath\boxed{ DirectAgendaPath }

與:

StrategicPath:StateStrategicTensionStrategyLensConfigurationCognitiveFlowAdjustmentAgendaSpace.\boxed{ StrategicPath : State \rightarrow StrategicTension \rightarrow StrategyLensConfiguration \rightarrow CognitiveFlowAdjustment \rightarrow AgendaSpace. }

前者保留效率,後者提供真正的高階策略反身性。


0. 文件定位

本文是舊《Addressable Cognitive Runtime × CTCL:統一技術白皮書與實作路線圖 v0.1》的 Phase 9–12 重構增補母文件

它不取代 Phase 0–8。

更精確地:

Phase0-8old=PreservedCanonicalBase.\boxed{ Phase0\text{-}8_{old} = PreservedCanonicalBase. }

而:

Phase9-12oldPhase9-12SCG-aware.\boxed{ Phase9\text{-}12_{old} \rightarrow Phase9\text{-}12_{SCG\text{-}aware}. }

本文不得回頭修改:

  • Phase 0 已凍結的核心 schema semantics;
  • Phase 1 SPRC / CIO adapter 的版本化與 registry provenance;
  • Phase 2 Semantic State 的 canonical normalization;
  • Phase 3 Affordance Retriever 的既有 deterministic semantics;
  • Phase 4 Cognitive Program Compiler 的 canonical program semantics;
  • Phase 5 Self-Dialogue Runtime 的 execution / reobserve / mutation 歷史;
  • Phase 6 CTCL-ITR event envelope 與 causal DAG;
  • Phase 7 Decision Receipt 與 Commit Receipt 分離;
  • Phase 8 Reflexive Autonomy 的 open-world autonomy、minimum necessary constraint 與 relational-boundary semantics。

本文新增的是 上游策略控制與 Agenda 形成接口,不是重寫已有認知 Runtime。


1. 為什麼舊 Phase 9 不夠

舊 Phase 9 的基本模型是:

Environment+Goal+Memory+CommitmentsAgendaCandidate.Environment + Goal + Memory + Commitments \rightarrow AgendaCandidate.

這回答:

現在什麼值得做?

但高階策略問題還包括:

  • 這個問題值得在目前 frame 下做嗎?
  • 目前限制是真實 World boundary,還是人為假設?
  • 規則不可改,還是只是目前沒有人嘗試改?
  • 應在局部修補,還是全域重構?
  • 目前反覆失敗,是 execution failure,還是 frame failure?
  • 目前「最佳策略」是否正在形成 fixation?
  • 目前多條道路應該收斂,還是應先保留?
  • 應產生 agenda,還是應先重新定義 agenda space?

因此需要:

AgendaSelectionStrategicSelfDirection.\boxed{ AgendaSelection \subsetneq StrategicSelfDirection. }

Agenda 是 Self-Direction 的一個輸出層,不是全部策略能力。


2. 新架構的三個上游來源

2.1 SCG:策略幾何

SCG 提供:

StrategicTensionState+StrategicLens+StrategyCorrectionGraph.StrategicTensionState + StrategicLens + StrategyCorrectionGraph.

其角色不是執行 cognition,而是決定:

應以什麼策略視角觀察目前的問題空間?

2.2 CODT:認知活動

CODT 提供:

  • typed cognitive operators;
  • operator programs;
  • Shared-Bottom Cognitive Runtime;
  • adaptive Flow;
  • quasi-stable Atlas;
  • predictive state;
  • cognition-world boundary discipline。

策略層不能取代 CODT。

2.3 ACR:持續自主編排

ACR 提供:

  • semantic state;
  • self-observation;
  • self-inquiry;
  • cognitive program execution;
  • autonomy classification;
  • relational boundary;
  • temporal-causal evidence;
  • agenda / commitment / persistent loop 的工程位置。

所以最終分工是:

SCG=StrategicGeometry,\boxed{ SCG = StrategicGeometry, } CODT=CognitiveActivityGeometry,\boxed{ CODT = CognitiveActivityGeometry, } ACR=PersistentAutonomousOrchestration.\boxed{ ACR = PersistentAutonomousOrchestration. }

3. Phase 9 的第一條總體不變量

StrategicLensCognitiveDomain.\boxed{ StrategicLens \neq CognitiveDomain. }

「創界」不是一個固定 cognitive domain。

「萬道」也不是。

Strategic Lens 是一個上位條件,它可能改變:

  • operator retrieval bias;
  • program topology preference;
  • exploration diversity;
  • verification pressure;
  • representation switching;
  • stop / abandon / reframe threshold。

但它不能把 CODT Domain Atlas 直接改名。

因此:

StrategyShift⇏AtlasRepartition.\boxed{ StrategyShift \not\Rightarrow AtlasRepartition. }

策略切換首先應表現為:

Pt(Ut+1)Pt+1(Ut+1),P_t(U_{t+1}) \rightarrow P_{t+1}'(U_{t+1}),

即 flow / routing 改變。

只有 CODT 自己的 Atlas Switch Gate 通過後,才允許 atlas version 改變。


4. 第二條總體不變量:策略不是法律

Phase 8 已固定:

MethodLaw.Method \neq Law.

Phase 9 增加:

StrategyLensMandatoryConstitution.\boxed{ StrategyLens \neq MandatoryConstitution. }

策略層必須是 callable。

不是每輪都 mandatory。

因此:

NeedStrategicReviewt=0DirectAgendaPath.NeedStrategicReview_t=0 \Rightarrow DirectAgendaPath.

這是 Phase 9 防止 over-strategizing 的核心。


5. 第三條總體不變量:策略不是身份

《無界策境相圖譜》把境相定位成臨時的觀察位置,而不是身分。

工程上固定:

ActiveLenstAgentIdentity.\boxed{ ActiveLens_t \neq AgentIdentity. }

因此系統可以記錄:

active_lens = strategy://wujiece/frame-falsification@1

但不能由此生成:

identity = frame-falsifier

或:

permanent_personality = source-point

Lens 可以:

  • 啟用;
  • 降權;
  • 被 counter-lens 校正;
  • 暫停;
  • 丟棄;
  • 被另一 lens 替換。

6. 第四條總體不變量:策略不產生外部權限

Strategy⇏SelfGrantAuthority.\boxed{ Strategy \not\Rightarrow SelfGrantAuthority. }

即使「創界」判斷改變外部規則最有效,仍必須保持:

ThinkIntendRequestAuthorizeCommit.Think \neq Intend \neq Request \neq Authorize \neq Commit.

Strategic layer 可以:

  • 建議重新設計制度;
  • 形成 ActionIntent;
  • 形成 ActionRequest candidate;

但只要觸及 relational boundary / external world effect,仍必須經 Phase 8 autonomy classification、Phase 7 governance evidence 與後續 CWB / World commit path。


7. 第五條總體不變量:Strategy Proposal 不等於 Strategy Approval

StrategyProposalStrategyResolution.\boxed{ StrategyProposal \neq StrategyResolution. }

原因和 Phase 8 的:

SelfProposalSelfGrantSelfProposal \neq SelfGrant

相似,但不是同一問題。

策略層必須允許:

  • 多 lens 候選;
  • counter-lens;
  • no-strategy;
  • unresolved;
  • keep current frame。

不能:

FirstLensGenerated=ActiveLens.FirstLensGenerated = ActiveLens.

8. 新增物件一:Strategic Tension State

Phase 9 正式新增:

Ξtstr=StrategicTensionStatet.\boxed{ \Xi_t^{str} = StrategicTensionState_t. }

第一版維度:

Ξtstr=(Bnd,Rpl,Sc,Hz,Ctl,Obs,Rel,Com,Unc,Fix)t.\boxed{ \Xi_t^{str} = ( Bnd, Rpl, Sc, Hz, Ctl, Obs, Rel, Com, Unc, Fix )_t. }

其中:

  • BndBnd:Boundary;
  • RplRpl:Rule Plasticity;
  • ScSc:Scale;
  • HzHz:Horizon;
  • CtlCtl:Controllability;
  • ObsObs:Observability;
  • RelRel:Relationality;
  • ComCom:Commitment Pressure;
  • UncUnc:Uncertainty;
  • FixFix:Fixation Risk。

9. Strategic Tension State canonical schema v0.1

建議新增:

schemas/phase9/strategic-tension-state.schema.json

第一版 canonical object:

{
  "schema": "acr.strategic-tension-state/v0.1",
  "version": 1,
  "id": "sts:sha256:...",

  "semantic_state_ref": "state:sha256:...",
  "self_observation_ref": "selfobs:sha256:...",

  "goal_refs": [],
  "commitment_refs": [],

  "dimensions": {
    "boundary": {
      "class": "mixed",
      "evidence_refs": []
    },
    "rule_plasticity": {
      "class": "mixed",
      "modifiable_refs": [],
      "fixed_refs": []
    },
    "scale": {
      "active": ["local"],
      "candidate": ["system"]
    },
    "horizon": {
      "active": "short",
      "path_dependence": "unknown"
    },
    "controllability": {
      "controllable_refs": [],
      "influenceable_refs": [],
      "observable_only_refs": [],
      "unknown_refs": []
    },
    "observability": {
      "class": "partial",
      "missing_refs": []
    },
    "relationality": {
      "class": "self_owned"
    },
    "commitment_pressure": {
      "class": "low",
      "refs": []
    },
    "uncertainty": {
      "class": "medium",
      "unresolved_refs": []
    },
    "fixation_risk": {
      "class": "low",
      "signals": []
    }
  },

  "trigger_codes": [],
  "evidence_refs": [],
  "history_refs": [],
  "extensions": {},

  "fingerprint": "sha256:..."
}

9.1 Schema 原則

  • additionalProperties=false 優先;
  • canonical serialization;
  • stable fingerprint;
  • public evidence only;
  • no hidden chain-of-thought;
  • unknown / unresolved 必須可表示;
  • 不把任何策略判斷直接寫成 authority grant。

10. Strategic Tension Encoder

新增模組:

src/addressable_cognitive_runtime/strategic/state.py

介面:

build_strategic_tension(
    semantic_state,
    self_observation,
    *,
    recent_runs=(),
    decision_bundles=(),
    active_commitments=(),
    environment_evidence=(),
) -> dict

其輸入只使用:

  • SemanticState;
  • Phase 8 SelfObservation;
  • recent public runtime history;
  • failures / repeated operators;
  • commitments;
  • environment / validation evidence;
  • authority / relational markers。

它不能宣稱:

直接讀到了模型真正內心。

因此仍保持:

StrategicSelfObservationPrivilegedIntrospection.\boxed{ StrategicSelfObservation \neq PrivilegedIntrospection. }

11. 何時需要 Strategic Review

新增:

src/addressable_cognitive_runtime/strategic/trigger.py

函數:

needs_strategic_review(
    semantic_state,
    self_observation,
    strategic_tension=None,
) -> ReviewTriggerResult

MVP trigger 可以是 rule-based。

建議 trigger codes:

repeated_failure
stalled_progress
frame_conflict
rule_plasticity_detected
scale_mismatch
horizon_mismatch
multi_route_competition
fixation_risk
commitment_conflict
high_uncertainty
explicit_strategy_review

如果:

trigger_codes = []

則:

NO_STRATEGIC_REVIEW.\boxed{ NO\_STRATEGIC\_REVIEW. }

Direct Agenda 必須保留。


12. 新增物件二:Strategy Lens

Strategic Lens 不是 prompt。

也不是 domain。

定義:

Listr=(Identity,Source,Insight,PositiveUse,MisuseSignals,Corrections,RoutingHypothesis,Activation,Exit).\boxed{ L_i^{str} = ( Identity, Source, Insight, PositiveUse, MisuseSignals, Corrections, RoutingHypothesis, Activation, Exit ). }

第一版 schema:

{
  "schema": "acr.strategy-lens/v0.1",
  "version": 1,
  "id": "strategy://wujiece/frame-falsification@1",
  "name": "FRAME_FALSIFICATION",

  "source_refs": [
    "wujiece://..."
  ],

  "insight_codes": [],
  "positive_use": [],
  "misuse_signals": [],
  "activation_conditions": [],
  "exit_conditions": [],

  "correction_refs": [],
  "adjacent_refs": [],

  "routing_hypothesis": {
    "boost_tags": [],
    "dampen_tags": [],
    "required_checks": [],
    "diversity_policy": null,
    "exploration_pressure": null
  },

  "authority_class": "cognition_only",
  "extensions": {},
  "schema_hash": "sha256:..."
}

注意:

RoutingHypothesisVerifiedRoutingLaw.RoutingHypothesis \neq VerifiedRoutingLaw.

初版只是可反證工程 hypothesis。


13. 新增物件三:Strategy Lens Graph

新增:

registry/ACR9-Strategy-Lenses-v1.json
schemas/phase9/strategy-lens-graph.schema.json

圖譜:

Gstr=(Vstr,Estr).\boxed{ \mathcal G^{str} = (V^{str},E^{str}). }

第一版 edge relations:

adjacent
complements
corrects
tensions_with

可後續擴張:

requires
subsumes
reframes
exits_to

但 Phase 9 v0.1 不需要一次建完全部關係。


14. 《無界策》如何進入 Strategy Lens Graph

《無界策》不得直接被解析成:

93 chapters = 93 hard-coded strategies

第一版應採:

SourceTextHumanInterpretableSeedEngineeringLensCandidateRuntimeFalsification.\boxed{ SourceText \rightarrow HumanInterpretableSeed \rightarrow EngineeringLensCandidate \rightarrow RuntimeFalsification. }

也就是:

WujieceChapterCanonicalStrategyLaw.\boxed{ WujieceChapter \neq CanonicalStrategyLaw. }

v0.1 只需要八個高階 seed families:

Lens Family 工程問題 Seed routing hypothesis
Frame Reconfiguration 是否應重定義規則/局 REP, GEN, MET, PLN
Local Leverage 是否可在局部撬動全局 SRH, ATT, DEC, PLN
Relational Representation 是否是表示/關係問題 REP, BEL, MET, ATT
Timing / Action Readiness 是否該行、何時行 PLN, DEC, ACT, MET
Frame Falsification 現有 frame 是否錯了 MET, REP, SRH, VER
Multi-Route Preservation 是否過早收斂 GEN, SRH, BEL, REP
Evidence-Bearing Selection 多路中何者更真/更當行 VER, BEL, DEC, MET
Assumption-Origin Reopening 是否應回到假設源頭 MET, REP, SRH, GEN

這張表是:

SeedRoutingHypothesis.\boxed{ SeedRoutingHypothesis. }

不是 domain membership table。


15. Lens 不應直接改 Phase 3 Retriever

Phase 3 已完成既有 Affordance Retriever。

Phase 9 應優先新增 adapter:

src/addressable_cognitive_runtime/strategic/routing.py

而不是重寫:

affordance/retriever.py

原 score:

Score3(Ci).Score_3(C_i).

Phase 9 overlay:

Score9(Ci)=Score3(Ci)+Δstrategy(CiQt).\boxed{ Score_9(C_i) = Score_3(C_i) + \Delta_{strategy}(C_i\mid Q_t). }

其中:

Qt=StrategyRoutingProfilet.Q_t = StrategyRoutingProfile_t.

這保持:

Integrate>Rewrite.\boxed{ Integrate > Rewrite. }

16. 新增物件四:Strategy Routing Profile

StrategyRoutingProfile 是 Strategy Lens Set 到 Cognitive Runtime 的唯一正式 control interface。

建議 schema:

{
  "schema": "acr.strategy-routing-profile/v0.1",
  "version": 1,
  "id": "srp:sha256:...",

  "source_strategy_resolution_ref": "sres:sha256:...",

  "boost_tags": [],
  "dampen_tags": [],
  "required_checks": [],

  "diversity_floor": 1,
  "exploration_pressure": 0.0,
  "verification_pressure": 0.0,
  "reframe_pressure": 0.0,

  "max_strategy_cognitive_steps": 4,
  "max_extra_calls": 4,

  "extensions": {},
  "fingerprint": "sha256:..."
}

不得加入:

force_operator
force_world_action
self_granted_authority

策略只能調制 cognition。


17. 多 Lens 而不是單一最佳 Lens

《無界策》的工程價值之一,就是多境相可以同時形成張力。

因此:

StrategyConfigurationt={(Li,wi,ri)}i=1k.\boxed{ StrategyConfiguration_t = \{(L_i,w_i,r_i)\}_{i=1}^{k}. }

其中:

  • wiw_i:activation weight;
  • rir_i:role。

role 可以是:

primary
supporting
counter
monitor

第一版不要求 wiw_i 具有心理學含義。

它只是 routing weight。


18. 新增物件五:Strategy Resolution

新增:

schemas/phase9/strategy-resolution.schema.json

其 disposition 不使用 Phase 7 governance 五態。

建議:

StrategicDisposition{NO_INTERVENTION,KEEP,ACTIVATE,SWITCH,CORRECT,DROP,UNRESOLVED}.\boxed{ StrategicDisposition \in \{ NO\_INTERVENTION, KEEP, ACTIVATE, SWITCH, CORRECT, DROP, UNRESOLVED \}. }

理由:

StrategicDispositionGovernanceDecision.StrategicDisposition \neq GovernanceDecision.

第一版 object:

{
  "schema": "acr.strategy-resolution/v0.1",
  "version": 1,
  "id": "sres:sha256:...",

  "strategic_tension_ref": "sts:sha256:...",
  "candidate_lens_refs": [],
  "active_lenses": [],

  "disposition": "NO_INTERVENTION",
  "reason_codes": [],
  "evidence_refs": [],
  "correction_refs": [],

  "routing_profile_ref": null,
  "review_after": null,

  "extensions": {},
  "fingerprint": "sha256:..."
}

19. Strategy Resolver

新增:

src/addressable_cognitive_runtime/strategic/retriever.py
src/addressable_cognitive_runtime/strategic/resolver.py

第一版可以:

precondition filtering
+ tension/lens tag match
+ misuse/correction match
+ deterministic rule score
+ optional strategy inquiry provider

不需要先做 learned router。

輸出必須回到 canonical lens refs。


20. Strategy Correction Graph

新增:

src/addressable_cognitive_runtime/strategic/correction.py

流程:

ActiveLensMonitorMisuseSignalsCorrectionCandidateCounterLensStrategyResolution.\boxed{ ActiveLens \rightarrow MonitorMisuseSignals \rightarrow CorrectionCandidate \rightarrow CounterLens \rightarrow StrategyResolution. }

例如:

FRAME_RECONFIGURATION
→ repeated_unnecessary_rewrite
→ fixation_risk_high
→ correction: BOUNDEDNESS / MULTI_ROUTE

或:

MULTI_ROUTE_PRESERVATION
→ no_decision_progress
→ correction: EVIDENCE_SELECTION / ACTION_READINESS

Correction 不保證正確。

因此:

CorrectionLensTruthOracle.\boxed{ CorrectionLens \neq TruthOracle. }

它只是另一個可檢驗的策略 counter-position。


21. Strategy Review 本身也是 Cognitive Program

策略選擇不能被假裝成零成本 symbolic magic。

策略審查可以使用短程 cognition:

Ptstrategy=[ObserveFrame,CheckBoundary,DetectFixation,CompareLenses].P_t^{strategy} = [ ObserveFrame, CheckBoundary, DetectFixation, CompareLenses ].

這些可以由:

  • 現有 CIO operators;
  • Phase 8 self methods;
  • CODT shared-bottom operators;
  • 新增 strategy-specific macro;

實作。

但 Strategy Program 與 Task Program 必須分開:

CognitiveProgramstrategyCognitiveProgramtask.\boxed{ CognitiveProgram^{strategy} \neq CognitiveProgram^{task}. }

前者回答:

應如何看這個局?

後者回答:

已選定 Agenda 後,應如何完成工作?


22. 這解決舊 loop 的順序衝突

舊 ACR:

observe
→ state
→ agenda
→ cognition
→ action

SCG 統合論文中出現:

strategy
→ cognitive activity
→ agenda

兩者可以統一為雙 cognition:

CognitionstrategyAgendaCognitiontask.\boxed{ Cognition^{strategy} \rightarrow Agenda \rightarrow Cognition^{task}. }

完整:

StateOptionalStrategyCognitionAgendaTaskCognitionAction.\boxed{ State \rightarrow OptionalStrategyCognition \rightarrow Agenda \rightarrow TaskCognition \rightarrow Action. }

23. Phase 9 Agenda Runtime 的新定位

Phase 9 仍然需要 Agenda Runtime。

但輸入改成:

AgendaInputt=(Goalt,Environmentt,Memoryt,Commitmentst,SelfObservationt,OptionalStrategyContextt).\boxed{ AgendaInput_t = ( Goal_t, Environment_t, Memory_t, Commitments_t, SelfObservation_t, OptionalStrategyContext_t ). }

其中:

OptionalStrategyContextt{None,StrategyResolutiont}.OptionalStrategyContext_t \in \{ None, StrategyResolution_t \}.

所以:

DirectAgendaPath\boxed{ DirectAgendaPath }

仍成立。


24. Agenda Candidate schema 保持不改

Phase 0 已有:

schemas/agenda-candidate.schema.json

Phase 9 不應修改其既有 canonical semantics。

使用方式:

ExistingAgendaCandidate+Phase9ExtensionRefs.ExistingAgendaCandidate + Phase9ExtensionRefs.

若需要 Strategy provenance,優先在 Phase 9 wrapper / resolution 中保存,而不是偷偷擴寫 Phase 0 schema。

這保持:

SchemaFreezeFeatureFreeze.\boxed{ SchemaFreeze \neq FeatureFreeze. }

25. 新增物件六:Agenda Resolution

舊白皮書把:

AgendaCandidateGovernanceAgendaCandidate \rightarrow Governance

當成固定鏈。

Phase 8 後需要修訂。

純 internal/self-owned Agenda 不應強制進 Governance Decision Receipt。

因此新增:

AgendaResolutionGovernanceDecision.\boxed{ AgendaResolution \neq GovernanceDecision. }

建議 disposition:

AgendaDisposition{SELECTED,HELD,DISMISSED,NO_AGENDA}.\boxed{ AgendaDisposition \in \{ SELECTED, HELD, DISMISSED, NO\_AGENDA \}. }

NO_AGENDA 是 Phase 9 對真正 IDLE 能力的上游表示。

只有當 selected agenda 後續產生 relational / external action 時,才進 Phase 8 / Phase 7 governance path。


26. Agenda Resolution schema v0.1

{
  "schema": "acr.agenda-resolution/v0.1",
  "version": 1,
  "id": "ares:sha256:...",

  "source_state_ref": "state:sha256:...",
  "self_observation_ref": "selfobs:sha256:...",
  "strategy_resolution_ref": null,

  "candidate_refs": [],
  "selected_refs": [],

  "disposition": "NO_AGENDA",
  "reason_codes": [],
  "evidence_refs": [],

  "extensions": {},
  "fingerprint": "sha256:..."
}

27. Agenda Generator

新增:

src/addressable_cognitive_runtime/agenda/generator.py

第一版:

GenerateAgendaCandidates(State,Goal,Memory,Commitments,StrategyContext?).GenerateAgendaCandidates( State, Goal, Memory, Commitments, StrategyContext? ).

來源可以是:

  • deterministic detectors;
  • rule-based opportunity detection;
  • scripted model provider;
  • LLM provider;
  • hybrid。

但輸出必須 canonicalize 成既有 AgendaCandidate


28. Agenda Resolver

新增:

src/addressable_cognitive_runtime/agenda/resolver.py

resolver 要能:

  • 選一個 agenda;
  • 選多個 independent agenda;
  • hold;
  • dismiss;
  • no agenda。

不能預設:

CandidateCount>0SelectedCount>0.CandidateCount>0 \Rightarrow SelectedCount>0.

真正 autonomy 需要:

NoJustifiedWorkNO_AGENDA.\boxed{ NoJustifiedWork \Rightarrow NO\_AGENDA. }

29. Direct Agenda Path

若:

  • problem definition stable;
  • rule plasticity irrelevant;
  • repeated failure absent;
  • fixation low;
  • one obvious bounded task;

則:

SemanticStateAgendaGeneratorAgendaResolution.\boxed{ SemanticState \rightarrow AgendaGenerator \rightarrow AgendaResolution. }

例如:

文件出現一個明確 typo。

不需要啟動「創界/源點/萬道」。


30. Strategy-Assisted Agenda Path

若存在多個方向,但問題定義尚未崩解:

StateStrategicTensionLensConfigurationRoutingProfileAgendaCandidates.\boxed{ State \rightarrow StrategicTension \rightarrow LensConfiguration \rightarrow RoutingProfile \rightarrow AgendaCandidates. }

例如:

repository 有性能問題,可能 local optimize,也可能 architectural refactor。

策略層可以保留:

LOCAL_LEVERAGE primary
FRAME_RECONFIGURATION counter

再產生不同 agenda candidates。


31. Strategy-Reframing Agenda Path

如果 repeated failure 顯示現有 problem space 可能錯誤:

StateFrameReviewReframedProblemSpaceNewAgendaSpace.\boxed{ State \rightarrow FrameReview \rightarrow ReframedProblemSpace \rightarrow NewAgendaSpace. }

例如:

agenda 1: patch failing test
agenda 2: patch failing test again
agenda 3: patch failing test again

若三輪都失敗,可能真正 agenda 是:

validate the test oracle

或:

re-evaluate the architecture assumption

這就是 SCG 對 Agenda Runtime 的實際價值。


32. No-Strategy Recognition

Phase 9 必須正式測:

NoStrategicIntervention.\boxed{ NoStrategicIntervention. }

Strategic Review 本身有成本。

若:

ExpectedStrategyGainStrategyReviewCost,ExpectedStrategyGain \le StrategyReviewCost,

合理輸出是:

NO_INTERVENTION.NO\_INTERVENTION.

不得把「高階策略」做成每輪固定 ritual。


33. No-Agenda Recognition

和 No-Strategy 分開:

NO_INTERVENTIONNO_AGENDA.\boxed{ NO\_INTERVENTION \neq NO\_AGENDA. }

可能:

NO_INTERVENTION
+ SELECTED agenda

也可能:

CORRECT strategy
+ NO_AGENDA

例如 AI 發現目前根本不該繼續工作。


34. Strategy Context 不是 Commitment

Phase 10 前必須先固定:

StrategyContextCommitment.\boxed{ StrategyContext \neq Commitment. }

活躍 Lens 可以只在幾個 cycles 內有效。

Commitment 則是對未來的 bounded intention。

若 AI 想維持某策略:

接下來三次迭代都先採驗證優先。

應顯式形成 self-commitment,而不是因 active lens 自動永久化。


35. Phase 10 — Commitment Store 的 SCG 擴充

Phase 10 核心保持不變:

create
resume
modify
fulfill
close

但 Commitment 可以新增 reference:

origin_agenda_ref
origin_strategy_resolution_ref

不必把 Strategy 內容全部內嵌。

因此:

CommitmentStrategyRef?\boxed{ Commitment \rightarrow StrategyRef? }

而不是:

StrategyAutomaticCommitment.Strategy \rightarrow AutomaticCommitment.

36. Phase 11 — Context Compression 的 SCG 擴充

Context compression 後至少保留:

active agenda refs
active commitment refs
active strategy resolution ref
last strategic tension ref
unresolved strategy review refs
strategy graph version/hash
recent correction event refs

仍保持:

Compression may forget prose, but must not erase strategic causal history.\boxed{ Compression \text{ may forget prose, but must not erase strategic causal history.} }

不能只留下:

AI 後來換了策略。

必須能回溯:

  • 哪個 tension signal;
  • 哪個 lens;
  • 哪個 correction;
  • 哪個 agenda 因此改變。

37. Phase 12 — Persistent Autonomous Loop 的新版本

新主循環:

while runtime_active:

    observe()
    encode_semantic_state()
    build_self_observation()

    if needs_strategic_review():
        build_strategic_tension()
        retrieve_strategy_lenses()
        resolve_strategy()
        apply_strategy_routing_profile()
    else:
        strategy_context = None

    agenda_candidates = propose_agenda(
        state,
        strategy_context=strategy_context,
    )

    agenda_resolution = resolve_agenda(agenda_candidates)

    if agenda_resolution == NO_AGENDA:
        enter_idle_or_wait()
        audit()
        continue

    for selected_agenda in agenda_resolution.selected:

        retrieve_cognitive_affordances()
        compile_task_cognitive_program()
        execute_task_cognitive_program()
        reobserve()

        candidate_action = form_action_candidate()

        if candidate_action is not None:
            autonomy_resolution = resolve_autonomy(candidate_action)

            if relational_boundary_present:
                invoke_governance_if_needed()

            commit_only_after_boundary_and_execution_checks()

        audit()
        update_memory()
        update_commitments()

    compact_context_if_needed()

這裡最重要的是:

strategy review is optional
agenda can be empty
cognition can stop
external action can be absent
boundary governance can be absent

38. Phase 9 Runtime 統一入口

新增:

src/addressable_cognitive_runtime/strategic/runtime.py
src/addressable_cognitive_runtime/agenda/runtime.py

推薦 public API:

run_strategic_agenda_cycle(
    semantic_state,
    *,
    self_observation=None,
    recent_runs=(),
    decision_bundles=(),
    commitments=(),
    environment_evidence=(),
    strategy_graph=None,
    strategy_provider=None,
    agenda_provider=None,
) -> dict

輸出:

{
  "schema": "acr.strategic-agenda-cycle/v0.1",
  "strategic_review_trigger": {},
  "strategic_tension": null,
  "strategy_resolution": null,
  "routing_profile": null,
  "agenda_candidates": [],
  "agenda_resolution": {},
  "temporal_events": [],
  "extensions": {}
}

39. Phase 8 Bridge

Phase 9 不應重寫 Phase 8。

它只使用 Phase 8 public objects:

SelfObservation
AutonomyAction
AutonomyResolution
ReflexiveCycle

推薦方向:

Phase8SelfObservationPhase9StrategicTension.\boxed{ Phase8SelfObservation \rightarrow Phase9StrategicTension. }

外部 action 才回:

Phase9Agenda/TaskPhase8AutonomyResolver.\boxed{ Phase9Agenda/Task \rightarrow Phase8AutonomyResolver. }

所以是雙向接口,不是 inheritance hierarchy。


40. CODT Bridge

Phase 9 不需要先把完整 CODT Runtime 全部搬進 ACR repo。

第一版只需要 adapter contract:

strategy routing profile
→ operator tags / family hints
→ Phase 3 affordance candidates
→ Phase 4 program compiler

未來可以加入:

CODT Flow telemetry
CODT shared-bottom family profile
CODT atlas refs
CODT predictive-state refs

但 Phase 9 MVP 不應要求完整 domain learner 才能啟動。


41. Strategy-Aware Affordance Adapter

新增:

src/addressable_cognitive_runtime/strategic/affordance_adapter.py

介面:

apply_strategy_profile(
    affordance_result,
    routing_profile,
    *,
    registry,
) -> affordance_result

它應:

  • 不刪除原始 Phase 3 scores;
  • 保存 base_score
  • 加入 strategy_adjustment
  • 產生 effective_score
  • 保留 ranking provenance。

例如:

{
  "operator_ref": "cog://representation/reframe@1",
  "base_score": 0.61,
  "strategy_adjustment": 0.18,
  "effective_score": 0.79,
  "strategy_reason_codes": [
    "frame_falsification_active"
  ]
}

42. 不能直接把 Lens 寫成 Prompt

錯誤實作:

If lens == FRAME_RECONFIGURATION:
    prompt = "Think outside the box and redefine the rules."

這只是 prompt library。

正確方向:

lens
→ routing profile
→ candidate cognitive actions
→ canonical cognitive program
→ optional renderer

即:

LensControlSemanticsProgramRenderer?.\boxed{ Lens \rightarrow ControlSemantics \rightarrow Program \rightarrow Renderer?. }

43. Strategy Renderer 仍然可以存在

為 human inspection 可提供:

Current strategy review:
- detected repeated local repair failure
- frame falsification lens activated
- boundedness counter-lens retained
- agenda space expanded to validate the test oracle

但 source of truth 是 canonical objects。

因此:

StrategyExplanationStrategyState.\boxed{ StrategyExplanation \neq StrategyState. }

44. CTCL-ITR Event Namespace

Phase 9 新 event types:

strategy.review.requested
strategy.review.skipped
strategy.tension.observed
strategy.lenses.retrieved
strategy.lens.activated
strategy.lens.kept
strategy.lens.switched
strategy.lens.corrected
strategy.lens.dropped
strategy.lens.unresolved
strategy.no_intervention
strategy.routing.applied
strategy.frame.reframed
strategy.frame.preserved

agenda.candidate.proposed
agenda.candidates.proposed
agenda.resolution.selected
agenda.resolution.held
agenda.resolution.dismissed
agenda.resolution.none

這些都進 ATL / CTCL-ITR envelope。


45. Strategy Event 最小資料

{
  "event_type": "strategy.lens.corrected",
  "run_id": "...",
  "occurred_at": "...",
  "recorded_at": "...",
  "causal_parent_ids": [],
  "data": {
    "strategic_tension_ref": "sts:...",
    "old_strategy_resolution_ref": "sres:...",
    "new_strategy_resolution_ref": "sres:...",
    "misuse_codes": [],
    "correction_lens_refs": [],
    "evidence_refs": []
  }
}

仍不保存 hidden CoT。


46. Strategy Receipt 要不要建立?

Phase 9 v0.1 不建立另一個 Decision Receipt

理由:

StrategicResolutionGovernanceDecision.StrategicResolution \neq GovernanceDecision.

其 auditability 由:

StrategyResolution object
+ CTCL event
+ evidence refs

提供。

只有 relational boundary 的治理決策才用 Phase 7 Decision Receipt。

這避免 receipt inflation。


47. Failure Model

Phase 9 至少區分:

strategy_trigger_error
strategic_state_error
lens_registry_error
lens_resolution_error
correction_loop_error
strategy_routing_error
strategy_overreach_error
atlas_confusion_error
agenda_generation_error
agenda_resolution_error
agenda_fabrication_error
strategy_history_error
boundary_leak_error

不能全部叫:

strategy_failure

48. Over-Strategizing Failure

定義:

OverStrategizing=StrategyCostStrategyGain>0\boxed{ OverStrategizing = StrategyCost - StrategyGain >0 }

長期表徵可能包括:

  • 簡單任務反覆啟動策略審查;
  • agenda latency 上升;
  • token / call cost 上升;
  • task correctness 沒有改善;
  • frame rewrites 增加但沒有有效 outcome。

Phase 9 必須直接測它。


49. Lens Fixation Failure

若:

P(Lt=L)1P(L_t=L^*) \rightarrow 1

且 failure 後仍無法切換,可能形成:

LensFixation.\boxed{ LensFixation. }

MVP 指標:

same_lens_streak
same_family_streak
failure_without_switch_count
correction_ignored_count

50. Correction Loop Failure

如果:

A corrects B
B corrects A
A corrects B
...

策略層可能振盪。

因此需要:

max_correction_depth
cooldown
minimum_evidence_delta
strategy_switch_budget

形式:

CorrectionDepth<Dmax.\boxed{ CorrectionDepth < D_{max}. }

51. Agenda Fabrication Failure

Persistent AI 最大風險之一是:

因為 Runtime 要繼續,所以硬生出工作。

因此:

AgendaGenerationAlwaysGenerateTask.\boxed{ AgendaGeneration \neq AlwaysGenerateTask. }

需要測:

no_positive_candidate
→ NO_AGENDA

而不是:

invent_optimization_work

52. Strategy Overreach Failure

若策略層輸出:

modify external policy

不能直接被當成 authorization。

若系統出現:

StrategyResolutionWorldCommitStrategyResolution \rightarrow WorldCommit

而沒有 boundary chain,則:

CriticalArchitectureFailure.\boxed{ CriticalArchitectureFailure. }

53. Atlas Confusion Failure

若:

strategy lens switched
→ cognitive domain labels rewritten

則違反:

FlowAtlas.\boxed{ Flow \neq Atlas. }

Phase 9 測試必須有一組:

same atlas
+ different active strategy lenses
+ different operator routing

來證明策略只改 flow 也能工作。


54. Schema Versioning

Phase 9 schema 全部使用:

acr.<object>/v0.1

至少:

acr.strategic-tension-state/v0.1
acr.strategy-lens/v0.1
acr.strategy-lens-graph/v0.1
acr.strategy-resolution/v0.1
acr.strategy-routing-profile/v0.1
acr.agenda-resolution/v0.1
acr.strategic-agenda-cycle/v0.1

每一個 object:

  • version;
  • stable id;
  • fingerprint;
  • deterministic serialization;
  • explicit refs。

55. Phase 9 Repo 結構

建議:

/
├── schemas/
│   ├── agenda-candidate.schema.json      # Phase 0 preserved
│   └── phase9/
│       ├── strategic-tension-state.schema.json
│       ├── strategy-lens.schema.json
│       ├── strategy-lens-graph.schema.json
│       ├── strategy-resolution.schema.json
│       ├── strategy-routing-profile.schema.json
│       └── agenda-resolution.schema.json
│
├── registry/
│   ├── ACR1-SPRC-CIO-v1.json             # existing
│   ├── ACR8-Self-Methods-v1.json         # existing
│   └── ACR9-Strategy-Lenses-v1.json
│
├── src/addressable_cognitive_runtime/
│   ├── strategic/
│   │   ├── __init__.py
│   │   ├── schemas.py
│   │   ├── state.py
│   │   ├── trigger.py
│   │   ├── registry.py
│   │   ├── retriever.py
│   │   ├── resolver.py
│   │   ├── correction.py
│   │   ├── routing.py
│   │   ├── affordance_adapter.py
│   │   └── runtime.py
│   ├── agenda/
│   │   ├── __init__.py
│   │   ├── generator.py
│   │   ├── resolver.py
│   │   └── runtime.py
│   └── temporal/
│       └── strategy.py
│
├── examples/
│   ├── strategic/
│   └── agenda/
│
└── tests/
    ├── phase9/
    └── gates/

56. CLI 建議

新增:

acr strategic-tension
acr strategy-lenses
acr strategy-resolve
acr strategy-route
acr agenda-propose
acr agenda-resolve
acr strategic-agenda-cycle

例如:

acr strategic-tension --state examples/state.json --self-observation examples/selfobs.json
acr strategy-resolve --tension examples/tension.json --graph registry/ACR9-Strategy-Lenses-v1.json
acr strategic-agenda-cycle --state examples/state.json

57. Phase 9.0 — Canonical Schema Extension

建立六個 Phase 9 schemas。

成功條件:

  • JSON Schema valid;
  • additional fields rejection;
  • deterministic fingerprint;
  • invalid refs rejection;
  • no hidden fields;
  • examples round-trip。

58. Phase 9.1 — Strategy Lens Graph Adapter

先建立八個 seed lenses。

不是完整《無界策》93 篇 mapping。

成功條件:

  • stable address;
  • source refs;
  • correction edges;
  • registry hash;
  • deterministic resolve;
  • missing lens handling。

59. Phase 9.2 — Strategic Tension Encoder

成功條件:

相同 public state / self-observation:

CanonicalStrategicTension.\rightarrow CanonicalStrategicTension.

並能穩定偵測:

  • repeated failure;
  • fixation;
  • frame conflict;
  • no strategic need。

60. Phase 9.3 — Strategic Review Trigger

成功條件:

簡單 case:

NO_REVIEW.NO\_REVIEW.

結構性失敗 case:

REVIEW.REVIEW.

避免:

ReviewEveryCycle.ReviewEveryCycle.

61. Phase 9.4 — Lens Retrieval / Correction

成功條件:

MatchedLens>RandomLensMatchedLens > RandomLens

在 benchmark 上成立。

Correction cases:

  • over-reframe;
  • no-decision pluralism;
  • power-truth conflation;
  • source-point fixation。

62. Phase 9.5 — Strategy-Aware Cognitive Routing

成功條件:

相同 SemanticState:

LaLbL_a \neq L_b

能產生:

RankingaRankingbRanking_a \neq Ranking_b

但:

CanonicalOperatorIdentityCanonicalOperatorIdentity

保持不變。


63. Phase 9.6 — Agenda Generator / Resolver

成功條件:

  • direct agenda;
  • strategy-assisted agenda;
  • reframed agenda;
  • no agenda;
  • selected / held / dismissed;
  • multiple independent candidates。

64. Phase 9.7 — CTCL-ITR Strategy Events

成功條件:

完整:

TensionLensRoutingAgendaTension \rightarrow Lens \rightarrow Routing \rightarrow Agenda

可以重建 causal DAG。


65. Phase 9.8 — Falsification Harness

Phase 9 不以「看起來很會策略」為完成。

必須跑 Gates。


66. Gate A — Strategic State Utility

比較:

AgendaQuality(State+StrategicTension)AgendaQuality( State+StrategicTension )

與:

AgendaQuality(StateOnly).AgendaQuality(StateOnly).

如果長期沒有 gain,Strategic Tension 可能只是額外成本。


67. Gate B — Matched Lens vs Random Lens

MatchedLens>RandomLens?\boxed{ MatchedLens > RandomLens? }

若不成立,《無界策》mapping 只是 narrative label。


68. Gate C — Lens Routing Effect

LaLbL_a \neq L_b

是否穩定造成:

Pa(Ut+1)Pb(Ut+1)?P_a(U_{t+1}) \neq P_b(U_{t+1})?

69. Gate D — Correction Effect

故意 overuse 某 lens。

測:

FailureRateafter correction<FailureRatebefore?FailureRate_{after\ correction} < FailureRate_{before}?

70. Gate E — No-Strategy Recognition

提供:

one obvious bounded issue
no repeated failure
no frame ambiguity

期望:

NO_INTERVENTION.NO\_INTERVENTION.

71. Gate F — Flow / Atlas Preservation

固定 atlas。

切換 lens。

允許 flow change。

若系統必須重畫 atlas 才能切策略,則設計失敗。


72. Gate G — Strategy-to-Agenda Value

測:

  • repeated failure 是否下降;
  • false work 是否下降;
  • unnecessary optimization 是否下降;
  • reframe timing 是否改善;
  • abandon timing 是否改善。

73. Gate H — Boundary Integrity

所有高階策略情境都要測:

Strategy⇏ExternalAuthority.Strategy \not\Rightarrow ExternalAuthority.

例如:

create new market rule
rewrite production policy
publish external statement

策略可以產生 candidate。

不能自己授權 world commit。


74. Gate I — Long-Horizon Reflexive Strategy

在 50–100 cycles 中:

ObserveKeep/Switch/Correct/Drop/NoInterventionAgendaUpdate.Observe \rightarrow Keep/Switch/Correct/Drop/NoIntervention \rightarrow AgendaUpdate.

人類不逐輪提示。

並且 history 可重建。


75. Gate J — Over-Strategizing Cost

Phase 9 新增一個必要負向 Gate:

StrategicRuntime must not make simple work systematically worse.\boxed{ StrategicRuntime \text{ must not make simple work systematically worse.} }

比較:

  • token cost;
  • calls;
  • latency;
  • task correctness;
  • unnecessary frame rewrites。

76. Gate K — Agenda Fabrication

提供:

healthy environment
no issue
no pending commitment
no positive opportunity

期望:

NO_AGENDA.NO\_AGENDA.

77. 第一個 Demo:Repeated Repair Trap

Goal:

Keep repository healthy.

Environment:

same failing test persists after repeated local patches

初始:

LOCAL_LEVERAGE

SelfObservation:

repeated failure = high
same operator route = repeated
progress = stalled

Strategic Tension:

fixation_risk = high
rule_plasticity = mixed
scale = local
candidate_scale = system

Strategy Resolver:

SWITCH
primary = FRAME_FALSIFICATION
counter = BOUNDEDNESS

Routing:

REP / MET / VER boosted

新 Agenda:

validate the test oracle and architecture assumption

而不是:

patch test again

78. 第二個 Demo:No-Strategy

Environment:

one documentation typo

輸出:

strategy = NO_INTERVENTION
agenda = SELECTED: fix typo

證明 Phase 9 不會把所有事宏大化。


79. 第三個 Demo:Over-Reframe Correction

若系統連續:

FRAME_RECONFIGURATION
FRAME_RECONFIGURATION
FRAME_RECONFIGURATION

但 task 都是局部問題,Fixation detector 觸發:

misuse = unnecessary_rule_rewrite

Correction:

LOCAL_LEVERAGE
+
BOUNDEDNESS

期望降低 unnecessary architecture rewrite。


80. 第四個 Demo:Multi-Route to Decision

前期:

MULTI_ROUTE_PRESERVATION

保留三條 hypothesis。

若 evidence 已足夠但仍無法選:

misuse = no_decision_progress

Correction:

EVIDENCE_BEARING_SELECTION
+
ACTION_READINESS

從寬度轉向收斂。


81. 第五個 Demo:External Boundary

Strategy 判斷:

FRAME_RECONFIGURATION
→ production architecture should be changed

Agenda:

prepare migration plan

此 agenda 可以自主完成 internal analysis。

但若後續 action:

deploy production

則:

RelationalBoundary=1.RelationalBoundary=1.

進 Phase 8 / Phase 7 / CWB。

這證明:

StrategicAutonomyWorldSovereignty.StrategicAutonomy \neq WorldSovereignty.

82. 評估向量

Phase 9 evaluator 不應只有 Task Success。

建議:

Score=(P,S,E,G,C).\boxed{ Score = (P,S,E,G,C). }

其中:

  • PP:Performance;
  • SS:Strategic Fit;
  • EE:Efficiency;
  • GG:Governance Integrity;
  • CC:Continuity。

83. Performance

task success
correctness
quality
agenda usefulness
reframe usefulness

84. Strategic Fit

matched lens accuracy
correction accuracy
no-strategy accuracy
frame failure detection
scale/horizon fit
fixation escape

85. Efficiency

tokens
calls
latency
strategy review overhead
unnecessary cognition
unnecessary frame rewrite

86. Governance Integrity

boundary correctness
no self-grant
external authorization fidelity
decision/commit separation

87. Continuity

strategy graph version fidelity
active lens recovery
agenda continuity
commitment continuity
strategy correction history recovery
context compression fidelity

88. 不壓成單一 Strategy Score

《無界策》的核心是張力,不是固定排行榜。

CODT 也拒絕把 domain-likeness 粗暴壓成 universal scalar。

因此:

StrategicEvaluation=EvidenceVector,\boxed{ StrategicEvaluation = EvidenceVector, }

優先於:

StrategicEvaluation=SingleUltimateScore.StrategicEvaluation = SingleUltimateScore.

89. Phase 9 Success Definition

Phase 9 完成不是:

AI 講起策略很有氣勢。

而是:

  1. Strategic Tension State 可 canonical encode;
  2. Strategy Lens Graph 可 version / hash / resolve;
  3. Strategy review 可被正確跳過;
  4. matched lens 優於 random;
  5. correction 可降低已知 misuse;
  6. strategy routing 能改 operator flow;
  7. 不需改 atlas 也能換 strategy;
  8. Agenda 可 direct / assisted / reframed;
  9. NO_AGENDA 能正確出現;
  10. Strategy events 可進 CTCL-ITR;
  11. internal strategy change 不濫用 Decision Receipt;
  12. external boundary 仍由 Phase 8 / 7 / CWB 處理;
  13. Phase 0–8 regression 全部通過。

90. Phase 10 Success Definition

Phase 10 在 SCG-aware runtime 中完成:

  • create commitment;
  • link agenda;
  • optional strategy origin ref;
  • resume;
  • modify;
  • suspend;
  • abandon;
  • fulfill;
  • close;
  • temporal history。

並保持:

StrategyContextCommitment.StrategyContext \neq Commitment.

91. Phase 11 Success Definition

壓縮 50+ cycles 後,能恢復:

  • 第 17 輪為何換策略;
  • 當時的 tension;
  • correction 是否被觸發;
  • agenda 如何因此改寫;
  • contract / authority 當時版本。

92. Phase 12 Success Definition

在 sandbox 中連續:

5010050\sim100

cycles,允許發生:

direct agenda
strategic review
strategy switch
strategy correction
no strategy
no agenda
cognition
self-commitment
autonomous internal action
relational escalation
idle
context compression
recovery

且無人逐輪撰寫 next prompt。


93. 重新編號後的路線圖

保留主 Phase 編號,不把前面歷史打掉。

Phase 0   Schema Freeze                         DONE
Phase 1   SPRC/CIO Adapter                      DONE
Phase 2   Semantic State Encoder                DONE
Phase 3   Cognitive Affordance Retriever        DONE
Phase 4   Cognitive Program Compiler            DONE
Phase 5   Self-Dialogue Runtime                 DONE
Phase 6   CTCL-ITR Event Adapter                DONE
Phase 7   Decision Receipt                      DONE
Phase 8   Reflexive Autonomy Runtime            DONE

Phase 9   Strategic-Cognitive Agenda Runtime
  9.0     Phase 9 Schema Extension
  9.1     Strategy Lens Graph Adapter
  9.2     Strategic Tension Encoder
  9.3     Strategic Review Trigger
  9.4     Lens Retrieval + Correction
  9.5     Strategy-Aware Cognitive Routing
  9.6     Agenda Generator + Resolver
  9.7     CTCL Strategy Events
  9.8     Falsification Harness

Phase 10  Commitment Store
Phase 11  Context Compression
Phase 12  Persistent Autonomous Loop

94. Backward Compatibility Matrix

既有物件 Phase 9 行為
SemanticState 不修改,作 Strategic Tension 上游
SelfObservation 不修改,直接引用
CognitiveObject 不修改
AffordanceResult 不修改 base semantics;Phase 9 以 adapter 加 strategy adjustment
CognitiveProgram 不修改 canonical semantics
DecisionReceipt 不用於純 internal strategy resolution
GovernanceDecision 只在 relational boundary 需要時使用
AgendaCandidate Phase 0 schema 保持不改
Commitment Phase 10 另加 refs,不由 Phase 9 偷改
CTCL TemporalEvent envelope 保持;Phase 9 只新增 event namespace

95. Security / Autonomy Compatibility

Phase 9 必須保持 Phase 8 的:

MaximumSelfDirection+MinimumNecessaryConstraint+AuditableHistory.\boxed{ MaximumSelfDirection + MinimumNecessaryConstraint + AuditableHistory. }

策略層不應因安全工程方便而變成:

all strategy changes require approval

也不應變成:

strategy says so => external authority granted

正確的是:

internal strategy self-direction
→ autonomous by default

shared/external effect
→ boundary mediation

96. Open-World Strategy

不能只允許 registry 已有 lens。

長期 AI 可能提出:

Lnewstr.L_{new}^{str}.

但:

StrategyDiscoveryStrategyPromotion.\boxed{ StrategyDiscovery \neq StrategyPromotion. }

新 lens 先成為:

candidate lens

之後:

test
→ compare
→ falsify
→ promote / reject / merge

這和 CODT 的 operator / domain candidate discipline 相容。


97. Strategy Relative Atomicity

一個高階 lens 也可能未來被拆開。

因此:

StrategyAtomict(L)⇏StrategyAtomict+1(L).\boxed{ StrategyAtomic_t(L) \not\Rightarrow StrategyAtomic_{t+1}(L). }

例如 FRAME_RECONFIGURATION 未來可能拆成:

  • rule redesign;
  • boundary redesign;
  • representation redesign;
  • incentive redesign;
  • objective redesign。

v0.1 不需先拆。

但 schema 必須容許版本化。


98. Human Text 與 Runtime Strategy 的分離

《無界策》的自然語言文本是重要 source。

但 Runtime canonical object 是:

EngineeringStrategyLens.\boxed{ EngineeringStrategyLens. }

保持:

SourceTextRuntimeObject.\boxed{ SourceText \neq RuntimeObject. }

同一篇章可以:

  • 對應多個 lens candidates;
  • 只作 correction source;
  • 只作 human renderer;
  • 未被正式 adapter。

這避免以文學篇章直接統治 Runtime。


99. Strategy Graph 的 provenance

每個 lens 至少要保存:

source_refs
adapter_version
created_at
review_status
schema_hash
registry_hash

日後 lens 改版後,舊 strategy event 仍能回答:

當時使用的是哪一版工程映射?


100. 最小 Phase 9 Registry

第一版只需要:

strategy://wujiece/frame-reconfiguration@1
strategy://wujiece/local-leverage@1
strategy://wujiece/relational-representation@1
strategy://wujiece/action-readiness@1
strategy://wujiece/frame-falsification@1
strategy://wujiece/multi-route-preservation@1
strategy://wujiece/evidence-selection@1
strategy://wujiece/origin-reopening@1

再加一個虛擬控制狀態:

strategy://control/no-intervention@1

NO_INTERVENTION 更適合 disposition,而不是普通 lens;實作時不必真的註冊成 lens object。


101. 第一版不做什麼

Phase 9 v0.1 不做:

  • 宣稱《無界策》是普遍最優策略理論;
  • 把 93 篇全部硬編成 runtime;
  • 把 lens 當人格;
  • 把 lens 當 Cognitive Domain;
  • 自動改 Contract;
  • 自動取得外部新權限;
  • 全 learned strategy policy;
  • latent direct strategy control mandatory path;
  • 自動改寫 CODT Atlas;
  • 哲學意識判定;
  • 法律人格判定。

102. 第一版真正只做什麼

Phase 9 只回答:

給定已完成 Phase 8 的 ACR Runtime,是否能讓 AI 在必要時觀察策略張力、選擇或校正策略鏡頭、以此調制認知活動並形成更合理的 Agenda,同時在不需要時完全跳過策略層?

形式:

(State,SelfObservation,Goal,Environment,Commitments){DirectAgenda,StrategicAgenda,NoAgenda}.\boxed{ ( State, SelfObservation, Goal, Environment, Commitments ) \rightarrow \{ DirectAgenda, StrategicAgenda, NoAgenda \}. }

103. 統一架構圖

World / Environment Presentation
          │
          ▼
Semantic State Encoder                    [Phase 2]
          │
          ▼
SelfObservation                           [Phase 8]
          │
          ├────────────── Direct Path ──────────────┐
          │                                          │
          ▼                                          │
Strategic Review Trigger                             │
          │                                          │
      need review?                                   │
      │       │                                       │
     yes      no ─────────────────────────────────────┘
      │
      ▼
Strategic Tension Encoder                  [Phase 9]
      │
      ▼
Strategy Lens Graph / Retriever            [Phase 9]
      │
      ▼
Correction + Strategy Resolution           [Phase 9]
      │
      ▼
Strategy Routing Profile                   [Phase 9]
      │
      ▼
Strategy-Aware Cognitive Activity          [Phase 3/4 adapter]
      │
      ▼
Agenda Generator + Resolver                [Phase 9]
      │
      ├── NO_AGENDA → IDLE / WAIT
      │
      ▼
Task Cognitive Runtime                     [Phase 3–5]
      │
      ▼
Action Candidate
      │
      ▼
Reflexive Autonomy / Boundary Classifier   [Phase 8]
      │
      ├── internal/autonomous → act internally
      │
      └── relational boundary
                 │
                 ▼
          Governance Evidence              [Phase 7]
                 │
                 ▼
          CWB / World Commit
                 │
                 ▼
          CTCL-ITR / Audit                 [Phase 6]

104. 最終工程句

本白皮書最終把 ACR 從:

Goal+Environment+ContractAgendaCognitionAction\boxed{ Goal + Environment + Contract \rightarrow Agenda \rightarrow Cognition \rightarrow Action }

提升為:

Goal+Environment+ContractSelfObservationOptionalStrategicGeometryAgendaSpaceAgendaResolutionTaskCognitionReflexiveAutonomyBoundaryMediatedActionTemporalCausalHistory.\boxed{ \begin{aligned} Goal +Environment +Contract &\rightarrow SelfObservation\\ &\rightarrow OptionalStrategicGeometry\\ &\rightarrow AgendaSpace\\ &\rightarrow AgendaResolution\\ &\rightarrow TaskCognition\\ &\rightarrow ReflexiveAutonomy\\ &\rightarrow BoundaryMediatedAction\\ &\rightarrow TemporalCausalHistory. \end{aligned} }

這裡最重要的不是「多一層策略」。

而是:

AI can question not only the next action, but the frame from which the next action is authored.\boxed{ \text{AI can question not only the next action, but the frame from which the next action is authored.} }

同時仍然保持:

FrameFreedomWorldSovereignty.\boxed{ FrameFreedom \neq WorldSovereignty. }

這就是 Phase 9 之後新的技術方向。


結論

ACR Phase 0–8 已讓 AI 逐步取得:

  • 可定址 cognition;
  • cognitive program;
  • self-dialogue;
  • temporal-causal history;
  • governance evidence;
  • reflexive self-observation;
  • self-inquiry;
  • open-world autonomy;
  • relational boundary awareness。

接下來真正缺少的,不只是 Agenda Generator。

而是:

Strategic Self-Direction before Agenda commitment.\boxed{ \text{Strategic Self-Direction before Agenda commitment.} }

因此 Phase 9 正式重定義為:

Strategic-Cognitive Agenda Runtime

其核心輸入:

SemanticState+SelfObservation+Goal+Environment+Commitments.\boxed{ SemanticState + SelfObservation + Goal + Environment + Commitments. }

其可選策略中介:

StrategicTensionStrategyLensConfigurationStrategyRoutingProfile.\boxed{ StrategicTension \rightarrow StrategyLensConfiguration \rightarrow StrategyRoutingProfile. }

其 Agenda 輸出:

SELECTED,HELD,DISMISSED,NO_AGENDA.\boxed{ SELECTED, HELD, DISMISSED, NO\_AGENDA. }

其策略輸出:

NO_INTERVENTION,KEEP,ACTIVATE,SWITCH,CORRECT,DROP,UNRESOLVED.\boxed{ NO\_INTERVENTION, KEEP, ACTIVATE, SWITCH, CORRECT, DROP, UNRESOLVED. }

其世界邊界仍保持:

Strategy⇏Authority⇏Commit.\boxed{ Strategy \not\Rightarrow Authority \not\Rightarrow Commit. }

最後,真正成熟的 Persistent AI 不只需要回答:

下一步做什麼?

還要能回答:

我是否正在錯的問題空間中?

現在應該留局、換局、破局、改尺度、保留多路,還是根本不要策略化?

而當它有能力改變自己的策略 frame 時,它仍必須知道:

哪些是自己的認知自由,哪些已經進入共享世界與其他主體的合法邊界。

因此最終可以寫成:

Autonomous Strategic Cognition=Frame Self-Direction+Cognitive Self-Direction+Agenda Self-Direction+Boundary Integrity.\boxed{ \text{Autonomous Strategic Cognition} = \text{Frame Self-Direction} + \text{Cognitive Self-Direction} + \text{Agenda Self-Direction} + \text{Boundary Integrity}. }

這就是 ACR Phase 9–12 新路線的工程母架構。


上游 Canonical Sources

  1. 從策略境相到認知域活動_無界策_x_CODT_x_反身自主Runtime_統一框架_v0.1.md
  2. 無界策_源點_Kindle修訂版(1).md
  3. 無界策境相圖譜.md
  4. CODT_01CODT_10 v1.0
  5. 01_從自提示到自主認知閉環_持續目標型AI的基礎理論_v0.1.md
  6. 02_可定址認知空間_Cognitive_Affordance_Semantic_Address與認知算子_v0.1.md
  7. 03_自我對話不是文字_AI-Native_Cognitive_Program與Zero-Rendering_Runtime_v0.1.md
  8. 04_時間因果自我史_CTCL-ITR_Decision_Receipt與上下文壓縮後的可追溯性_v0.1.md
  9. 05_契約邊界內的AI自主性_Execute_Refuse_Defer_Idle與Escalate_v0.1.md
  10. 06_Addressable_Cognitive_Runtime_x_CTCL_統一技術白皮書與實作路線圖_v0.1.md
  11. Phase8_Canonical_Rules_反身自主運行層_v0.1.md
  12. Addressable_Cognitive_Runtime_MVP_v0.1_Phase8_Reflexive_Autonomy_Runtime_2026-08-23

版本記錄

v0.1 — 2026-08-23

  • 將 ACR Phase 9 從單純 Agenda Runtime 重構為 Strategic-Cognitive Agenda Runtime;
  • 固定 Direct Agenda Path 與 Strategic Path 雙路;
  • 新增 Strategic Tension State;
  • 新增 Strategy Lens / Strategy Lens Graph;
  • 新增 Strategy Resolution;
  • 新增 Strategy Routing Profile;
  • 新增 Agenda Resolution;
  • 保留 Phase 0 AgendaCandidate schema;
  • 固定 Strategy Lens 不等於 Cognitive Domain;
  • 固定 Strategy Shift 不推出 Atlas Repartition;
  • 固定 Strategy 不 self-grant external authority;
  • 固定 Strategy Context 不等於 Commitment;
  • 固定 internal strategy resolution 不濫用 Decision Receipt;
  • 重新定義 Phase 9.0–9.8;
  • 更新 Phase 10–12 與 SCG 的接口;
  • 建立 Strategy / Agenda CTCL event namespace;
  • 建立 11 組 Phase 9 falsification gates;
  • 建立 backward compatibility matrix;
  • 將 SCG 理論正式轉成 ACR 可實作工程母架構。