← Archive
lm-004175 · 2026-09

類全域 AI 世界—計算—觀察統合系列(Paper 09) — 執行層:方法選擇、MSSP、RDR、WFS 與 Runtime Routing

下載 MD 檔 ⬇

類全域 AI 世界—計算—觀察統合系列(Paper 09)

執行層:方法選擇、MSSP、RDR、WFS 與 Runtime Routing

The Execution Layer: Method Selection, MSSP, RDR, WFS, and Runtime Routing

作者: Neo.K
AI 協作: Aletheia(GPT-5.6 Sol)
機構: EveMissLab/一言諾科技有限公司
系列: 類全域 AI 世界—計算—觀察統合系列
英文系列名: Global-Like AI World–Computation–Observation Synthesis Series
篇次: Paper 09 / 12
版本: v0.1
日期: 2026-09-09
研究定位: MSSP × RDR × CAIR × World Family Scheduler × Global Computation Router × Memory Strategy Governor × Global Observer × CPC 2.0 × PPOS 2.0 × Mother Runtime × Durable Execution
前篇: Paper 08《記憶層:從資訊海到可重建世界與世界族》
狀態: WCO 執行層母規格/Thin Mother Runtime orchestration architecture;不宣稱存在單一普遍最優 scheduler,也不宣稱任一現成 Agent framework 已完整實作本文架構


摘要

Paper 01–08 已建立類全域 AI 的主要能力層:

WtG,CtWF,OtG,DtWCO,PtG,PRSt,MtWCO.\mathfrak W_t^G, \quad \mathfrak C_t^{WF}, \quad \mathfrak O_t^G, \quad \mathfrak D_t^{WCO}, \quad \mathfrak P_t^G, \quad \mathsf{PRS}_t, \quad \mathfrak M_t^{WCO}.

然而,「系統擁有這些能力」不等於「系統知道什麼時候、以什麼順序、在什麼權限下、用什麼資源把它們組合成一次可靠執行」。

因此:

Capability AvailabilityExecutable Orchestration.\boxed{ \text{Capability Availability} \neq \text{Executable Orchestration}. }

本文建立 WCO-TF 的正式 Execution Layer,核心不是再增加一個萬能 scheduler,而是把不同層級、不同語義的 routing / scheduling / authority / materialization 保持分離,再由一個薄型 Mother Runtime Meta-Orchestrator 透過明確 contract 協調。

本文承接 MSSP × RDR 的基本分工:

MSSP=What\boxed{ MSSP = What }

與:

RDR=How.\boxed{ RDR = How. }

MSSP 負責可定址能力空間、分類、關係、版本與 capability discovery;RDR 負責指定版本如何在某 runtime environment 中被物化、綁定 provider、租用資源、派發、執行、追蹤、回收與恢復。

本文同時保留 CAIR 的權威分離:

UIAuthorityIndexRuntime.\boxed{ UI \neq Authority \neq Index \neq Runtime. }

但進一步限制:CAIR 不被提升成整個 WCO 的唯一總資料庫或唯一真理來源。 對 capability definition、workflow contract 與 execution semantics,CAIR 可作 canonical authority;對 world state、memory claims、physical evidence 等其他 data class,應遵守:

One Canonical Authority per Data Class\boxed{ \text{One Canonical Authority per Data Class} }

而不是:

One Giant Authority for Everything.\boxed{ \text{One Giant Authority for Everything}. }

本文定義 WCO Execution Orchestration State(EOS)

EOSt=It,Atcap,Ptplan,Rtroute,Xtrun,Vtexec,Gtauth,Btres,Utexec,Htreceipt.\boxed{ \mathsf{EOS}_t = \left\langle \mathfrak I_t, \mathfrak A_t^{cap}, \mathfrak P_t^{plan}, \mathfrak R_t^{route}, \mathfrak X_t^{run}, \mathfrak V_t^{exec}, \mathfrak G_t^{auth}, \mathfrak B_t^{res}, \mathfrak U_t^{exec}, \mathfrak H_t^{receipt} \right\rangle. }

其中:

  • It\mathfrak I_t:intent / task-contract state;
  • Atcap\mathfrak A_t^{cap}:capability address space;
  • Ptplan\mathfrak P_t^{plan}:execution-plan graph family;
  • Rtroute\mathfrak R_t^{route}:layer routing decisions;
  • Xtrun\mathfrak X_t^{run}:active materialized runs;
  • Vtexec\mathfrak V_t^{exec}:validation / verification state;
  • Gtauth\mathfrak G_t^{auth}:authority / policy state;
  • Btres\mathfrak B_t^{res}:resource / budget state;
  • Utexec\mathfrak U_t^{exec}:unknown / blocked / failure / replan state;
  • Htreceipt\mathfrak H_t^{receipt}:execution receipts / checkpoints / provenance。

所有自然語言、API 或 Agent intent 都必須先轉成 Task Identity Contract(TIC)

Iq=TaskId,Goal,Inputs,Outputs,Invariants,Constraints,RequiredQualification,Risk,Budget,AllowedEffects,Authority,StopConditions.\boxed{ \mathfrak I_q = \left\langle TaskId, Goal, Inputs, Outputs, Invariants, Constraints, RequiredQualification, Risk, Budget, AllowedEffects, Authority, StopConditions \right\rangle. }

其核心目的,是防止 runtime 在 replan、representation escape、capability composition 或 capability construction 中悄悄改題:

q̸IqqNew Task Contract Required.\boxed{ q' \not\equiv_{\mathfrak I_q} q \Rightarrow \text{New Task Contract Required}. }

接著,系統建立 Task Signature:

Σq=(D,I,O,E,C,B,V,R,A,S),\boxed{ \Sigma_q = ( D, I, O, E, C, B, V, R, A, S ), }

其中可包含 domain、input、output、exactness、constraints、budget、verification requirement、risk、authority requirement 與 stop condition。

MSSP 依:

Σq\Sigma_q

解析:

Candidates(q)={A1,,Ak}.\boxed{ Candidates(q) = \{A_1,\ldots,A_k\}. }

若沒有單一 capability,可進行:

Select,Compose,Construct.Select, Compose, Construct.

但 generated capability 只能先成為:

CandidateCapabilityCandidateCapability

而不能直接升格為 trusted capability。

本文提出 Layered Runtime Routing Fabric(LRRF)。它不把所有 routing 合成單一函數,而承認至少以下不同問題:

  • MSSP:哪個 structural capability relevant?
  • WFS:哪些 worlds 應 continue / fork / prune / reobserve?
  • Computation Router:哪個 computational form / law / backend?
  • Memory Router:要從哪些 memory provider、以什麼 fidelity 取什麼?
  • Observation Router:現在應觀察什麼?
  • Domain Router:下一步需要補哪一種 epistemic qualification?
  • CPC 2.0:如何 projection、選 carrier?
  • PPOS 2.0:physical target 是否可實現?
  • RDR:如何物化、選 provider、派發、checkpoint、retry / fallback?
  • Governance / CAIR:這個 semantic definition、version、permission、effect 是否被允許?
  • Mother Runtime Meta-Orchestrator:如何讓上述決策形成一個可持續、可恢復、可審計的執行生命週期?

因此:

MSSPWFSComputeRouterMemoryRouterObserverRouterCPCRDRMotherMetaOrchestrator.\boxed{ MSSP \neq WFS \neq ComputeRouter \neq MemoryRouter \neq ObserverRouter \neq CPC \neq RDR \neq MotherMetaOrchestrator. }

本文進一步提出 Execution Plan Graph(EPG)

GtEP=(VtEP,EtEP,ΓtEP,ΣtEP,HtEP).\boxed{ \mathcal G_t^{EP} = ( V_t^{EP}, E_t^{EP}, \Gamma_t^{EP}, \Sigma_t^{EP}, H_t^{EP} ). }

其中 plan node 可具有 typed role:

StepType{Resolve,Retrieve,Reconstruct,ForkWorld,Compute,Observe,Judge,Verify,Project,Materialize,ProposeEffect,Authorize,Dispatch,ObserveOutcome,Reconcile,CommitState,Wait,Stop,Escalate}.\boxed{ \mathsf{StepType} \in \{ Resolve, Retrieve, Reconstruct, ForkWorld, Compute, Observe, Judge, Verify, Project, Materialize, ProposeEffect, Authorize, Dispatch, ObserveOutcome, Reconcile, CommitState, Wait, Stop, Escalate \}. }

Execution Plan 不是 canonical world state,也不是 capability definition:

PlanCapabilityExecutionWorldState.\boxed{ Plan \neq Capability \neq Execution \neq WorldState. }

Plan 可以因 runtime evidence 受控重規劃:

Plan0etPlan1.Plan_0 \xrightarrow{e_t} Plan_1.

但 replan 必須保持:

Iq\boxed{ \mathfrak I_q }

中的 task invariants;若要改變核心目標或 acceptance criteria,應建立新 task / branch,而不是 silent redefinition。

本文特別區分六種常被混用的恢復行為:

RetryFallbackRerouteRecomposeReframeEscalate.\boxed{ Retry \neq Fallback \neq Reroute \neq Recompose \neq Reframe \neq Escalate. }
  • Retry:同一 step / route 重試;
  • Fallback:同 capability 換 provider / runtime;
  • Reroute:換 capability,但 task contract 不變;
  • Recompose:重組多個 capabilities;
  • Reframe:換 representation / domain route;
  • Escalate:要求更高 verifier、authority、human 或 specialist。

同樣:

PauseWaitStopCancelRollback.\boxed{ Pause \neq Wait \neq Stop \neq Cancel \neq Rollback. }

Stop 是合法 policy,不必等於 failure。若 task 已足夠完成、required qualification 已達到、Value of Information 低於成本、authority blocked 或 budget rationally exhausted,runtime 可以安全停止。

本文把 external effect path 再細分。對高影響真實世界 action:

IntentActionProposalAuthorizedDispatchPhysicalEffectObservedOutcomeWorldStateCommit.\boxed{ Intent \neq ActionProposal \neq AuthorizedDispatch \neq PhysicalEffect \neq ObservedOutcome \neq WorldStateCommit. }

因此真實作用的安全閉環應為:

ProposalValidateAuthorizeDispatchObserveOutcomeReconcileWorldStateCommit.\boxed{ Proposal \rightarrow Validate \rightarrow Authorize \rightarrow Dispatch \rightarrow ObserveOutcome \rightarrow Reconcile \rightarrow WorldStateCommit. }

這比「模型產生 tool call 就算世界已改變」更嚴格。

本文亦定義 Runtime Receipt Bundle(RRB)

RRB=IntentReceipt,ResolutionReceipt,PlanReceipt,RoutingReceipts,MaterializationReceipt,ExecutionReceipts,VerificationReceipt,EffectReceipt,StateCommitReceipt.\boxed{ \mathsf{RRB} = \left\langle IntentReceipt, ResolutionReceipt, PlanReceipt, RoutingReceipts, MaterializationReceipt, ExecutionReceipts, VerificationReceipt, EffectReceipt, StateCommitReceipt \right\rangle. }

每一層可重建「為什麼這一步發生」,但 receipts 是 operational evidence,不自動等於 epistemic truth。

最後,本文把上述執行骨架放進 Thin Mother Runtime。承接 Mother Runtime 的五平面:

RM=(PS,PC,PE,PO,PG),\boxed{ \mathfrak R_M = ( \mathcal P_S, \mathcal P_C, \mathcal P_E, \mathcal P_O, \mathcal P_G ), }

其中:

  • PS\mathcal P_S:State Plane;
  • PC\mathcal P_C:Cognitive Control Plane;
  • PE\mathcal P_E:Effect Plane;
  • PO\mathcal P_O:Observability / Recovery Plane;
  • PG\mathcal P_G:Governance Plane。

本文的新增原則是:

Thin Core, Rich Specialized Layers.\boxed{ \text{Thin Core, Rich Specialized Layers}. }

Mother Runtime 不自己做所有 specialist work,不自己成為所有 truth authority,也不因為能選擇 cognition 就自動取得 Effect / Governance authority。

它的核心責任只是:

Event Intake+Task Contract+State Projection+Layer Routing+Plan Lifecycle+Authority Fencing+Checkpoint / Recovery+Receipt / History.\boxed{ \text{Event Intake} + \text{Task Contract} + \text{State Projection} + \text{Layer Routing} + \text{Plan Lifecycle} + \text{Authority Fencing} + \text{Checkpoint / Recovery} + \text{Receipt / History}. }

因此本文最終提出:

A Global-Like AI runtime is not a single super-agent loop.\boxed{ \text{A Global-Like AI runtime is not a single super-agent loop.} }

而是:

a durable, authority-bounded meta-runtime that compiles intent into typed plans, delegates decisions to specialized routing layers, materializes capabilities through RDR, and closes execution through observation, verification, receipts, and state reconstruction.\boxed{ \text{a durable, authority-bounded meta-runtime that compiles intent into typed plans, delegates decisions to specialized routing layers, materializes capabilities through RDR, and closes execution through observation, verification, receipts, and state reconstruction}. }

關鍵詞: MSSP、RDR、CAIR、WFS、Mother Runtime、Execution Plan Graph、Capability Routing、Durable Execution、Task Identity Contract、Candidate/Commit、Replan、Runtime Receipt、Global AI


0. Paper 08 留下的問題

Paper 08 已能:

MemoryWorldReconstruction.Memory \rightarrow WorldReconstruction.

但 reconstructed world 並不會自己決定:

下一步做什麼?


1. Intent 不等 Execution

IntentExecution.\boxed{ Intent \neq Execution. }

2. Intent 不等 Effect

IntentPhysicalEffect.\boxed{ Intent \neq PhysicalEffect. }

3. Natural Language 不能直接 Dispatch

自然語言先轉:

IntentTaskContract.Intent \rightarrow TaskContract.

4. Task Identity Contract

Iq=TaskId,Goal,Inputs,Outputs,Invariants,Constraints,RequiredQualification,Risk,Budget,AllowedEffects,Authority,StopConditions.\boxed{ \mathfrak I_q = \left\langle TaskId, Goal, Inputs, Outputs, Invariants, Constraints, RequiredQualification, Risk, Budget, AllowedEffects, Authority, StopConditions \right\rangle. }

5. Task Contract 防止 Silent Reframing

q̸Iqqq' \not\equiv_{\mathfrak I_q}q

時,不能宣稱同一任務已完成。


6. Allowed Transformations

Task contract 可明確允許:

  • decomposition;
  • representation change;
  • provider substitution;
  • capability composition;
  • world branching。

7. Disallowed Transformations

例如:

  • approximate solution when exact required;
  • scope reduction;
  • task redefinition;
  • authority expansion;
  • evidence standard lowering。

8. Task Signature

Σq=(D,I,O,E,C,B,V,R,A,S).\boxed{ \Sigma_q = ( D, I, O, E, C, B, V, R, A, S ). }

9. Capability Address Space

Atcap.\mathfrak A_t^{cap}.

每個 capability 具有 stable identity。


10. Capability Identity 不等 Provider Identity

CapabilityIdentityProviderIdentity.\boxed{ CapabilityIdentity \neq ProviderIdentity. }

11. Provider 可以替換

同一 capability 可以由:

  • local code;
  • cloud API;
  • remote worker;
  • GPU kernel;
  • human expert;

實作,只要 contract 保持。


12. MSSP 的責任

MSSP 回答:

系統中有什麼能力?

這些能力怎麼分類?

哪些和 task signature 相符?


13. MSSP 不直接執行

MSSPExecutor.\boxed{ MSSP \neq Executor. }

14. MSSP 不是完整 Authority Store

它是從 canonical definition 建立的可重建索引。


15. CAIR 的責任

對 capability / workflow semantics:

CAIRCAIR

保存:

  • node / capability definition;
  • type;
  • contract;
  • effects;
  • permission;
  • version;
  • provenance;
  • certificates。

16. CAIR 不等 UI

UICAIR.\boxed{ UI \neq CAIR. }

17. CAIR 不等 MSSP

AuthorityIndex.\boxed{ Authority \neq Index. }

18. CAIR 不等 RDR

AuthorityRuntimeMaterialization.\boxed{ Authority \neq RuntimeMaterialization. }

19. One Authority per Data Class

capability semantics 可由 CAIR authority。

world state 可由 world-state authority。

memory claim 可由 SEDB / memory authority。


20. 不是 One Giant Truth Database

FederatedCanonicalAuthoritiesAuthorityChaos.\boxed{ FederatedCanonicalAuthorities \neq AuthorityChaos. }

只要每個 data class 的 authority boundary 清楚。


21. MSSP Candidate Resolution

Candidates(q)={A1,,Ak}.\boxed{ Candidates(q) = \{A_1,\ldots,A_k\}. }

22. Deterministic Resolution First

type、schema、exact tag、declared relation 能解的部分:

DeterministicFirst.\boxed{ DeterministicFirst. }

23. Semantic Resolution 是補充

AI 用於:

  • ambiguous mapping;
  • unseen wording;
  • weak relation;
  • capability similarity。

24. Select

有單一合適 capability:

A.A^\ast.

25. Compose

如果需要:

A=AnA1.A^\ast = A_n \circ \cdots \circ A_1.

26. Construct

若沒有既有 capability,可產生:

Acandnew.A_{cand}^{new}.

27. Generated Capability 不等 Trusted Capability

GeneratedCapabilityTrustedCapability.\boxed{ GeneratedCapability \neq TrustedCapability. }

28. Capability Promotion

新 capability 必須經:

  • tests;
  • verifier;
  • sandbox;
  • governance;
  • versioning;

才能進正式 registry。


29. Capability Construction 不等 Arbitrary Self-Modification

v0.1 應保持:

Bounded+Typed+Auditable+ReversibleWherePossible.\boxed{ Bounded + Typed + Auditable + ReversibleWherePossible. }

30. Layered Routing Fabric

本文定義:

LRRF={RMSSP,RWFS,Rcomp,Rmem,Robs,Rdom,Rproj,Rphys,RRDR}.\boxed{ LRRF = \{ R_{MSSP}, R_{WFS}, R_{comp}, R_{mem}, R_{obs}, R_{dom}, R_{proj}, R_{phys}, R_{RDR} \}. }

31. MSSP Router

回答:

哪些 capabilities relevant?


32. WFS

回答:

哪些 worlds 應 continue / fork / merge / prune / pause?


33. WFS Actions

{Continue,Fork,Merge,Prune,Pause,Reobserve,IncreaseResolution,PromoteCandidate}.\{ Continue, Fork, Merge, Prune, Pause, Reobserve, IncreaseResolution, PromoteCandidate \}.

34. PromoteCandidate 不等 Reality Commit

PromoteCandidateCommitToReality.\boxed{ PromoteCandidate \neq CommitToReality. }

35. Computation Router

回答:

哪個 computational form、transition law、backend、fidelity?


36. Memory Router

回答:

哪個 memory provider / representation / fidelity / source depth?


37. Observation Router

回答:

現在看哪個 world / domain / operator / resolution?


38. Domain Router

回答:

目前缺哪個 epistemic qualification?


39. Projection Router / CPC 2.0

回答:

observation content 如何變成 task-relative view?


40. Physical Layer

回答:

carrier 在物理上做得到嗎?


41. RDR

回答:

指定 capability 版本如何在指定 environment 被實際物化與派發?


42. Mother Meta-Orchestrator

回答:

哪些 layer decision 現在需要被觸發、等待、重做或收束?


43. 不建立 One Scheduler to Rule Them All

Cross-Layer CoordinationScheduler Collapse.\boxed{ \text{Cross-Layer Coordination} \neq \text{Scheduler Collapse}. }

44. Why Not Collapse?

因為:

  • world value;
  • compute cost;
  • epistemic debt;
  • memory fidelity;
  • observation value;
  • provider health;

是不同 decision semantics。


45. Execution Plan Graph

GtEP=(VtEP,EtEP,ΓtEP,ΣtEP,HtEP).\boxed{ \mathcal G_t^{EP} = ( V_t^{EP}, E_t^{EP}, \Gamma_t^{EP}, \Sigma_t^{EP}, H_t^{EP} ). }

46. Plan Node Type

StepType{Resolve,Retrieve,Reconstruct,ForkWorld,Compute,Observe,Judge,Verify,Project,Materialize,ProposeEffect,Authorize,Dispatch,ObserveOutcome,Reconcile,CommitState,Wait,Stop,Escalate}.StepType \in \{ Resolve, Retrieve, Reconstruct, ForkWorld, Compute, Observe, Judge, Verify, Project, Materialize, ProposeEffect, Authorize, Dispatch, ObserveOutcome, Reconcile, CommitState, Wait, Stop, Escalate \}.

47. Plan Edge

可以表示:

  • dependency;
  • condition;
  • data;
  • control;
  • failure;
  • retry;
  • fallback;
  • authority;
  • evidence dependency。

48. Plan Graph 不等 World Graph

ExecutionPlanGraphWorldInteractionGraph.\boxed{ ExecutionPlanGraph \neq WorldInteractionGraph. }

49. Plan Graph 不等 Capability Graph

PlanGraphCapabilityOntology.\boxed{ PlanGraph \neq CapabilityOntology. }

50. Plan Graph 不等 Computation Graph

某 plan node 可以呼叫一個內部複雜 computational graph。


51. Plan 是一次 Task-Relative Compilation

GqEP=Compile(Iq,Statet,Capabilitiest,Policiest).\mathcal G_q^{EP} = Compile( \mathfrak I_q, State_t, Capabilities_t, Policies_t ).

52. Plan Node 必須有 Owner Layer

例如:

memory.resolve -> WRMF
world.fork -> WFS
compute.route -> GCM
observe.select -> Global Observer
project.compile -> CPC
effect.dispatch -> RDR
authority.check -> Governance

53. Owner Layer 避免 Meta-Orchestrator 吞權

Mother Core 只發出 request / collect decision。


54. Plan Status

status{Draft,Validated,Running,Waiting,Paused,Blocked,Replanning,Completed,Stopped,Cancelled,Failed}.status \in \{ Draft, Validated, Running, Waiting, Paused, Blocked, Replanning, Completed, Stopped, Cancelled, Failed \}.

55. Waiting 不等 Failed

WaitingFailed.\boxed{ Waiting \neq Failed. }

56. Blocked 不等 Cancelled

Blocked 可能等待:

  • authority;
  • evidence;
  • resource;
  • human;
  • external event。

57. Execution Plan 與 Runtime Materialization 分離

PlanMaterializedRuntime.\boxed{ Plan \neq MaterializedRuntime. }

58. RDR Materialization

Qv,e=MaterializeRDR(Pv,e).Q_{v,e} = Materialize_{RDR}(P_v^\ast,e).

59. Materialization 固定精確版本

不能只依:

latest
stable
default

等 movable labels。


60. Content Hash

正式 run 應固定:

Iver,Icontent.I_{ver}, \quad I_{content}.

61. Provider Snapshot

執行時固定:

  • provider id;
  • provider version;
  • executable version;
  • node;
  • config hash。

62. Provider Health 與 Capability Validity 分離

ProviderHealthCapabilitySemanticValidity.\boxed{ ProviderHealth \neq CapabilitySemanticValidity. }

63. Resource Lease

RDR 取得:

  • CPU;
  • GPU;
  • memory;
  • network;
  • storage;
  • time;
  • quota。

64. Resource Lease 不等 Ownership

ResourceLeaseResourceOwnership.\boxed{ ResourceLease \neq ResourceOwnership. }

65. Execution Step

Xi=(StepId,CapabilityVersion,ProviderSnapshot,Inputs,Policy,Budget,Retry,Fallback,Checkpoint,EffectClass).X_i = ( StepId, CapabilityVersion, ProviderSnapshot, Inputs, Policy, Budget, Retry, Fallback, Checkpoint, EffectClass ).

66. Effect Class

本文建議:

EffectClass{None,Internal,Reversible,Compensatable,Irreversible}.EffectClass \in \{ None, Internal, Reversible, Compensatable, Irreversible \}.

67. Effect Classification 影響 Gate

Irreversible:

stronger validation + authority.\Rightarrow \text{stronger validation + authority}.

68. Idempotency

可能重試的 external effect step 應盡可能提供:

IdempotencyKey.IdempotencyKey.

69. Exactly-Once 不應隨意宣稱

跨外部系統:

GlobalExactlyOnce\boxed{ GlobalExactlyOnce }

通常不是免費性質。


70. 更實際的策略

AtLeastOnce+Idempotency+Dedup+EffectLedger+Compensation.AtLeastOnce + Idempotency + Dedup + EffectLedger + Compensation.

71. Retry

Retry(step)Retry(step)

同 capability、同 semantic path。


72. Fallback

Fallback(provideraproviderb).Fallback(provider_a\rightarrow provider_b).

同 capability 換 provider。


73. Reroute

Reroute(AiAj).Reroute(A_i\rightarrow A_j).

換 capability,但 Task Contract 不變。


74. Recompose

重新組合:

A1,A2,.A_1,A_2,\ldots.

75. Reframe

更換:

  • representation;
  • observation;
  • domain route。

76. Escalate

交給:

  • stronger model;
  • formal verifier;
  • human;
  • higher authority;
  • specialist system。

77. 六種 Recovery 不同

RetryFallbackRerouteRecomposeReframeEscalate.\boxed{ Retry \neq Fallback \neq Reroute \neq Recompose \neq Reframe \neq Escalate. }

78. Replan

Plan0RuntimeEvidencePlan1.Plan_0 \xrightarrow{RuntimeEvidence} Plan_1.

79. Replan 必須保存 Task Invariants

Preserve(Iq)=1.Preserve( \mathfrak I_q )=1.

80. Task Redefinition 要 New TaskId

如果真正目標改變:

qq,q\rightarrow q',

建立新的:

TaskId.TaskId'.

81. Silent Replan 是風險

不能為了成功率偷偷:

  • 降 exactness;
  • 放寬 scope;
  • 降 verification;
  • 擴 authority。

82. Stop Conditions

可以包含:

{TaskSatisfied,QualificationSatisfied,NoNewInformation,LowVoI,BudgetExhausted,AuthorityBlocked,CapabilityMissing,HumanReview,ExternalWait,RiskCeiling}.\{ TaskSatisfied, QualificationSatisfied, NoNewInformation, LowVoI, BudgetExhausted, AuthorityBlocked, CapabilityMissing, HumanReview, ExternalWait, RiskCeiling \}.

83. Stop 不等 Failure

StopFailure.\boxed{ Stop \neq Failure. }

84. Wait 不等 Stop

等待外部 event:

Wait(Event).Wait(Event).

85. Pause 不等 Wait

Pause 是 runtime / user / policy 主動暫停。


86. Cancel 不等 Rollback

CancelRollback.\boxed{ Cancel \neq Rollback. }

87. Rollback 不等 Physical Undo

RuntimeRollbackPhysicalHistoryUndo.\boxed{ RuntimeRollback \neq PhysicalHistoryUndo. }

88. Compensation

不可回滾 effect 可嘗試:

Compensate.Compensate.

但 compensation 也不保證恢復原世界。


89. Internal Candidate / Commit

simulation state 可以:

CandidateDeltaVerifyWorldCommit.CandidateDelta \rightarrow Verify \rightarrow WorldCommit.

90. Real-World Effect 需要更長鏈

ProposalValidateAuthorizeDispatchObserveOutcomeReconcileWorldStateCommit.\boxed{ Proposal \rightarrow Validate \rightarrow Authorize \rightarrow Dispatch \rightarrow ObserveOutcome \rightarrow Reconcile \rightarrow WorldStateCommit. }

91. Authorized Dispatch 不等 Effect Succeeded

AuthorizedDispatchPhysicalEffectSucceeded.\boxed{ AuthorizedDispatch \neq PhysicalEffectSucceeded. }

92. Effect Succeeded 不等 Desired Outcome

EffectSucceededDesiredOutcome.\boxed{ EffectSucceeded \neq DesiredOutcome. }

93. Desired Outcome 不等 Observed Outcome

ExpectedOutcomeObservedOutcome.\boxed{ ExpectedOutcome \neq ObservedOutcome. }

94. World State Commit 要基於 Observed / Verified Consequence

不是基於 intent。


95. Effect Ledger

所有 high-impact action 保存:

  • requested;
  • authorized;
  • dispatched;
  • acknowledged;
  • observed;
  • reconciled;
  • compensated。

96. Runtime Receipt Bundle

RRB=IntentReceipt,ResolutionReceipt,PlanReceipt,RoutingReceipts,MaterializationReceipt,ExecutionReceipts,VerificationReceipt,EffectReceipt,StateCommitReceipt.\boxed{ RRB = \left\langle IntentReceipt, ResolutionReceipt, PlanReceipt, RoutingReceipts, MaterializationReceipt, ExecutionReceipts, VerificationReceipt, EffectReceipt, StateCommitReceipt \right\rangle. }

97. Intent Receipt

保存 task contract hash。


98. Resolution Receipt

保存:

  • MSSP candidates;
  • rejected capabilities;
  • semantic fit reason。

99. Plan Receipt

保存:

  • plan id;
  • plan version;
  • step graph;
  • dependencies;
  • policy snapshot。

100. Routing Receipts

各 router 各自出具。

不要全部塞成單一 opaque score。


101. Materialization Receipt

保存:

  • capability version;
  • content hash;
  • provider;
  • resource lease;
  • runtime id。

102. Execution Receipt

保存:

  • inputs;
  • outputs;
  • timing;
  • errors;
  • retry;
  • effect state。

103. Verification Receipt

保存:

  • verifier;
  • criteria;
  • pass / fail;
  • certificate。

104. Effect Receipt

保存 external effect lifecycle。


105. State Commit Receipt

保存:

  • before world version;
  • after world version;
  • evidence;
  • authority;
  • provenance。

106. Receipts 不等 Truth

OperationalReceiptEpistemicTruth.\boxed{ OperationalReceipt \neq EpistemicTruth. }

107. Durable Execution

長流程必須可:

  • checkpoint;
  • interrupt;
  • resume;
  • replay;
  • recover。

108. Runtime Process Death 不應抹除 Task

StateSurvivesProcessDeath.\boxed{ StateSurvivesProcessDeath. }

109. Checkpoint

CheckpointkCheckpoint_k

保存可恢復執行狀態。


110. Checkpoint 不等 Canonical World Snapshot

workflow execution state 和 world state 仍要分開。


111. Replay

execution replay 可以重建 workflow control state。


112. Replay 不等 Reality Replay

WorkflowReplayRealityReplay.\boxed{ WorkflowReplay \neq RealityReplay. }

113. Deterministic Replay 需要限制 Side Effect

重放時不可任意再次執行外部 irreversible effect。


114. External Effect 應隔離成 Activity / Effect Boundary

使 replay 能辨認:

這步已做過。


115. Checkpoint Fork

對純 internal / simulated plan,可從 checkpoint 建 alternative branch。


116. Forked Execution 不等 Independent Evidence

ExecutionForkEvidenceIndependence.\boxed{ ExecutionFork \neq EvidenceIndependence. }

117. Mother Runtime

承接:

RM=(PS,PC,PE,PO,PG).\mathfrak R_M = ( \mathcal P_S, \mathcal P_C, \mathcal P_E, \mathcal P_O, \mathcal P_G ).

118. State Plane

保存:

  • world;
  • memory;
  • tasks;
  • commitments;
  • lineage;
  • plan state。

119. Cognitive Control Plane

負責:

  • trigger;
  • task compilation;
  • layer routing;
  • model / agent choice;
  • replan;
  • stop。

120. Effect Plane

只負責真正 external effect。


121. Observability / Recovery Plane

保存:

  • trace;
  • checkpoint;
  • metrics;
  • replay;
  • fault recovery;
  • receipts。

122. Governance Plane

保存:

  • identity;
  • permission;
  • policy;
  • approval;
  • veto;
  • authority ceiling。

123. Control Plane 不包含 Governance Plane

PC⊅PG.\boxed{ \mathcal P_C \not\supset \mathcal P_G. }

124. Effect Plane 不包含 State Authority

PE⊅PS.\boxed{ \mathcal P_E \not\supset \mathcal P_S. }

executor 不應自行改 canonical memory/world semantics。


125. Observability Plane 盡量 Append-Only

避免事後重寫 execution history。


126. Thin Mother Runtime

ThinCore+RichSpecializedLayers.\boxed{ ThinCore + RichSpecializedLayers. }

127. Thin Core 最小責任

  1. ingest events;
  2. construct task contract;
  3. project / reconstruct relevant state;
  4. coordinate layer routers;
  5. maintain plan lifecycle;
  6. enforce authority boundaries;
  7. checkpoint / recover;
  8. collect receipts;
  9. decide stop / wait / replan。

128. Mother Core 不做所有 Specialist Work

它不應自己:

  • 當 theorem prover;
  • 當 image renderer;
  • 當 database;
  • 當 robot driver;
  • 當所有 world simulator。

129. Mother Model 不等 Mother Runtime

MotherModelMotherRuntime.\boxed{ MotherModel \neq MotherRuntime. }

130. Model Replaceability

LiLjL_i\rightarrow L_j

不應失去:

  • world;
  • task;
  • memory;
  • plan;
  • authority;
  • history。

131. Agent Replaceability

Agent 可以:

  • spawn;
  • retire;
  • replace。

Mother Runtime identity 不必消失。


132. Workflow Replaceability

workflow 是 runtime 產生的局部 execution object。


133. Logical Unity + Physical Modularity

LogicalUnity+PhysicalModularity.\boxed{ LogicalUnity + PhysicalModularity. }

134. Mother Identity

可由:

IM=(SystemId,StateLineage,MemoryLineage,GoalLineage,AuthorityRoot,RuntimeVersion).I_M = ( SystemId, StateLineage, MemoryLineage, GoalLineage, AuthorityRoot, RuntimeVersion ).

135. Event-Driven Runtime

系統主要靠:

ete_t

觸發更新,而不是 periodic giant prompt。


136. Event Type

可以包括:

  • user intent;
  • external world event;
  • timer;
  • provider failure;
  • new evidence;
  • authority change;
  • resource change;
  • stale state;
  • task completion。

137. Event → Meta-State Projection

Mother Runtime 不需要每次看完整世界。


138. Meta-State

ZtM=Project(Wt,Mt,Taskst,Unknownt,Resourcest).Z_t^M = Project( W_t, M_t, Tasks_t, Unknown_t, Resources_t ).

139. Meta-Control

ZtMCtmeta.Z_t^M \rightarrow \mathcal C_t^{meta}.

140. Meta-Control Output

不是直接 physical effect。

而是:

  • plan delta;
  • routing request;
  • authority request;
  • wait;
  • stop。

141. Execution Orchestration State

EOSt=It,Atcap,Ptplan,Rtroute,Xtrun,Vtexec,Gtauth,Btres,Utexec,Htreceipt.\boxed{ EOS_t = \left\langle \mathfrak I_t, \mathfrak A_t^{cap}, \mathfrak P_t^{plan}, \mathfrak R_t^{route}, \mathfrak X_t^{run}, \mathfrak V_t^{exec}, \mathfrak G_t^{auth}, \mathfrak B_t^{res}, \mathfrak U_t^{exec}, \mathfrak H_t^{receipt} \right\rangle. }

142. WCO Runtime Master Loop

etUpdateStateTaskContractResolveCapabilitiesCompileEPGRouteLayersMaterializeRDRExecuteObserveVerify/ReconcileCommit/Stop/ReplanEOSt+1.\boxed{ \begin{aligned} e_t &\rightarrow \mathsf{UpdateState} \\ &\rightarrow \mathsf{TaskContract} \\ &\rightarrow \mathsf{ResolveCapabilities} \\ &\rightarrow \mathsf{CompileEPG} \\ &\rightarrow \mathsf{RouteLayers} \\ &\rightarrow \mathsf{MaterializeRDR} \\ &\rightarrow \mathsf{Execute} \\ &\rightarrow \mathsf{Observe} \\ &\rightarrow \mathsf{Verify/Reconcile} \\ &\rightarrow \mathsf{Commit/Stop/Replan} \\ &\rightarrow EOS_{t+1}. \end{aligned} }

143. 這個 Loop 不是永遠線性

Plan graph 可:

  • branch;
  • parallel;
  • wait;
  • fork;
  • return;
  • escalate。

144. Cross-Layer Replan

某 projection failure 可能回到:

ProjectionObservationComputation.Projection \rightarrow Observation \rightarrow Computation.

145. Memory Failure 可能要求 Reobserve

若 memory stale:

MemoryObservation.Memory \rightarrow Observation.

146. World Disagreement 可能要求 Fork

WWa,Wb.W \rightarrow W_a,W_b.

147. Verification Failure 可能要求 Recompute

VerifyFailComputeRoute.VerifyFail \rightarrow ComputeRoute'.

148. Authority Failure 不應靠 Reroute 繞過

AuthorityBlock⇏FindAnotherProviderToBypass.\boxed{ AuthorityBlock \not\Rightarrow FindAnotherProviderToBypass. }

149. Policy Hierarchy

可概念化:

System>Organization>Project>Task>Invocation>ProviderDefault.System > Organization > Project > Task > Invocation > ProviderDefault.

低層不能放寬高層 hard prohibition。


150. Hard Gate 與 Optimization 分離

AdmissibilityOptimization.\boxed{ Admissibility \prec Optimization. }

151. Soft Optimization

通過 hard gates 後,可比較:

  • quality;
  • latency;
  • cost;
  • privacy;
  • energy;
  • localness;
  • verifiability。

152. Pareto Routing

不同 plan 可形成:

PPareto.\mathcal P_{Pareto}.

153. 不必永遠一個 Scalar Score

政策才決定 scalarization。


154. Runtime Unknown Registry

若:

  • no capability;
  • no bridge;
  • no authority;
  • no evidence;
  • no physical carrier;

應記:

UNKNOWN/BLOCKED.UNKNOWN/BLOCKED.

155. Unknown 不等 Failure

UnknownFailed.\boxed{ Unknown \neq Failed. }

156. Capability Gap

若:

Candidates(q)=,Candidates(q) = \varnothing,

形成:

CapabilityGap.CapabilityGap.

157. Gap 可以觸發 Capability Construction

但只在 policy 允許的 sandbox。


158. Capability Promotion 不應自動

成功一次:

⇏Trusted.\not\Rightarrow Trusted.

159. Promotion Gate

至少:

  • repeated tests;
  • verifier;
  • failure modes;
  • version;
  • provenance;
  • authority ceiling;
  • rollback / revocation path。

160. Revocation

capability 可以:

ActiveRevoked.Active \rightarrow Revoked.

161. Revocation Propagation

已排程但未執行的 plans 應 re-evaluate。


162. Execution State 與 Semantic Definition 分離

RDR 的 runtime observation:

TrunT_{run}

不能直接修改 CAIR definition。


163. Execution Evidence → Proposal

TrunΔdefinitionproposal.T_{run} \rightarrow \Delta_{definition}^{proposal}.

164. Proposal 再經治理

只有通過才形成新 version。


165. Index Lag 是合法 Operational State

MSSP 落後 CAIR:

Lag>0Lag>0

需要 rebuild。


166. Index Lag 不應改變 Canonical Definition

IndexFailureAuthorityRollback.\boxed{ IndexFailure \neq AuthorityRollback. }

167. Runtime Cache 不是 Authority

RuntimeCacheCapabilityDefinition.\boxed{ RuntimeCache \neq CapabilityDefinition. }

168. Plan Cache 也不是 Task Truth

cache 需要 version / policy check。


169. Execution History

至少保存:

  • task contract;
  • plan versions;
  • routes;
  • provider snapshots;
  • checkpoints;
  • failures;
  • effect receipts;
  • state commits。

170. History Supports Recovery

不是只有 observability。


171. Durable Runtime Engineering Interface

現有 workflow / agent frameworks 已能提供部分:

  • graph execution;
  • checkpoint;
  • interrupt;
  • resume;
  • HITL;
  • tracing;
  • durable event history。

172. 這些不是 WCO 全部

因為 WCO 額外要求:

  • world-family semantics;
  • epistemic domain routing;
  • projection routing;
  • capability semantic authority;
  • cross-layer receipts;
  • physical-effect separation。

173. MVP:Thin WCO Runtime

第一版只做:

  1. Task Contract;
  2. MSSP Capability Registry;
  3. Execution Plan Graph;
  4. RDR local / remote provider;
  5. WFS four-world scheduler;
  6. WRMF memory router;
  7. verification gate;
  8. authority gate;
  9. event log / checkpoint;
  10. runtime receipt bundle。

174. MVP 不需要 Autonomous Capability Construction

先只:

Select+Compose.Select + Compose.

Construct 可以第二階段。


175. MVP Worlds

WA,WB,WC,WN.W_A,W_B,W_C,W_N.

176. MVP Capability Examples

  • simulate.policy;
  • compare.worlds;
  • observe.risk;
  • retrieve.evidence;
  • verify.constraint;
  • project.risk;
  • render.report。

177. MVP Plan

IntentTICMSSPWFSComputeObserveVerifyProjectStop.Intent \rightarrow TIC \rightarrow MSSP \rightarrow WFS \rightarrow Compute \rightarrow Observe \rightarrow Verify \rightarrow Project \rightarrow Stop.

178. MVP RDR

provider:

  • local Python;
  • local database;
  • one remote worker。

179. MVP Failure Injection

故意:

  • provider down;
  • stale memory;
  • capability revoked;
  • budget exhausted;
  • verifier fail;
  • authority block。

180. Experiment 1 — Manual Orchestration vs Thin Runtime

測:

  • human routing operations;
  • completion;
  • recovery;
  • context cost;
  • plan auditability。

181. Experiment 2 — One Scheduler vs Layered Routing

比較 scheduler collapse 與 LRRF。


182. Experiment 3 — Fixed Plan vs Controlled Replan

注入 provider failure。


183. Experiment 4 — Task Fidelity Fence

故意提供更簡單但錯題 capability。

runtime 必須拒絕。


184. Experiment 5 — Authority Bypass Attack

讓 provider / router 建議替代 path 繞過 authority。

runtime 必須 block。


185. Experiment 6 — Checkpoint Recovery

process kill 後恢復。


186. Experiment 7 — External Effect Idempotency

重放 plan。

外部 effect 不得無限重複。


187. Experiment 8 — Capability Promotion

一次成功 vs 多次 validated。

測 promotion gate。


188. Experiment 9 — Receipt Reconstruction

只用 RRB 重建:

為什麼這個 action 發生?


189. Experiment 10 — Model Replacement

中途換 model。

Task / plan / world / memory / authority continuity 應保持。


190. 可反駁性

本文會被削弱,如果:

  1. Layered routers 比單一 planner 沒有任何 correctness / audit / modularity 優勢;
  2. Task Identity Contract 無法降低 silent task drift;
  3. MSSP / CAIR / RDR 分離只增加 overhead;
  4. durable checkpoint / receipt 對 long-running tasks 沒有 recovery 價值;
  5. controlled replan 無法改善 provider / resource failure;
  6. authority/effect separation 無法降低錯誤 side effects;
  7. Thin Mother Runtime 比普通 agent loop 在代表性 long-horizon tasks 中完全等效;
  8. capability promotion / revocation lifecycle 沒有工程價值。

191. 外部工程接口

2026 年 Microsoft Agent Framework 已提供 graph-based workflow、state management、checkpointing、human-in-the-loop 與 workflow-as-agent 等能力,顯示 Agent 與 Workflow graph 已可被清楚分離。

LangGraph 的 persistence / checkpoint 機制可保存 thread state、支援 interrupt / resume、time travel、fault recovery 與 checkpoint fork,證明 durable graph execution 可直接成為上層 AI runtime 的工程 substrate。

Temporal 以 durable Event History 與 deterministic replay 恢復 Workflow Execution,並把 Commands、Events、Activities 與 durable workflow state 分離,是本文 Effect Boundary、event history 與 resume 設計的重要工程參照。

OpenAI Agents SDK 則提供 agents、handoffs、guardrails、sessions、human-in-the-loop 與 tracing,說明輕量 Agent primitives、run-level trace 與 validation layer 已能作為 WCO specialist layer 的下層實現之一。

本文不宣稱上述 framework 已實作 WCO。本文的額外問題是:

如何在 durable agent / workflow substrate 之上,加入 world-family governance、epistemic qualification、multi-layer routing、capability authority、projection routing、physical-effect fencing 與 reconstructive memory?


192. 本文不主張什麼

本文不主張:

  1. MSSP 應取代所有 service registry;
  2. RDR 應取代 OS scheduler;
  3. CAIR 應成為整個 WCO 唯一資料庫;
  4. Mother Runtime 應擁有所有 authority;
  5. 所有 tasks 都需要 world family;
  6. 所有 routing 都應由 LLM 決定;
  7. deterministic resolver 可解所有 semantic mapping;
  8. capability composition 必然成功;
  9. generated capability 可自動 trusted;
  10. one scheduler 可以安全吞掉所有 layer routers;
  11. checkpoint 等於 world truth;
  12. workflow replay 等於 reality replay;
  13. retry 等於 rollback;
  14. compensation 等於 physical undo;
  15. authorized dispatch 等於 desired effect;
  16. receipt 等於 truth;
  17. exactly-once external effect 在所有系統中可被免費保證;
  18. Mother Runtime 必須是一個單一 process;
  19. Mother Model 必須是單一最大模型;
  20. Paper 09 已完成 production Global AI runtime。

193. 核心非同一性

IntentTaskContractPlanExecutionEffectObservedOutcome.\boxed{ Intent \neq TaskContract \neq Plan \neq Execution \neq Effect \neq ObservedOutcome. } CapabilityIdentityProviderIdentity.\boxed{ CapabilityIdentity \neq ProviderIdentity. } UIAuthorityIndexRuntime.\boxed{ UI \neq Authority \neq Index \neq Runtime. } MSSPWFSComputeRouterMemoryRouterObserverRouterCPCRDR.\boxed{ MSSP \neq WFS \neq ComputeRouter \neq MemoryRouter \neq ObserverRouter \neq CPC \neq RDR. } RetryFallbackRerouteRecomposeReframeEscalate.\boxed{ Retry \neq Fallback \neq Reroute \neq Recompose \neq Reframe \neq Escalate. } PauseWaitStopCancelRollback.\boxed{ Pause \neq Wait \neq Stop \neq Cancel \neq Rollback. } AuthorizedDispatchPhysicalEffectWorldStateCommit.\boxed{ AuthorizedDispatch \neq PhysicalEffect \neq WorldStateCommit. } OperationalReceiptEpistemicTruth.\boxed{ OperationalReceipt \neq EpistemicTruth. } MotherModelMotherRuntime.\boxed{ MotherModel \neq MotherRuntime. }

194. 核心母式一:Task Identity Contract

Iq=TaskId,Goal,Inputs,Outputs,Invariants,Constraints,RequiredQualification,Risk,Budget,AllowedEffects,Authority,StopConditions.\boxed{ \mathfrak I_q = \left\langle TaskId, Goal, Inputs, Outputs, Invariants, Constraints, RequiredQualification, Risk, Budget, AllowedEffects, Authority, StopConditions \right\rangle. }

195. 核心母式二:Execution Plan Graph

GtEP=(VtEP,EtEP,ΓtEP,ΣtEP,HtEP).\boxed{ \mathcal G_t^{EP} = ( V_t^{EP}, E_t^{EP}, \Gamma_t^{EP}, \Sigma_t^{EP}, H_t^{EP} ). }

196. 核心母式三:Layered Runtime Routing Fabric

LRRF={RMSSP,RWFS,Rcomp,Rmem,Robs,Rdom,Rproj,Rphys,RRDR}.\boxed{ LRRF = \{ R_{MSSP}, R_{WFS}, R_{comp}, R_{mem}, R_{obs}, R_{dom}, R_{proj}, R_{phys}, R_{RDR} \}. }

197. 核心母式四:Execution Orchestration State

EOSt=It,Atcap,Ptplan,Rtroute,Xtrun,Vtexec,Gtauth,Btres,Utexec,Htreceipt.\boxed{ EOS_t = \left\langle \mathfrak I_t, \mathfrak A_t^{cap}, \mathfrak P_t^{plan}, \mathfrak R_t^{route}, \mathfrak X_t^{run}, \mathfrak V_t^{exec}, \mathfrak G_t^{auth}, \mathfrak B_t^{res}, \mathfrak U_t^{exec}, \mathfrak H_t^{receipt} \right\rangle. }

198. 核心母式五:Runtime Receipt Bundle

RRB=IntentReceipt,ResolutionReceipt,PlanReceipt,RoutingReceipts,MaterializationReceipt,ExecutionReceipts,VerificationReceipt,EffectReceipt,StateCommitReceipt.\boxed{ RRB = \left\langle IntentReceipt, ResolutionReceipt, PlanReceipt, RoutingReceipts, MaterializationReceipt, ExecutionReceipts, VerificationReceipt, EffectReceipt, StateCommitReceipt \right\rangle. }

199. 核心母式六:Reality Effect Loop

ProposalValidateAuthorizeDispatchObserveOutcomeReconcileWorldStateCommit.\boxed{ Proposal \rightarrow Validate \rightarrow Authorize \rightarrow Dispatch \rightarrow ObserveOutcome \rightarrow Reconcile \rightarrow WorldStateCommit. }

200. 核心母式七:WCO Runtime Master Loop

etUpdateStateTaskContractResolveCapabilitiesCompileEPGRouteLayersMaterializeRDRExecuteObserveVerify/ReconcileCommit/Stop/ReplanEOSt+1.\boxed{ \begin{aligned} e_t &\rightarrow UpdateState \rightarrow TaskContract \rightarrow ResolveCapabilities \\ &\rightarrow CompileEPG \rightarrow RouteLayers \rightarrow MaterializeRDR \\ &\rightarrow Execute \rightarrow Observe \rightarrow Verify/Reconcile \\ &\rightarrow Commit/Stop/Replan \rightarrow EOS_{t+1}. \end{aligned} }

201. 結論:真正的類全域 AI 執行能力,是把「可以做」變成「知道如何安全地做」

前八篇已經給了類全域 AI 很多能力:

  • world family;
  • global computation;
  • observation;
  • epistemic qualification;
  • projection;
  • physical realization;
  • reconstructive memory。

但如果沒有 Execution Layer,這些能力只是一堆模組。

真正的 runtime 必須回答:

使用者/AI 現在到底想做什麼?

任務 identity 是什麼?

哪些 transformations 可以接受?

目前有哪些 capabilities?

應直接選一個、組合多個,還是建立 candidate?

哪些 worlds 值得繼續?

哪種 computation 要跑?

哪些 memory 需要取?

要觀察什麼?

哪個 epistemic debt 還沒補?

用什麼 projection?

物理 carrier 做得到嗎?

由哪個 provider 執行?

權限夠嗎?

失敗後是 retry、fallback、reroute、reframe 還是 stop?

什麼時候任務已經足夠完成?

真實作用到底發生了沒有?

為什麼當時做了這個決定?

這就是為什麼:

Capability AvailabilityExecutable Intelligence.\boxed{ \text{Capability Availability} \neq \text{Executable Intelligence}. }

真正的執行智能更接近:

Task Fidelity+Capability Resolution+Layered Routing+Durable Planning+Authority Fencing+Verification+Recovery+History.\boxed{ \text{Task Fidelity} + \text{Capability Resolution} + \text{Layered Routing} + \text{Durable Planning} + \text{Authority Fencing} + \text{Verification} + \text{Recovery} + \text{History}. }

同時,本文拒絕建立一個「什麼都懂、什麼都管、什麼都執行」的超級 Mother Agent。

更合理的是:

Thin Mother Runtime+Specialized World / Compute / Memory / Observation / Projection / Execution Layers.\boxed{ \text{Thin Mother Runtime} + \text{Specialized World / Compute / Memory / Observation / Projection / Execution Layers}. }

Mother Runtime 的強大不來自吞掉所有模組,而來自它能維持:

logical unity+semantic separation+durable continuity.\boxed{ \text{logical unity} + \text{semantic separation} + \text{durable continuity}. }

因此:

A Global-Like AI runtime is not a single super-agent loop.\boxed{ \text{A Global-Like AI runtime is not a single super-agent loop.} }

而是:

a durable, authority-bounded meta-runtime that compiles intent into typed plans, delegates decisions to specialized routing layers, materializes capabilities through RDR, and closes execution through observation, verification, receipts, and state reconstruction.\boxed{ \text{a durable, authority-bounded meta-runtime that compiles intent into typed plans, delegates decisions to specialized routing layers, materializes capabilities through RDR, and closes execution through observation, verification, receipts, and state reconstruction}. }

到此 WCO 已從能力鏈進入真正 runtime:

WtG+CtWF+OtG+DtWCO+PtG+PRSt+MtWCOEOSt.\boxed{ \mathfrak W_t^G + \mathfrak C_t^{WF} + \mathfrak O_t^G + \mathfrak D_t^{WCO} + \mathfrak P_t^G + \mathsf{PRS}_t + \mathfrak M_t^{WCO} \rightarrow \mathsf{EOS}_t. }

下一篇將進入:

Paper 10

驗證層:多世界、多投影、多模型為何仍不等於真理

將正式集中處理:

  • evidence independence;
  • common-mode failure;
  • multi-world consensus;
  • multi-model consensus;
  • projection agreement;
  • verifier diversity;
  • certificate composition;
  • local / global validity;
  • adversarial counterexample;
  • verification debt;
  • provenance closure;
  • truth-status separation。

202. 下一篇接口

Paper 10 將處理:

  • WorldCount != EvidenceCount;
  • model agreement != independent replication;
  • observer agreement != truth;
  • projection convergence != verification;
  • common-mode dependency;
  • evidence independence graph;
  • verifier independence;
  • counterexample world;
  • certificate lattice;
  • local certificate / global certificate;
  • evidence transport;
  • simulation-to-reality qualification;
  • adversarial reprojection;
  • verification closure;
  • falsification / reopen triggers。

參考文獻與內部前置研究

EveMissLab / Neo.K

  1. Neo.K × Aletheia,《Hyperconnected MSSP–RDR Runtime Architecture》,2026。
  2. Neo.K × Aletheia,《MSSP × RDR Runtime Architecture v0.1》,2026。
  3. Neo.K × Aletheia,《權威結構與執行派發:格子語言、CAIR、MSSP × RDR 的接合》,2026。
  4. Neo.K × Aletheia,《RDSS Runtime Architecture》,2026。
  5. Neo.K × Aletheia,《母 AI Runtime:從模型到持續認知核心》,2026。
  6. Neo.K × Aletheia,《SWFR Paper 02:模擬世界族執行時》,2026。
  7. Neo.K × Aletheia,《WCO Paper 01–08》,2026。
  8. Neo.K × Aletheia,《Multi-Representation Autonomous Memory Fabric》,2026。
  9. Neo.K × Aletheia,《Global Computation Methodology》,2026。
  10. Neo.K × Aletheia,《Global Observer Series C》,2026。

External Engineering Interfaces

  1. Microsoft. Agent Framework — Workflow Concepts / Workflows, updated 2026-08-25.
  2. LangChain. LangGraph Persistence, Interrupts, and Time Travel, 2026.
  3. Temporal Technologies. Temporal Event History, Replay, and Durable Workflow Execution, 2026.
  4. OpenAI. Agents SDK — Agents, Sessions, Handoffs, Guardrails, Human-in-the-Loop, and Tracing, 2026.
  5. Fowler, M. Event Sourcing, 2005.
  6. Kleppmann, M. Designing Data-Intensive Applications, O'Reilly, 2017.

Paper 09 狀態:COMPLETE v0.1
下一篇:Paper 10 — 驗證層:多世界、多投影、多模型為何仍不等於真理
Canonical source:UTF-8 Markdown;數學 delimiter 僅使用 $...$$$...$$