← Archive
lm-003668 · 2026-09

全域計算、局部顯現:投影路由、解析度場與物化決策

下載 MD 檔 ⬇

PNCW Paper 05

全域計算、局部顯現:投影路由、解析度場與物化決策

Compute Globally, Materialize Selectively:

Projection Routing, Resolution Fields, and Materialization Decisions

版本:v0.1
日期:2026-08-28
系列:Projection-Native Computational World Series / 投影原生計算世界系列
定位:Series Paper 05 / Global Routing, Resolution, and Materialization Governance Layer
依賴:PNCW Paper 00–04、Global Computation Methodology(GCM)Series、GCM Phase A Reference Runtime、GCM Phase B Deterministic Compute Allocator
作者: Neo.K
機構: EveMissLab/一言諾科技有限公司


摘要

本文延續 Projection-Native Computational World(PNCW)Series,建立 PNCW 的 Global Compute / Selective Materialization Governance Layer,回答前四篇尚未統一解決的 Runtime 問題:

當 canonical world、Context World、projected carrier、visual canvas 與外部 resources 同時存在時,究竟由誰決定「現在要算什麼、在哪裡算、用什麼 representation、用多高解析度、哪些部分值得 materialize、哪些應維持 dormant、哪些應該被投影給哪個 observer」?

本文以 Global Computation Methodology(GCM)作為主要治理框架,將其核心原則:

ComputationObservation\boxed{ \text{Computation} \neq \text{Observation} } Global DependencyFull Materialization\boxed{ \text{Global Dependency} \neq \text{Full Materialization} } Recursive GlobalityRecursive Full Expansion\boxed{ \text{Recursive Globality} \neq \text{Recursive Full Expansion} }

以及:

Finite Active Realization+Unbounded Extensibility\boxed{ \text{Finite Active Realization} + \text{Unbounded Extensibility} }

正式嵌入 PNCW。

本文將 PNCW Runtime state 定義為:

Wt=Wt,Dt,P,L,Λt,Ct,St,Πt,Ht.\boxed{ \mathfrak W_t = \left\langle W_t, \mathcal D_t, \mathfrak P, \mathfrak L, \Lambda_t, \mathcal C_t, \mathcal S_t, \Pi_t, \mathcal H_t \right\rangle. }

其中:

  • WtW_t:canonical world state;
  • Dt\mathcal D_t:active/potential computational domains;
  • P\mathfrak P:computational form / representation space;
  • L\mathfrak L:transition-law family;
  • Λt\Lambda_t:resolution / materialization field;
  • Ct\mathcal C_t:global constraints / couplings;
  • St\mathcal S_t:scheduling / routing / allocation;
  • Πt\Pi_t:observer / carrier / context projections;
  • Ht\mathcal H_t:history / provenance。

本文進一步將 GCM 的 Materialization Policy 形式化為:

MatPolicyt:(Wt,q,O,R,B,Risk)Mt,\boxed{ \mathsf{MatPolicy}_t : ( W_t, q, O, R, B, Risk ) \rightarrow \mathcal M_t, }

其中 Mt\mathcal M_t 是此時此刻值得 active realization 的 representation set。

本文特別區分:

λcomputeλobserveλcarrierλrender.\boxed{ \lambda^{compute} \neq \lambda^{observe} \neq \lambda^{carrier} \neq \lambda^{render}. }

也就是 compute resolution、observer resolution、carrier resolution 與 render resolution 不必相同。

本文亦將 GCM Phase B 的 deterministic allocation 思想納入 PNCW,使「顯示哪個 projection」不再由 UI 任意決定,而可以經:

ConstraintFeasible PlansObjective EvaluationPareto FrontierExplicit Policy Selection\boxed{ \text{Constraint} \rightarrow \text{Feasible Plans} \rightarrow \text{Objective Evaluation} \rightarrow \text{Pareto Frontier} \rightarrow \text{Explicit Policy Selection} }

選出一個可稽核的 active realization plan。

本文最終提出:

Global CoherenceGlobal Simultaneous Materialization.\boxed{ \text{Global Coherence} \neq \text{Global Simultaneous Materialization}. }

以及:

Compute Globally, Materialize Selectively, Observe Relatively.\boxed{ \text{Compute Globally, Materialize Selectively, Observe Relatively}. }

這使 PNCW 的「一口氣看到」獲得真正的全域 runtime 基礎:不是把世界全部一次 render,而是讓 canonical world 維持全域一致,讓 Runtime 依 task、dependency、budget、risk、observer 與 resource state,選擇當下最值得 active realization 的部分。

關鍵詞: GCM、Selective Materialization、Resolution Field、Projection Routing、Finite Active Realization、Pareto Allocation、Global Computation、PNCW


0. 研究目的

PNCW Paper 00–04 已經建立:

MtotalCqactiveEkPkVq,k.\boxed{ \mathcal M^{total} \rightarrow C_q^{active} \rightarrow \mathcal E_k \rightarrow P_k \rightarrow V_{q,k}. }

但仍缺一個更高層問題:

誰決定這條 projection chain 什麼時候啟動、啟動多大、走哪一條 representation route、用多少 resource?

本文引入:

GlobalRuntimeGovernor.\boxed{ \mathsf{GlobalRuntimeGovernor}. }

1. Global 不等於 Everything Everywhere

PNCW 接受 GCM 的核心重定義:

Global=coherence relative to a designated World boundary.\boxed{ \text{Global} = \text{coherence relative to a designated World boundary}. }

它不等於:

  • all-to-all;
  • same algorithm;
  • same representation;
  • same timestep;
  • full materialization;
  • global barrier。

2. Global Computation

對 domains:

D1,,Dn,D_1,\ldots,D_n,

每個 domain 可使用:

pi(t)P,p_i(t)\in\mathfrak P, i(t)L.\ell_i(t)\in\mathfrak L.

局部演化:

Φipi,i.\Phi_i^{p_i,\ell_i}.

全域演化:

ΦG(t)=ComposeCG(t)(Φ1p1,1,,Φnpn,n).\boxed{ \Phi_G(t) = \operatorname{Compose}_{\mathcal C_G(t)} \left( \Phi_1^{p_1,\ell_1}, \ldots, \Phi_n^{p_n,\ell_n} \right). }

並:

Wt+1=ΦG(t)(Wt).W_{t+1} = \Phi_G(t)(W_t).

3. Computational Form Space

令:

P={sequence,graph,matrix,tensor,symbolic,simulation,image-carrier,canvas,}.\boxed{ \mathfrak P = \{ \text{sequence}, \text{graph}, \text{matrix}, \text{tensor}, \text{symbolic}, \text{simulation}, \text{image-carrier}, \text{canvas}, \ldots \}. }

PNCW 不要求所有 domain 使用同一 representation。


4. Representation Routing

對 domain DiD_i 與 task (q):

pi=RouteRep(Di,q,Constraints,Cost,Risk).\boxed{ p_i^\star = \mathsf{RouteRep} ( D_i, q, Constraints, Cost, Risk ). }

因此:

Global ComputationOne Representation Everywhere.\boxed{ \text{Global Computation} \neq \text{One Representation Everywhere}. }

5. Transition-Law Family

不同 domain 可以使用不同:

iL.\ell_i \in \mathfrak L.

例如:

  • deterministic update;
  • probabilistic transition;
  • graph propagation;
  • optimization step;
  • symbolic rewrite;
  • neural inference;
  • event-driven transition。

6. Domain

定義 domain:

Di=State,Representation,Law,Resolution,Dependencies,Authority,History.\boxed{ D_i = \left\langle State, Representation, Law, Resolution, Dependencies, Authority, History \right\rangle. }

Domain 是 GCM / PNCW 的基本 routing unit。


7. Domain 不等於 Process

一個 process 可以跨多 domain。

一個 domain 也可以由多 executor 共同處理。

因此:

DomainProcess.\boxed{ \text{Domain} \neq \text{Process}. }

8. Global Dependency 不等於 Full Materialization

如果:

DiDjD_i \rightarrow D_j

存在 dependency,

不表示:

Di,DjD_i, D_j

所有 detail 必須同時 materialize。

因此:

Global DependencyFull Materialization.\boxed{ \text{Global Dependency} \neq \text{Full Materialization}. }

9. Materialization

定義:

Materialize:XXresident.\boxed{ \mathsf{Materialize} : X \rightarrow X^{resident}. }

其中 resident 可以是:

  • RAM;
  • VRAM;
  • local DB;
  • HDSRC tile;
  • Context Packet;
  • Canvas region;
  • rendered viewport。

10. Materialization 不等於 Existence

若:

XMt,X\notin M_t,

不表示:

XX

不存在。

所以:

Not MaterializedNonexistent.\boxed{ \text{Not Materialized} \neq \text{Nonexistent}. }

11. 五種 Runtime State

對 object (X):

RuntimeState(X){ACTIVE,MATERIALIZED,DORMANT,ARCHIVED,POTENTIAL}.\boxed{ \mathsf{RuntimeState}(X) \in \{ \mathsf{ACTIVE}, \mathsf{MATERIALIZED}, \mathsf{DORMANT}, \mathsf{ARCHIVED}, \mathsf{POTENTIAL} \}. }

12. Active

XAtX\in A_t

表示目前參與 computation。


13. Materialized

XMtX\in M_t

表示細節已展開、可直接 access / render。


14. Dormant

存在 canonical state,但暫不 high-frequency update。


15. Archived

只保存:

  • checkpoint;
  • seed;
  • summary;
  • history pointer;
  • compressed carrier。

16. Potential

structure 尚未 materialize,但 grammar / generator / rule 允許未來生成。


17. Active 不等於 Materialized

一般:

AtMt.\boxed{ A_t \neq M_t. }

可能:

XAt,XMt,X\in A_t, \quad X\notin M_t,

例如 hidden low-resolution solver。

也可能:

XMt,XAt,X\in M_t, \quad X\notin A_t,

例如 frozen UI snapshot。


18. Finite Active Realization

真實 Runtime 必須:

At<t.\boxed{ |A_t|<\infty \qquad \forall t. }

更強可要求:

AtBA.|A_t| \le B_A.

19. Unbounded Extensibility

World 可以:

W0W1W2W_0 \rightarrow W_1 \rightarrow W_2 \rightarrow \cdots

而不預設固定最大深度。

但:

Unbounded Extensibility⇏Infinite Simultaneous Computation.\boxed{ \text{Unbounded Extensibility} \not\Rightarrow \text{Infinite Simultaneous Computation}. }

20. Recursive Globality / Recursive Full Expansion Non-Collapse

Recursive GlobalityRecursive Full Expansion.\boxed{ \text{Recursive Globality} \neq \text{Recursive Full Expansion}. }

這直接適用於:

  • Context World;
  • Recursive Canvas;
  • HDSRC multi-scale carrier;
  • nested simulation;
  • project graph。

21. Bounded Frontier

定義 active frontier:

At\boxed{ \partial A_t }

為目前 active world 與 potential / dormant world 的邊界。

Runtime 只需管理 bounded frontier。


22. Rolling Horizon

對長期 evolution:

Ht=[t,t+h].H_t = [t,t+h].

只 materialize有限 rolling horizon:

Ht<.\boxed{ |H_t|<\infty. }

23. Checkpoint Reconstruction

如果某 region dormant / archived:

XtX_t

可以從:

Checkpoint+Seed+EventsCheckpoint + Seed + Events

重建。

因此:

PersistenceFull Residency.\boxed{ \text{Persistence} \neq \text{Full Residency}. }

24. Resolution Field

本文引入:

Λt:DtLλ.\boxed{ \Lambda_t : \mathcal D_t \rightarrow \mathcal L_\lambda. }

其中:

Λt(Di)=λi(t).\Lambda_t(D_i) = \lambda_i(t).

25. Resolution 是 Field,不是 Global Constant

不同 domain:

λiλj\lambda_i \neq \lambda_j

完全合法。

因此 World 可以同時:

  • one domain coarse;
  • one domain fine;
  • one dormant;
  • one fully materialized;
  • one checkpoint-only。

26. Compute Resolution

定義:

λicompute.\lambda_i^{compute}.

表示 domain 真正計算所需 fidelity。


27. Observe Resolution

定義:

λiobserve.\lambda_i^{observe}.

表示 observer projection fidelity。


28. Carrier Resolution

定義:

λicarrier.\lambda_i^{carrier}.

表示 projected carrier 的 spatial / semantic / quantization fidelity。


29. Render Resolution

定義:

λirender.\lambda_i^{render}.

表示最後 viewport / human output render fidelity。


30. 四種 Resolution Non-Collapse

因此:

λcomputeλobserveλcarrierλrender.\boxed{ \lambda^{compute} \neq \lambda^{observe} \neq \lambda^{carrier} \neq \lambda^{render}. }

它們可以相等,但不是理論必然。


31. Hidden High-Fidelity Compute

可以:

λcompute>λobserve.\lambda^{compute} > \lambda^{observe}.

例如內部高精度 simulation,但 UI 只顯示 summary。


32. High-Resolution View / Low-Fidelity Compute

也可以:

λrender>λcompute.\lambda^{render} > \lambda^{compute}.

例如 frozen data 以高解析圖形 presentation,但 underlying compute 已停止。


33. Adaptive Refinement

如果:

Risk(Di)Risk(D_i)

或:

Uncertainty(Di)Uncertainty(D_i)

升高,

可以:

λi.\lambda_i \uparrow.

34. De-Refinement

如果 relevance 下降:

λi.\lambda_i \downarrow.

因此:

Refinement is reversible runtime policy.\boxed{ \text{Refinement} \text{ is reversible runtime policy}. }

35. Materialization Policy

定義:

MatPolicyt:(Wt,q,O,R,B,Risk)Mt.\boxed{ \mathsf{MatPolicy}_t : ( W_t, q, O, R, B, Risk ) \rightarrow \mathcal M_t. }

36. Materialization Decision Factors

可包含:

  • query frequency;
  • observer demand;
  • dependency;
  • cost;
  • predicted future use;
  • safety criticality;
  • uncertainty;
  • storage pressure;
  • latency;
  • provenance requirement;
  • authority。

37. What Deserves Active Realization Now?

GCM materialization policy 的核心是:

What representation deserves active realization now?\boxed{ \text{What representation deserves active realization now?} }

這也是 PNCW Paper 05 的中心問題。


38. Context Materialization

對 PNCW Paper 02:

MatContext:MtotalCqresident/active.\mathsf{MatContext} : \mathcal M^{total} \rightarrow C_q^{resident/active}.

39. Carrier Materialization

對 Paper 03:

MatCarrier:CqactivePkactiveregions.\mathsf{MatCarrier} : C_q^{active} \rightarrow P_k^{active-regions}.

40. Visual Materialization

對 Paper 04:

MatVisual:PkVviewport.\mathsf{MatVisual} : P_k \rightarrow V_{viewport}.

41. Materialization Stack

因此:

Mat=MatContextMatCarrierMatVisual.\boxed{ \mathsf{Mat} = \mathsf{MatContext} \circ \mathsf{MatCarrier} \circ \mathsf{MatVisual}. }

這些不是同一層。


42. Materialization / Projection Non-Collapse

Projection:

Π(X)\Pi(X)

可以只產生 logical view。

Materialization:

Mat(Π(X))\mathsf{Mat}(\Pi(X))

才把所需 detail resident。

所以:

ProjectionMaterialization.\boxed{ \text{Projection} \neq \text{Materialization}. }

43. Projection / Observation Non-Collapse

同樣:

ProjectionObservation.\boxed{ \text{Projection} \neq \text{Observation}. }

一個 projection 可以先 machine-ready,但還沒 human-visible。


44. Observation / Presentation Non-Collapse

Observer 取得 state 後,還可有不同 presentation。

因此:

ObservationPresentation.\boxed{ \text{Observation} \neq \text{Presentation}. }

45. GCM Projection Operator

對 observer (O):

ΠO:WtYO.\boxed{ \Pi_O: W_t \rightarrow Y_O. }

Projection 可以:

  • hide;
  • aggregate;
  • filter;
  • quantize;
  • resample;
  • transform modality;
  • summarize history;
  • select relation layer。

46. Read-Side by Default

一般:

ΠO(Wt) is read-side by default.\boxed{ \Pi_O(W_t) \text{ is read-side by default}. }

Projection 不應偷偷 mutate world。


47. Observer Operation

UO:OO.\mathcal U_O: O \rightarrow O'.

例如:

  • zoom;
  • focus;
  • layer toggle;
  • resolution change。

通常:

ΔW=0.\Delta W=0.

48. World Operation

UW:WtWt+1.\mathcal U_W: W_t \rightarrow W_{t+1}.

例如:

  • state update;
  • topology event;
  • database mutation;
  • agent action;
  • transaction commit。

49. Observer / World Operation Non-Collapse

UOUW.\boxed{ \mathcal U_O \neq \mathcal U_W. }

這與 MRMIC Canvas 的 view operation / world operation 分離一致。


50. Compute Globally, Observe Selectively

本文採用:

Compute Globally, Observe Selectively.\boxed{ \text{Compute Globally, Observe Selectively}. }

更精確:

global dependency、constraint、history 與 domain evolution 由 canonical Runtime 管理;每個 observer 只取得所需 projection。


51. PNCW 擴張版

本文進一步提出:

Compute Globally, Materialize Selectively, Observe Relatively.\boxed{ \text{Compute Globally, Materialize Selectively, Observe Relatively}. }

52. Global Coherence / Global Materialization Non-Collapse

Global CoherenceGlobal Simultaneous Materialization.\boxed{ \text{Global Coherence} \neq \text{Global Simultaneous Materialization}. }

53. Projection Route

定義:

ρ=SourceDomain,Representation,Resolution,Carrier,Observer,Resource,Policy.\boxed{ \rho = \left\langle SourceDomain, Representation, Resolution, Carrier, Observer, Resource, Policy \right\rangle. }

54. Route 不是單一 path

Route 可以是一個 DAG:

D0{D1,D2}D3.D_0 \rightarrow \{D_1,D_2\} \rightarrow D_3.

因此:

Projection RouteSingle Linear Pipeline.\boxed{ \text{Projection Route} \neq \text{Single Linear Pipeline}. }

55. Route Candidate

ρiRcand.\rho_i \in \mathcal R^{cand}.

每個 candidate 有:

  • feasibility;
  • latency;
  • cost;
  • quality;
  • risk;
  • migration cost;
  • materialization cost。

56. Constraint First

先求:

Rfeasible={ρi:Constraints(ρi)=PASS}.\boxed{ \mathcal R^{feasible} = \{ \rho_i: Constraints(\rho_i)=PASS \}. }

再做 optimization。


57. Feasible 不等於 Preferred

FeasiblePreferred.\boxed{ \text{Feasible} \neq \text{Preferred}. }

58. Objective Vector

對 route:

o(ρi)=(Latency,Memory,IO,Energy,Cost,Quality,Risk,Migration,VisibilityLatency).\boxed{ \mathbf o(\rho_i) = ( Latency, Memory, IO, Energy, Cost, Quality, Risk, Migration, VisibilityLatency ). }

59. Pareto Frontier

定義:

P=ParetoFront(Rfeasible).\boxed{ \mathcal P = \operatorname{ParetoFront} ( \mathcal R^{feasible} ). }

60. Pareto 不等於 Weighted Score

ParetoWeighted Score.\boxed{ \text{Pareto} \neq \text{Weighted Score}. }

61. Explicit Policy Selection

對 frontier:

P,\mathcal P,

由 policy:

PP

選:

ρ=P(P).\boxed{ \rho^\star = P(\mathcal P). }

62. Preferred 不等於 Universal Optimum

Policy PreferredUniversally Optimal.\boxed{ \text{Policy Preferred} \neq \text{Universally Optimal}. }

63. Selection / Execution / Commit Non-Collapse

即使:

ρ\rho^\star

被選中,

也:

Selected RouteExecution AuthorityCommit Authority.\boxed{ \text{Selected Route} \neq \text{Execution Authority} \neq \text{Commit Authority}. }

64. Allocation Plan

定義:

APlan=Tasks,Resources,Bindings,Capacity,Constraints,Objectives,Policy,Provenance.\boxed{ \mathsf{APlan} = \left\langle Tasks, Resources, Bindings, Capacity, Constraints, Objectives, Policy, Provenance \right\rangle. }

65. Physical Resource Binding

可能綁定:

  • CPU;
  • GPU;
  • VRAM;
  • RAM;
  • SSD;
  • NAS;
  • remote compute;
  • browser process;
  • agent session。

66. Resource Snapshot

allocation 必須基於:

Rtsnapshot.R_t^{snapshot}.

不能基於模糊的「大概有資源」。


67. Frozen Evaluation Context

多 objective 比較必須在:

Eeval\boxed{ \mathcal E^{eval} }

中固定:

  • resource snapshot;
  • objective profile;
  • units;
  • policy;
  • bounded candidate set。

68. Same Frozen Inputs → Same Selection

要求:

Same Frozen Inputs+Same ObjectiveProfile+Same Evaluation Context+Same PolicySame Result.\boxed{ \text{Same Frozen Inputs} + \text{Same ObjectiveProfile} + \text{Same Evaluation Context} + \text{Same Policy} \Rightarrow \text{Same Result}. }

69. PNCW Projection Plan

本文定義:

PPlant=ContextPlan,CarrierPlan,VisualPlan,ResourcePlan,RevealPlan.\boxed{ \mathsf{PPlan}_t = \left\langle ContextPlan, CarrierPlan, VisualPlan, ResourcePlan, RevealPlan \right\rangle. }

70. Context Plan

決定:

  • active objects;
  • page-in;
  • working set;
  • context budget;
  • query overlay。

71. Carrier Plan

決定:

  • HDSRC profile;
  • scale;
  • relation representation;
  • materialized tiles;
  • integrity mode。

72. Visual Plan

決定:

  • viewport;
  • panel set;
  • portals;
  • visual overlays;
  • render resolution。

73. Resource Plan

決定:

  • executor;
  • memory tier;
  • IO path;
  • compute budget。

74. Reveal Plan

決定:

  • stream;
  • batch;
  • atomic artifact;
  • canvas;
  • hybrid。

75. Plan Non-Collapse

Context PlanCarrier PlanVisual PlanReveal Plan.\boxed{ \text{Context Plan} \neq \text{Carrier Plan} \neq \text{Visual Plan} \neq \text{Reveal Plan}. }

76. Hierarchical Planning

可先:

GlobalPlan\mathsf{GlobalPlan}

再:

DomainPlani.\mathsf{DomainPlan}_i.

因此:

Planning can be hierarchical.\boxed{ \text{Planning} \text{ can be hierarchical}. }

77. Global Plan / Local Autonomy

Global plan 管:

  • budget;
  • hard constraints;
  • dependencies;
  • authority。

Local domain 可以在範圍內自治。


78. Constraint Envelope

定義:

Eiconstraint\boxed{ \mathcal E_i^{constraint} }

作為 domain 的合法操作 envelope。


79. Local Freedom Within Envelope

只要:

FiEiconstraint,F_i \in \mathcal E_i^{constraint},

domain 可自行:

  • refine;
  • route;
  • cache;
  • rearrange local tiles。

80. Authority Envelope

同樣:

Eiauthority\boxed{ \mathcal E_i^{authority} }

限定:

  • read;
  • propose;
  • execute;
  • commit。

81. Materialization Budget

對每 domain:

Bimat.\boxed{ B_i^{mat}. }

要求:

Costmat(Di)Bimat.Cost_{mat}(D_i) \le B_i^{mat}.

82. Active Support Budget

Ai(t)Biactive.\boxed{ |A_i(t)| \le B_i^{active}. }

83. Resolution Budget

λiBiresolution.\boxed{ \lambda_i \le B_i^{resolution}. }

84. Visibility Budget

對 observer:

VOBOview.\boxed{ |V_O| \le B_O^{view}. }

所以:

observer overload\text{observer overload}

也可以是 runtime constraint。


85. Multi-Observer Allocation

如果:

O1,,Om,O_1,\ldots,O_m,

各自要求不同 projection,

Runtime 不必 duplicate 整個 world。

可共享:

  • canonical state;
  • carrier regions;
  • cached context;
  • portal state。

86. Shared Substrate / Observer-Specific Projection

Shared Substrate+Observer-Specific Projection.\boxed{ \text{Shared Substrate} + \text{Observer-Specific Projection}. }

87. Human / AI Resource Competition

Human viewport 與 AI query 可能同時要求同一 GPU / carrier region。

GCM 可做:

Resource Arbitration.\boxed{ \text{Resource Arbitration}. }

88. Priority

對 request:

rir_i

可有:

Priority(ri).Priority(r_i).

但 priority 不得繞過 hard authority / safety constraints。


89. Safety Criticality

某些 region:

SafetyCritical=1SafetyCritical=1

可提高:

  • resolution;
  • redundancy;
  • verification;
  • residency priority。

90. Prediction of Future Use

Materialization Policy 可以包含:

U^future(X).\hat U_{future}(X).

例如 HDSRC predictive materialization / Context prefetch。


91. Prediction 不等於 Authority

Predicted UsefulAuthorized.\boxed{ \text{Predicted Useful} \neq \text{Authorized}. }

92. Predictive Materialization / GCM Routing

GCM 可以使用:

C^mat,L^,U^.\hat C_{mat}, \quad \hat L, \quad \hat U.

但 high uncertainty 時回 oracle / conservative path。


93. Fast / Oracle Selection

SelectMode{FAST,ORACLE}.\boxed{ \mathsf{SelectMode} \in \{ \mathsf{FAST}, \mathsf{ORACLE} \}. }

94. Selection Regret

Regret=C(ρselected)C(ρbestknown).\boxed{ Regret = C(\rho_{selected}) - C(\rho_{best-known}). }

95. System Regret / Prediction Error Non-Collapse

如果 predictor miss,但 fallback 防止 bad route:

Prediction ErrorSystem Regret.\boxed{ \text{Prediction Error} \neq \text{System Regret}. }

96. Replanning

當:

  • resource stale;
  • context change;
  • carrier debt;
  • task change;
  • portal loss;
  • observer change;

可以:

Replan.\boxed{ \mathsf{Replan}. }

97. Replanning / Commit Non-Collapse

新 plan:

P1P_1

不自動替代:

P0.P_0.

需要 lifecycle / authority。


98. Stale Plan

如果:

ResourceSnapshot(P0)Rt,ResourceSnapshot(P_0) \neq R_t,

則:

P0=STALE.\boxed{ P_0 = \mathsf{STALE}. }

99. Plan Lineage

P0P1P2P_0 \rightarrow P_1 \rightarrow P_2

必須記錄 supersession / cause。


100. Projection Schedule

定義:

ΣPNCW=(ObserveWorld,BuildPlan,Allocate,ProjectContext,FreezeCarrier,MaterializeCarrier,BuildVisualWorld,Verify,Reveal).\boxed{ \Sigma_{PNCW} = ( \mathsf{ObserveWorld}, \mathsf{BuildPlan}, \mathsf{Allocate}, \mathsf{ProjectContext}, \mathsf{FreezeCarrier}, \mathsf{MaterializeCarrier}, \mathsf{BuildVisualWorld}, \mathsf{Verify}, \mathsf{Reveal} ). }

101. Schedule Non-Commutativity

一般:

ABBA.A\circ B \neq B\circ A.

例如先 Reveal 再 Verify 是非法的。


102. No Premature Reveal

RevealVerifyis forbidden for authoritative mode.\boxed{ \mathsf{Reveal} \prec \mathsf{Verify} \quad \text{is forbidden for authoritative mode}. }

103. No Premature Full Materialization

同樣:

Materialize Everything First\boxed{ \text{Materialize Everything First} }

不應成為 default。


104. Lazy Materialization

優先:

PlanMaterialize Required SupportExpand on Demand.\boxed{ \text{Plan} \rightarrow \text{Materialize Required Support} \rightarrow \text{Expand on Demand}. }

105. Speculative Materialization

Runtime 可以 speculative prefetch:

Xspec.X^{spec}.

但:

Speculative ResidencySemantic Activation.\boxed{ \text{Speculative Residency} \neq \text{Semantic Activation}. }

106. Eviction

如果:

Costkeep(X)>ExpectedBenefit(X),Cost_{keep}(X) > ExpectedBenefit(X),

可以:

Evict(X).\boxed{ \mathsf{Evict}(X). }

107. Evict / Forget Non-Collapse

EvictForget.\boxed{ \text{Evict} \neq \text{Forget}. }

evict 只改 residency。


108. Archive

如果長期不 active:

Archive(X)\boxed{ \mathsf{Archive}(X) }

保留重建所需資訊。


109. Archive / Delete Non-Collapse

ArchiveDelete.\boxed{ \text{Archive} \neq \text{Delete}. }

110. PNCW Materialization Ledger

記錄:

eM=Object,OldState,NewState,Reason,Cost,Policy,Time.\boxed{ e_M = \left\langle Object, OldState, NewState, Reason, Cost, Policy, Time \right\rangle. }

111. Resolution Ledger

eλ=Domain,OldResolution,NewResolution,Cause,Risk,Time.\boxed{ e_\lambda = \left\langle Domain, OldResolution, NewResolution, Cause, Risk, Time \right\rangle. }

112. Projection Routing Ledger

eρ=Request,Candidates,FeasibleSet,Frontier,Policy,SelectedRoute,Evidence.\boxed{ e_\rho = \left\langle Request, Candidates, FeasibleSet, Frontier, Policy, SelectedRoute, Evidence \right\rangle. }

113. History / Current State Non-Collapse

即使:

CurrentPlan=P2,CurrentPlan=P_2,

仍保留:

P0,P1.P_0,P_1.

所以:

Current RouteRouting History.\boxed{ \text{Current Route} \neq \text{Routing History}. }

114. GCM × Context VM

GCM 決定:

  • budget;
  • resource;
  • representation;
  • global dependency。

Context MMU 決定:

  • page;
  • working set;
  • active context。

因此:

GCM+Context MMU=Cognitive Materialization Governance.\boxed{ \text{GCM} + \text{Context MMU} = \text{Cognitive Materialization Governance}. }

115. GCM × SPET/HDSRC

GCM 決定:

  • carrier candidate;
  • scale;
  • materialization budget;
  • resource route。

SPET/HDSRC 決定:

  • frame validity;
  • carrier semantics;
  • native operator correctness。

116. GCM × MRMIC/NVCL

GCM 決定:

  • viewport budget;
  • portal priority;
  • render resources;
  • active panels。

MRMIC/NVCL 提供:

  • visual world;
  • interaction;
  • resource portals;
  • observer projection。

117. PNCW Global Runtime Architecture

WtGCMPlanPPlantContextMMUCqactiveSPET/HDSRCPkMRMIC/NVCLVq,kRevealPolicyUq,k.\boxed{ \begin{aligned} W_t &\xrightarrow{\mathsf{GCMPlan}} \mathsf{PPlan}_t \\ &\xrightarrow{\mathsf{ContextMMU}} C_q^{active} \\ &\xrightarrow{\mathsf{SPET/HDSRC}} P_k \\ &\xrightarrow{\mathsf{MRMIC/NVCL}} V_{q,k} \\ &\xrightarrow{\mathsf{RevealPolicy}} U_{q,k}. \end{aligned} }

118. Global Runtime Invariant

No layer may silently assume authority owned by another layer.\boxed{ \text{No layer may silently assume authority owned by another layer}. }

119. Planner / Executor Non-Collapse

PlannerExecutor.\boxed{ \text{Planner} \neq \text{Executor}. }

120. Executor / Committer Non-Collapse

ExecutorCommitter.\boxed{ \text{Executor} \neq \text{Committer}. }

121. Observer / Planner Non-Collapse

Observer demand 可以影響 materialization,但:

Observer RequestGlobal Plan Authority.\boxed{ \text{Observer Request} \neq \text{Global Plan Authority}. }

122. AI Proposal / Deterministic Runtime Non-Collapse

未來 AI 可提議:

Policy,Objectives,Representation,Resolution.Policy, Objectives, Representation, Resolution.

但:

AI ProposalDeterministic Runtime Decision.\boxed{ \text{AI Proposal} \neq \text{Deterministic Runtime Decision}. }

123. Phase C / Future AI Orchestration

只有 deterministic substrate 成熟後,才值得研究:

AI-assisted Compute Orchestrator.\boxed{ \text{AI-assisted Compute Orchestrator}. }

它應提出 plan,不直接繞過 hard constraints。


124. PNCW Global Objective

定義:

J=wLL+wMM+wII+wEE+wRR+wQQ+wVV.\boxed{ J = w_L L + w_M M + w_I I + w_E E + w_R R + w_Q Q + w_V V. }

但 scalar objective 只是一種 policy,不是 universal truth。


125. Multi-Objective Default

因此更穩妥的是:

FeasibilityParetoExplicit Policy.\boxed{ \text{Feasibility} \rightarrow \text{Pareto} \rightarrow \text{Explicit Policy}. }

126. Proposition 1 — Global Coherence Does Not Require Global Materialization

只要 canonical dependency / constraints 可被 Runtime 管理,observer/task 所需 support 有效,則:

Global Coherence⇏Full Materialization.\boxed{ \text{Global Coherence} \not\Rightarrow \text{Full Materialization}. }

127. Proposition 2 — Different Domains May Use Different Resolutions

存在:

Di,DjD_i,D_j

使:

λiλj\lambda_i\neq\lambda_j

而 global constraints 仍成立。


128. Proposition 3 — Observer Resolution Does Not Determine Compute Resolution

存在:

λobserveλcompute.\lambda^{observe} \neq \lambda^{compute}.

因此 UI zoom 不應直接綁定 solver fidelity。


129. Proposition 4 — Projection Route Selection Can Be Deterministic Without Being Unique in Objective Space

同一 feasible frontier 可有多個 nondominated routes,但 explicit policy 可 deterministically 選一個。

因此:

Multiple Valid AlternativesNondeterministic System.\boxed{ \text{Multiple Valid Alternatives} \neq \text{Nondeterministic System}. }

130. Proposition 5 — Materialization State Can Change Without Semantic State Change

如果:

X:ResidentArchivedX: Resident \rightarrow Archived

但 canonical content 不變,

則:

ΔResidency⇏ΔSemantics.\boxed{ \Delta Residency \not\Rightarrow \Delta Semantics. }

131. Proposition 6 — Visibility Can Change Without Recomputing the World

不同 observer projection / viewport 可:

V1V2V_1 \rightarrow V_2

同時:

WtW_t

不變。


132. PNCW Paper 05 規範 v0.1

PNCW-G1 — Global / Everywhere Non-Collapse

Global 不得被解釋為 all-to-all/full materialization。

PNCW-G2 — Dependency / Materialization Separation

存在 dependency 不等於同時展開全部 detail。

PNCW-G3 — Finite Active Realization

任一真實 Runtime 的 active support 必須 finite / bounded。

PNCW-G4 — Unbounded / Infinite Active Separation

無界可展開不等於無限同時活動。

PNCW-G5 — Resolution Typing

compute / observe / carrier / render resolution 必須分開。

PNCW-G6 — Projection / Materialization Separation

logical projection 不等於 physical residency。

PNCW-G7 — Constraint Before Optimization

所有 route selection 必須先通過 hard constraints。

PNCW-G8 — Feasible / Preferred / Optimal Separation

不得把 policy-selected route 宣稱 universal optimum。

PNCW-G9 — Plan / Execution / Commit Separation

選中 plan 不等於有 execution / commit authority。

PNCW-G10 — Explicit Materialization Policy

active realization 必須有可解釋 policy / evidence。

PNCW-G11 — Replanning Lineage

stale / superseded plan 必須保存 lineage。

PNCW-G12 — Observer Relativity

不同 observer 可以合法取得不同 projection / resolution。


133. 對「一口氣看到」的真正全域解釋

到了 Paper 05,「一口氣看到」可以被更精確地描述成:

Global world remains coherent,\boxed{ \text{Global world remains coherent,} } only task-relevant structures are materialized,\boxed{ \text{only task-relevant structures are materialized,} } and the observer receives a logically complete projection of that active world.\boxed{ \text{and the observer receives a logically complete projection of that active world}. }

134. 它不是全世界一次展開

不是:

WtRender Everything.\boxed{ W_t \rightarrow \text{Render Everything}. }

而是:

WtPlanAtMtVO.\boxed{ W_t \rightarrow \mathsf{Plan} \rightarrow A_t \rightarrow M_t \rightarrow V_O. }

135. Active / Materialized / Visible Chain

WtAtMttaskVOvisible\boxed{ W_t \supseteq A_t \supseteq M_t^{task} \supseteq V_O^{visible} }

注意這只是常見 inclusion,不是所有架構必須嚴格如此。


136. Global Structure First

Observer 可以先拿到:

  • manifest;
  • domain graph;
  • active regions;
  • panel identities;
  • unresolved areas;
  • projection routes。

再按需 materialize detail。


137. First Useful Global View

因此新的 latency 指標:

Lglobalview.\boxed{ L_{global-view}. }

它不同於:

Lfirsttoken.L_{first-token}.

138. First Actionable State

另一個指標:

Lfirstactionable.\boxed{ L_{first-actionable}. }

表示 observer 何時能對整體 artifact 做有意義操作。


139. Sequence vs Projection Benchmark

Sequence mode 優化:

Lfirsttoken.L_{first-token}.

Projection mode可能優化:

Lglobalview,Lfirstactionable.L_{global-view}, L_{first-actionable}.

兩者不是同一產品目標。


140. PNCW Vertical Slice — Global Phase

完整最小實驗:

  1. 建立 10410510^4\sim10^5 node canonical world;
  2. 建多個 task/query;
  3. GCM 建 candidate routes;
  4. constraints filter;
  5. Pareto frontier;
  6. explicit policy select;
  7. Context MMU materialize active support;
  8. SPET Freeze;
  9. HDSRC materialize selected carrier regions;
  10. MRMIC 建 visual world;
  11. Atomic Reveal;
  12. observer zoom / query;
  13. dynamic refinement;
  14. resource pressure 觸發 eviction;
  15. task change 觸發 replanning。

141. Metrics

MG=ActiveFraction,MaterializedFraction,GlobalLatency,FirstGlobalView,FirstActionable,PeakMemory,IO,ReplanCount,MigrationCost,SelectionRegret,StaleRejectRate,ReuseRate.\boxed{ \mathbf M_G = \left\langle ActiveFraction, MaterializedFraction, GlobalLatency, FirstGlobalView, FirstActionable, PeakMemory, IO, ReplanCount, MigrationCost, SelectionRegret, StaleRejectRate, ReuseRate \right\rangle. }

142. Resolution Metrics

Mλ=ComputeRes,CarrierRes,ObserveRes,RenderRes,RefinementCount,DeRefinementCount.\boxed{ \mathbf M_\lambda = \left\langle ComputeRes, CarrierRes, ObserveRes, RenderRes, RefinementCount, DeRefinementCount \right\rangle. }

143. Materialization Metrics

MM=PageIn,PageOut,CacheHit,PrefetchHit,Eviction,Archive,Reconstruct,ResidentBytes.\boxed{ \mathbf M_M = \left\langle PageIn, PageOut, CacheHit, PrefetchHit, Eviction, Archive, Reconstruct, ResidentBytes \right\rangle. }

144. Failure Conditions

若:

  • active support bounded 但 runtime 還每 tick scan 全 world;
  • lazy materialization 導致 page-fault storm;
  • replan churn;
  • resolution oscillation;
  • observer view inconsistent;
  • route predictor distribution shift 無 fallback;
  • migration cost 超過收益;
  • global constraint 被 local optimization 破壞;

則 PNCW global-runtime claim 必須降低。


145. Bounded Semantics / Bounded Runtime Cost Non-Collapse

即使:

AtBA,|A_t|\le B_A,

如果每 tick 還:

  • scan dormant;
  • clone full world;
  • render all;
  • rehash all;

仍不能說 cost bounded。

所以:

Bounded Active SemanticsBounded Runtime Cost.\boxed{ \text{Bounded Active Semantics} \neq \text{Bounded Runtime Cost}. }

146. Runtime Closure Requirement

真正 bounded runtime 還要:

Bounded Active+Bounded Scan+Bounded Materialization+Bounded Commit.\boxed{ \text{Bounded Active} + \text{Bounded Scan} + \text{Bounded Materialization} + \text{Bounded Commit}. }

147. 與 Paper 06 的接口

Paper 05 回答:

誰決定什麼被算、被物化、被投影、被顯示?

下一篇 Paper 06 將整個系列收束成:

一個真正可實作的 Non-Sequential AI Output Architecture 應如何把 GCM、Context VM、SPET/HDSRC、MRMIC/NVCL、Visibility Commit 變成一條完整 runtime pipeline?

即:

WorldPlanActive ContextCarrierCanvasAtomic Visible Result.\boxed{ \text{World} \rightarrow \text{Plan} \rightarrow \text{Active Context} \rightarrow \text{Carrier} \rightarrow \text{Canvas} \rightarrow \text{Atomic Visible Result}. }

148. 系列位置

P00:Projection-Native World FoundationsP01:Visibility / Atomic RevealP02:Virtual Context ProjectionP03:Stable High-D Projection CarrierP04:Visual Computational CanvasP05:Global Compute / Local MaterializationP06:Non-Sequential AI Output Architecture\boxed{ \begin{aligned} P00 &: \text{Projection-Native World Foundations}\\ P01 &: \text{Visibility / Atomic Reveal}\\ P02 &: \text{Virtual Context Projection}\\ P03 &: \text{Stable High-D Projection Carrier}\\ P04 &: \text{Visual Computational Canvas}\\ P05 &: \text{Global Compute / Local Materialization}\\ P06 &: \text{Non-Sequential AI Output Architecture} \end{aligned} }

149. 結論

本文把 GCM 正式放回 PNCW 的最上層 runtime governance。

它的核心不是:

把所有東西都算完再顯示。

而是:

Global Coherence+Finite Active Realization+Selective Materialization+Observer-Relative Projection.\boxed{ \text{Global Coherence} + \text{Finite Active Realization} + \text{Selective Materialization} + \text{Observer-Relative Projection}. }

因此:

Global ComputationGlobal Simultaneous Materialization.\boxed{ \text{Global Computation} \neq \text{Global Simultaneous Materialization}. }

一個世界可以保持全域一致,但:

  • 只有少數 domain active;
  • 只有部分 active domain materialized;
  • 只有部分 materialized state 進 carrier;
  • 只有部分 carrier region 進 viewport;
  • 不同 observer 取得不同 projection。

本文也正式建立:

λcomputeλobserveλcarrierλrender.\boxed{ \lambda^{compute} \neq \lambda^{observe} \neq \lambda^{carrier} \neq \lambda^{render}. }

因此「看得更細」不自動要求「算得更細」,「算得更細」也不自動要求「全部顯示」。

透過 GCM Phase B 類型的 deterministic allocation,PNCW 可以進一步將:

ConstraintFeasible RoutesPareto FrontierExplicit Policy Selection\boxed{ \text{Constraint} \rightarrow \text{Feasible Routes} \rightarrow \text{Pareto Frontier} \rightarrow \text{Explicit Policy Selection} }

用於 projection/materialization planning。

這使 PNCW 不只是一套 visual metaphor,而開始具有一個真正可實作的 global runtime selection semantics。

本文最終提出:

Compute Globally, Materialize Selectively, Observe Relatively.\boxed{ \text{Compute Globally, Materialize Selectively, Observe Relatively}. }

這句話正是「一口氣看到」在 PNCW 中最準確的全域版本。

不是讓整個宇宙同時 render。

而是:

全域世界保持一致;Runtime 只把現在真正有用的部分算到適合的解析度、物化成適合的 representation,再讓 observer 一次取得完整而可操作的當前世界投影。


內部理論與工程血統

本文主要承接:

  1. PNCW Paper 00–04;
  2. Global Computation Methodology Series 00–06;
  3. GCM Paper 04 — Computation ≠ Observation / Selective Materialization;
  4. GCM Paper 05 — Finite Active Realization and Unbounded Extensibility;
  5. GCM Phase A — Reference Runtime MVP;
  6. GCM Phase B — Deterministic Compute Allocator;
  7. B1 — Intent & Deterministic Planning Contract;
  8. B2 — Resource Snapshot & Binding;
  9. B3 — Capacity-Constrained Allocation;
  10. B4 — Multi-Objective / Pareto Allocation;
  11. AI Context Virtual Memory / Context MMU;
  12. SPET / HDSRC carrier routing;
  13. MRMIC/NVCL visual computational surface。

本文不宣稱 GCM Phase B 全部完成;目前以已驗證的 Phase A 與已完成之 B1–B4 engineering evidence 作為可引用工程基礎,B5 lifecycle/replanning、B6 final conformance、Phase C AI-assisted orchestration 與 Phase D native compute fabric 仍屬後續階段。