類全域 AI 世界—計算—觀察統合系列(Paper 03)
計算層:世界族上的全域異質計算
The Computation Layer: Heterogeneous Global Computation over Governed World Families
作者: Neo.KAI 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 03 / 12版本: v0.1日期: 2026-09-08研究定位: Global Computation Methodology × Governed World Family × Computational Configuration Space × Dynamic Computational Routing × CDI/AIVS × PNCW × Heterogeneous Runtime × Candidate/Commit前篇: Paper 02《世界層:從單一 World Model 到 Governed World Family》狀態: 計算層母規格/Runtime composition framework;不宣稱本文提供普遍最優 scheduler,也不宣稱 24/72 分類已被證明為計算形態的完備基底
摘要
Paper 02 已將類全域 AI 的世界層由單一 world model 擴張為受治理世界族:
W t G . \mathfrak W_t^G. W t G .
然而,擁有很多可運行 worlds 並不等於知道如何計算它們。一個類全域 AI 可能同時面對 actual-linked world、digital twin、simulation、counterfactual、replay、synthetic 與 nested worlds;每個 world 又可能包含連續物理、離散制度、圖結構、符號約束、機率過程、資料流、神經模型與外部程式。若所有 world、所有 domain 都被強迫使用同一算法、同一 timestep、同一 representation、同一 precision、同一 accelerator 與同一同步方式,所謂「全域計算」反而退化成全域同質化。
本文提出 WCO-TF 的正式 Computation Layer ,核心命題沿用並擴張既有 Global Computation Methodology:
Global Computation = Globally Coherent Heterogeneous Computation . \boxed{
\text{Global Computation}
=
\text{Globally Coherent Heterogeneous Computation}.
} Global Computation = Globally Coherent Heterogeneous Computation .
本文將 Paper 01 的計算族:
C t \mathfrak C_t C t
細化為:
C t = ( F t c o m p , L t , R t e x e c , Q t , S t , V t ) , \boxed{
\mathfrak C_t
=
\left(
\mathfrak F_t^{comp},
\mathfrak L_t,
\mathfrak R_t^{exec},
\mathfrak Q_t,
\mathfrak S_t,
\mathfrak V_t
\right),
} C t = ( F t co m p , L t , R t e x ec , Q t , S t , V t ) ,
其中:
F t c o m p \mathfrak F_t^{comp} F t co m p :computational-form family;
L t \mathfrak L_t L t :transition-law family;
R t e x e c \mathfrak R_t^{exec} R t e x ec :resource/backend family;
Q t \mathfrak Q_t Q t :precision / fidelity / resolution family;
S t \mathfrak S_t S t :scheduling / synchronization modes;
V t \mathfrak V_t V t :available validation / reconciliation procedures。
本文特別將既有 GCM 中常以 P \mathfrak P P 表示的 computational form space 改記為:
F c o m p \mathfrak F^{comp} F co m p
以避免與 WCO 系列中 projection family:
P \mathfrak P P
發生符號坍縮。這個重新命名本身反映本系列的核心非同一性:
Computational Form ≠ Projection Form . \boxed{
\text{Computational Form}
\neq
\text{Projection Form}.
} Computational Form = Projection Form .
對受治理世界 W i W_i W i 的 computational domain D i j D_{ij} D ij ,本文定義計算配置:
κ i j , t = ⟨ W i , D i j , f i j , ℓ i j , λ i j c o m p , r i j , τ i j , q i j , σ i j ⟩ , \boxed{
\kappa_{ij,t}
=
\left\langle
W_i,
D_{ij},
f_{ij},
\ell_{ij},
\lambda_{ij}^{comp},
r_{ij},
\tau_{ij},
q_{ij},
\sigma_{ij}
\right\rangle,
} κ ij , t = ⟨ W i , D ij , f ij , ℓ ij , λ ij co m p , r ij , τ ij , q ij , σ ij ⟩ ,
其中:
f i j ∈ F c o m p f_{ij}\in\mathfrak F^{comp} f ij ∈ F co m p :computational form;
ℓ i j ∈ L \ell_{ij}\in\mathfrak L ℓ ij ∈ L :transition law;
λ i j c o m p \lambda_{ij}^{comp} λ ij co m p :compute resolution / materialization level;
r i j r_{ij} r ij :resource/backend target;
τ i j \tau_{ij} τ ij :execution order / temporal mode;
q i j q_{ij} q ij :precision / quality target;
σ i j \sigma_{ij} σ ij :synchronization / consistency mode。
計算路由器因此不只回答「什麼時候、在哪裡跑」,而回答:
what + how + at what resolution + on which backend + under which consistency contract . \boxed{
\text{what}
+
\text{how}
+
\text{at what resolution}
+
\text{on which backend}
+
\text{under which consistency contract}.
} what + how + at what resolution + on which backend + under which consistency contract .
本文定義 World-Family Computational Router:
R t W F C : ( W t G , D i j , T , B t , R i s k t , Γ t ) ↦ κ i j , t . \boxed{
\mathcal R_t^{WFC}
:
(
\mathfrak W_t^G,
D_{ij},
T,
B_t,
Risk_t,
\Gamma_t
)
\mapsto
\kappa_{ij,t}.
} R t W F C : ( W t G , D ij , T , B t , R i s k t , Γ t ) ↦ κ ij , t .
其中 T T T 是 task / intent, B t B_t B t 是資源預算, R i s k t Risk_t R i s k t 是風險狀態, Γ t \Gamma_t Γ t 是 authority / admissibility constraints。
然而,真正的全域性不在 routing 本身,而在 Global Composition and Reconciliation 。各 domain 的 local executor 可以異質地計算:
Δ i j c a n d = Φ i j f i j , ℓ i j ( S i j , I i j ) , \Delta_{ij}^{cand}
=
\Phi_{ij}^{f_{ij},\ell_{ij}}
\left(
S_{ij},
I_{ij}
\right), Δ ij c an d = Φ ij f ij , ℓ ij ( S ij , I ij ) ,
但 local success 不得直接改寫 canonical world。所有候選 delta 必須進入:
R e c o n c i l e G ( { Δ i j c a n d } , C t G , W t G , H t ) \boxed{
\mathsf{Reconcile}_G
\left(
\{\Delta_{ij}^{cand}\},
\mathcal C_t^G,
\mathfrak W_t^G,
H_t
\right)
} Reconcile G ( { Δ ij c an d } , C t G , W t G , H t )
檢查:
cross-domain dependencies;
cross-world coupling;
lineage constraints;
evidence mode;
state versions;
authority;
resource receipts;
semantic conflicts;
global invariants;
accumulated projection / computation debt。
只有通過 reconciliation 的候選狀態才可:
C a n d i d a t e → C o m m i t . Candidate
\rightarrow
Commit. C an d i d a t e → C o mmi t .
因此:
Local Execution Success ⇏ Global Commit Admissibility . \boxed{
\text{Local Execution Success}
\not\Rightarrow
\text{Global Commit Admissibility}.
} Local Execution Success ⇒ Global Commit Admissibility .
本文進一步建立 World-Family Global Computation Contract(WF-GCC) :
W F - G C C = ⟨ B o u n d a r y , D o m a i n s , C o n f i g u r a t i o n s , D e p e n d e n c i e s , B u d g e t s , C a n d i d a t e D e l t a s , R e c o n c i l i a t i o n , C o m m i t , H i s t o r y , C e r t i f i c a t e s ⟩ . \boxed{
\mathsf{WF\text{-}GCC}
=
\left\langle
Boundary,
Domains,
Configurations,
Dependencies,
Budgets,
CandidateDeltas,
Reconciliation,
Commit,
History,
Certificates
\right\rangle.
} WF - GCC = ⟨ B o u n d a r y , D o main s , C o n f i g u r a t i o n s , D e p e n d e n c i es , B u d g e t s , C an d i d a t eD e l t a s , R eco n c i l ia t i o n , C o mmi t , H i s t or y , C er t i f i c a t es ⟩ .
其中 Boundary 特別重要。全域不是對所有已存在 worlds 無限擴張,而是相對指定任務與 world boundary 建立 coherence:
Global B = coherence over a designated computation boundary B . \boxed{
\text{Global}_B
=
\text{coherence over a designated computation boundary }B.
} Global B = coherence over a designated computation boundary B .
本文也把 GCM 的 Finite Active Realization + Unbounded Extensibility 推進到 world-family level。令所有可潛在計算的 world-domain pair 為:
U t c o m p , \mathcal U_t^{comp}, U t co m p ,
當下 active support 為:
A t c o m p ⊆ U t c o m p . \mathcal A_t^{comp}
\subseteq
\mathcal U_t^{comp}. A t co m p ⊆ U t co m p .
任何實際 runtime 必須有:
∣ A t c o m p ∣ < ∞ , \boxed{
|\mathcal A_t^{comp}|<\infty,
} ∣ A t co m p ∣ < ∞ ,
但允許:
U t + 1 c o m p ⊃ U t c o m p . \mathcal U_{t+1}^{comp}
\supset
\mathcal U_t^{comp}. U t + 1 co m p ⊃ U t co m p .
所以:
Finite Active Computation + Open-Ended Computational Extensibility . \boxed{
\text{Finite Active Computation}
+
\text{Open-Ended Computational Extensibility}.
} Finite Active Computation + Open-Ended Computational Extensibility .
本文同時保留 PNCW 的四解析度分離:
λ c o m p u t e ≠ λ o b s e r v e ≠ λ c a r r i e r ≠ λ r e n d e r . \boxed{
\lambda^{compute}
\neq
\lambda^{observe}
\neq
\lambda^{carrier}
\neq
\lambda^{render}.
} λ co m p u t e = λ o b ser v e = λ c a r r i er = λ r e n d er .
一個 world 可以高精度計算但只向 observer 顯示粗粒度摘要;也可以低成本 coarse simulation 先篩選 branches,再只對高價值 branch 提升 computational fidelity。這使 multi-fidelity world family 成為全域計算的自然結果,而不是例外。
最後,本文把計算歷史提升為一級狀態。即使兩條計算路徑最後得到相同 state:
S a = S b , S_a=S_b, S a = S b ,
只要它們的 backend、seed、approximation、proof obligation、data version、resource receipt 或 intermediate failure 不同,就不應直接認為 computation history 相同:
State Equality ⇏ Computation-History Equality . \boxed{
\text{State Equality}
\not\Rightarrow
\text{Computation-History Equality}.
} State Equality ⇒ Computation-History Equality .
這與 MWT 的非交換歷史保存及動態不動點系列直接接合。
本文最終提出:
Global-Like AI computation is not one giant calculation. \boxed{
\text{Global-Like AI computation is not one giant calculation.}
} Global-Like AI computation is not one giant calculation.
而是:
a governed routing-and-reconciliation process over heterogeneous computations distributed across a family of worlds . \boxed{
\text{a governed routing-and-reconciliation process
over heterogeneous computations distributed across a family of worlds}.
} a governed routing-and-reconciliation process over heterogeneous computations distributed across a family of worlds .
關鍵詞: Global Computation、Heterogeneous Computing、World Family、Computational Routing、24/72、Transition Law、Finite Active Support、Selective Materialization、Candidate/Commit、Global Reconciliation、CDI、AIVS、PNCW
0. Paper 02 留下的計算問題
Paper 02 已建立:
W t G . \mathfrak W_t^G. W t G .
但世界族只是可被計算的對象集合與治理結構。
它本身沒有回答:
哪些 world 現在值得算?
world 裡哪個 domain 應該算?
要用 deterministic、probabilistic、symbolic、graph、simulation 還是 learned model?
哪個 backend?
哪個 precision?
哪些可以 parallel?
哪些必須 serial?
哪些只是 candidate?
哪些可以 commit?
這就是 Paper 03。
1. 全域計算不是一個超大 Algorithm
本文拒絕:
Global Computation = One Universal Algorithm Everywhere . \boxed{
\text{Global Computation}
=
\text{One Universal Algorithm Everywhere}.
} Global Computation = One Universal Algorithm Everywhere .
2. 全域計算是世界相對的 Coherence
Global B = coherence relative to designated boundary B . \boxed{
\text{Global}_B
=
\text{coherence relative to designated boundary }B.
} Global B = coherence relative to designated boundary B .
B B B 可以是:
一個 world;
一組 worlds;
一組 domains;
一個 organization;
一個 simulation portfolio;
一個 research problem。
3. Boundary 不必等於全部 World Family
若:
W t G = { W 1 , … , W 1000 } , \mathfrak W_t^G
=
\{W_1,\ldots,W_{1000}\}, W t G = { W 1 , … , W 1000 } ,
一個 task 可能只需要:
B T = { W 3 , W 8 , W 21 } . B_T
=
\{W_3,W_8,W_{21}\}. B T = { W 3 , W 8 , W 21 } .
4. Global Dependency 不等 Full Activation
Global Dependency ≠ Full Active Computation . \boxed{
\text{Global Dependency}
\neq
\text{Full Active Computation}.
} Global Dependency = Full Active Computation .
5. World Family 與 Computation Family
本文建立:
( W t G , C t ) . \boxed{
(\mathfrak W_t^G,\mathfrak C_t).
} ( W t G , C t ) .
這是 Paper 03 的核心耦合物件。
6. Computation Family 正式拆解
C t = ( F t c o m p , L t , R t e x e c , Q t , S t , V t ) . \boxed{
\mathfrak C_t
=
\left(
\mathfrak F_t^{comp},
\mathfrak L_t,
\mathfrak R_t^{exec},
\mathfrak Q_t,
\mathfrak S_t,
\mathfrak V_t
\right).
} C t = ( F t co m p , L t , R t e x ec , Q t , S t , V t ) .
7. Computational Form 與 Projection Form 分離
既有研究有時使用:
P \mathfrak P P
表示 computational paradigms。
但 WCO 已把:
P \mathfrak P P
保留給 projection family。
因此本文改用:
F c o m p . \boxed{
\mathfrak F^{comp}.
} F co m p .
8. 核心非同一性
Computational Form ≠ Projection Form . \boxed{
\text{Computational Form}
\neq
\text{Projection Form}.
} Computational Form = Projection Form .
一個 graph algorithm 可以輸出 text projection。
一個 neural computation 也可以輸出 symbolic projection。
9. Computational Forms
F c o m p \mathfrak F^{comp} F co m p 可以包含:
sequence;
parallel;
concurrent;
dataflow;
graph traversal;
symbolic deduction;
constraint solving;
optimization;
stochastic sampling;
simulation;
retrieval;
theorem proving;
learned inference;
mixed / hybrid forms。
10. 24/72 的位置
本文不把 24/72 當成已證明完備分類。
而把它定位為:
addressable computational configuration vocabulary . \boxed{
\text{addressable computational configuration vocabulary}.
} addressable computational configuration vocabulary .
11. Taxonomy 不等 Runtime
Computational Taxonomy ≠ Computational Routing . \boxed{
\text{Computational Taxonomy}
\neq
\text{Computational Routing}.
} Computational Taxonomy = Computational Routing .
12. Configuration Space
K c o m p = F c o m p × L × Q × R e x e c × S . \mathcal K^{comp}
=
\mathfrak F^{comp}
\times
\mathfrak L
\times
\mathfrak Q
\times
\mathfrak R^{exec}
\times
\mathfrak S. K co m p = F co m p × L × Q × R e x ec × S .
13. Transition-Law Family
L = { ℓ 1 , ℓ 2 , … } . \mathfrak L
=
\{\ell_1,\ell_2,\ldots\}. L = { ℓ 1 , ℓ 2 , … } .
可包含:
discrete deterministic;
continuous dynamics;
stochastic transition;
event-driven;
hybrid;
learned transition;
rule-based transition。
14. 同一 World 可以多 Transition Laws
例如:
W i = physics + market rules + agent behavior . W_i
=
\text{physics}
+
\text{market rules}
+
\text{agent behavior}. W i = physics + market rules + agent behavior .
不同 domain 不必共享一個 ℓ \ell ℓ 。
15. Computational Domain
對 world W i W_i W i :
D i c o m p = { D i j } j . \mathcal D_i^{comp}
=
\{D_{ij}\}_{j}. D i co m p = { D ij } j .
16. Domain 不是硬體分區
D i j D_{ij} D ij 可以是:
spatial region;
subsystem;
semantic domain;
causal component;
algorithmic subproblem;
agent population;
proof obligation。
17. Domain 可以跨 World
某比較任務可能建立:
D c r o s s ⊂ W i × W j . D^{cross}
\subset
W_i\times W_j. D cr oss ⊂ W i × W j .
18. Cross-World Computation
例如:
C o m p a r e ( W A , W B ) Compare(W_A,W_B) C o m p a r e ( W A , W B )
本身就是跨 world computational domain。
19. 計算配置
本文定義:
κ i j , t = ⟨ W i , D i j , f i j , ℓ i j , λ i j c o m p , r i j , τ i j , q i j , σ i j ⟩ . \boxed{
\kappa_{ij,t}
=
\left\langle
W_i,
D_{ij},
f_{ij},
\ell_{ij},
\lambda_{ij}^{comp},
r_{ij},
\tau_{ij},
q_{ij},
\sigma_{ij}
\right\rangle.
} κ ij , t = ⟨ W i , D ij , f ij , ℓ ij , λ ij co m p , r ij , τ ij , q ij , σ ij ⟩ .
20. f i j f_{ij} f ij
computational form。
21. ℓ i j \ell_{ij} ℓ ij
transition law。
22. λ i j c o m p \lambda_{ij}^{comp} λ ij co m p
compute resolution / active materialization。
23. r i j r_{ij} r ij
backend / resource target。
例如:
CPU;
GPU;
NPU;
FPGA;
theorem prover;
database;
remote cluster;
external simulator。
24. τ i j \tau_{ij} τ ij
execution order / temporal mode。
例如:
immediate;
deferred;
batch;
streaming;
speculative;
background;
deadline-bound。
25. q i j q_{ij} q ij
precision / quality / confidence target。
26. σ i j \sigma_{ij} σ ij
synchronization / consistency mode。
例如:
strict barrier;
local consistency;
eventual reconciliation;
snapshot isolation;
asynchronous relay。
27. Scheduling 是 Routing 的子集
Scheduling ⊂ Computational Routing . \boxed{
\text{Scheduling}
\subset
\text{Computational Routing}.
} Scheduling ⊂ Computational Routing .
Scheduling 主要處理:
Routing 還要處理:
how;
which form;
which law;
which precision;
which materialization;
which consistency mode。
28. World-Family Computational Router
R t W F C : ( W t G , D , T , B t , R i s k t , Γ t ) ↦ κ t . \boxed{
\mathcal R_t^{WFC}
:
(
\mathfrak W_t^G,
D,
T,
B_t,
Risk_t,
\Gamma_t
)
\mapsto
\kappa_t.
} R t W F C : ( W t G , D , T , B t , R i s k t , Γ t ) ↦ κ t .
29. Router 不等 Executor
R o u t e r ≠ E x e c u t o r . \boxed{
\mathsf{Router}
\neq
\mathsf{Executor}.
} Router = Executor .
30. Router 不等 Governor
R o u t e r ≠ G o v e r n o r . \boxed{
\mathsf{Router}
\neq
\mathsf{Governor}.
} Router = Governor .
Governor 決定 admissibility。
Router 在 admissible space 中選方案。
31. Router 不等 Observer
R o u t e C o m p u t e ≠ S e l e c t O b s e r v a t i o n . \boxed{
\mathsf{RouteCompute}
\neq
\mathsf{SelectObservation}.
} RouteCompute = SelectObservation .
Paper 04 才正式處理 Observation Family。
32. Route 的輸入包含 World Mode
同一 computation 對:
S I M SIM S I M
與:
A L AL A L
可能採不同安全與 precision policy。
33. Pure Simulation 可以更激進
例如:
speculative compute;
destructive branch;
rollback;
coarse approximation;
massive parallelism。
34. Reality-Coupled World 更嚴格
需要:
version checks;
freshness;
authority;
effect barriers;
stronger validation。
35. Local Executor
對配置:
κ i j , t , \kappa_{ij,t}, κ ij , t ,
執行:
Δ i j c a n d = Φ i j f i j , ℓ i j ( S i j , I i j ) . \Delta_{ij}^{cand}
=
\Phi_{ij}^{f_{ij},\ell_{ij}}
(S_{ij},I_{ij}). Δ ij c an d = Φ ij f ij , ℓ ij ( S ij , I ij ) .
36. Executor 先產生 Candidate
Execution Result = Candidate Delta \boxed{
\text{Execution Result}
=
\text{Candidate Delta}
} Execution Result = Candidate Delta
而不是直接 canonical mutation。
37. Candidate 不等 Commit
C a n d i d a t e ≠ C o m m i t . \boxed{
Candidate
\neq
Commit.
} C an d i d a t e = C o mmi t .
38. Candidate Store
所有:
Δ i j c a n d \Delta_{ij}^{cand} Δ ij c an d
可以先進 candidate store。
39. Why Candidate First?
因為 local executor 不知道完整 global constraints。
40. Local Success 不等 Global Admissibility
Local Execution Success ⇏ Global Commit Admissibility . \boxed{
\text{Local Execution Success}
\not\Rightarrow
\text{Global Commit Admissibility}.
} Local Execution Success ⇒ Global Commit Admissibility .
41. Global Constraint Set
C t G \mathcal C_t^G C t G
可以包含:
cross-domain dependencies;
cross-world causal coupling;
authority;
invariants;
budget;
temporal ordering;
evidence boundaries;
version conditions;
safety limits。
42. Global Reconciliation
R e c o n c i l e G ( { Δ i j c a n d } , C t G , W t G , H t ) \boxed{
\mathsf{Reconcile}_G
\left(
\{\Delta_{ij}^{cand}\},
\mathcal C_t^G,
\mathfrak W_t^G,
H_t
\right)
} Reconcile G ( { Δ ij c an d } , C t G , W t G , H t )
43. Reconcile 不等 Average
不同 candidates 衝突時,不是取平均。
44. Reconciliation Outcomes
可以是:
{ A c c e p t , R e j e c t , D e f e r , F o r k , R e t r y , R e c o m p u t e , E s c a l a t e } . \{
Accept,
Reject,
Defer,
Fork,
Retry,
Recompute,
Escalate
\}. { A cce pt , R e j ec t , D e f er , F or k , R e t r y , R eco m p u t e , E sc a l a t e } .
45. Accept
candidate 可進指定 world commit。
46. Reject
candidate 不符合 contract。
47. Defer
等待 dependency / evidence。
48. Fork
如果兩個合法 candidate 不可同時合併,可以產生新 worlds。
49. Retry
同配置重試。
50. Recompute
改 computational form、precision 或 backend。
51. Escalate
交給:
stronger solver;
proof system;
human;
external validator。
52. World Commit 不等 Reality Actuation
World Commit ≠ Reality Actuation . \boxed{
\text{World Commit}
\neq
\text{Reality Actuation}.
} World Commit = Reality Actuation .
simulation world commit 只改 simulation state。
53. Local Commit
C o m m i t W i . Commit_{W_i}. C o mmi t W i .
54. Cross-World Commit
如果同時更新多個 coupled worlds:
C o m m i t { W i , W j } . Commit_{\{W_i,W_j\}}. C o mmi t { W i , W j } .
需要更強 reconciliation。
55. Reality-Facing Commit
即使 actual-linked world 更新,
也不表示 physical effect 已發生。
56. Candidate / Commit Pipeline
R e s o l v e → R o u t e → E x e c u t e → C a n d i d a t e → R e c o n c i l e → C o m m i t → O b s e r v e . \boxed{
Resolve
\rightarrow
Route
\rightarrow
Execute
\rightarrow
Candidate
\rightarrow
Reconcile
\rightarrow
Commit
\rightarrow
Observe.
} R eso l v e → R o u t e → E x ec u t e → C an d i d a t e → R eco n c i l e → C o mmi t → O b ser v e .
57. WF-GCC
本文正式提出:
W F - G C C = ⟨ B o u n d a r y , D o m a i n s , C o n f i g u r a t i o n s , D e p e n d e n c i e s , B u d g e t s , C a n d i d a t e D e l t a s , R e c o n c i l i a t i o n , C o m m i t , H i s t o r y , C e r t i f i c a t e s ⟩ . \boxed{
\mathsf{WF\text{-}GCC}
=
\left\langle
Boundary,
Domains,
Configurations,
Dependencies,
Budgets,
CandidateDeltas,
Reconciliation,
Commit,
History,
Certificates
\right\rangle.
} WF - GCC = ⟨ B o u n d a r y , D o main s , C o n f i g u r a t i o n s , D e p e n d e n c i es , B u d g e t s , C an d i d a t eD e l t a s , R eco n c i l ia t i o n , C o mmi t , H i s t or y , C er t i f i c a t es ⟩ .
58. Boundary
指定這次 global coherence 的範圍。
59. Domains
本次 active computational domains。
60. Configurations
每個 domain 的:
κ i j , t . \kappa_{ij,t}. κ ij , t .
61. Dependencies
跨 domain / world dependency graph。
62. Budgets
包含:
compute;
memory;
energy;
bandwidth;
latency;
token/model;
human attention。
63. CandidateDeltas
未提交結果。
64. Reconciliation
全域一致性與衝突處理。
65. Commit
正式 world-state transition。
66. History
完整計算路徑。
67. Certificates
為何這條 route / commit 合法。
68. Finite Active Support
所有潛在可計算 pair:
U t c o m p . \mathcal U_t^{comp}. U t co m p .
active support:
A t c o m p ⊆ U t c o m p . \mathcal A_t^{comp}
\subseteq
\mathcal U_t^{comp}. A t co m p ⊆ U t co m p .
69. Runtime 必須有限
∣ A t c o m p ∣ < ∞ . \boxed{
|\mathcal A_t^{comp}|<\infty.
} ∣ A t co m p ∣ < ∞.
70. 但 Potential Space 可以擴張
U t + 1 c o m p ⊃ U t c o m p \mathcal U_{t+1}^{comp}
\supset
\mathcal U_t^{comp} U t + 1 co m p ⊃ U t co m p
可以成立。
71. 這不是矛盾
Finite Active Realization + Unbounded Extensibility . \boxed{
\text{Finite Active Realization}
+
\text{Unbounded Extensibility}.
} Finite Active Realization + Unbounded Extensibility .
72. Recursive Globality 不等 Recursive Full Expansion
nested worlds 可以存在,
但不要求:
expand every child world recursively . \boxed{
\text{expand every child world recursively}.
} expand every child world recursively .
73. Recursive Budget
對深度 d d d :
B d . B_d. B d .
可以隨深度下降。
74. World-Family Compute Budget
B t W F = ( B C , B M , B E , B L , B N ) . B_t^{WF}
=
(
B_C,
B_M,
B_E,
B_L,
B_N
). B t W F = ( B C , B M , B E , B L , B N ) .
概念上分別是 compute、memory、energy、latency、network budget。
75. Resource Allocation
∑ ( i , j ) ∈ A t c o m p b i j ≤ B t W F . \sum_{(i,j)\in\mathcal A_t^{comp}}
b_{ij}
\leq
B_t^{WF}. ( i , j ) ∈ A t co m p ∑ b ij ≤ B t W F .
76. Multi-Fidelity Computation
不同 worlds:
q i ≠ q j . q_i\neq q_j. q i = q j .
77. Coarse-to-Fine
先:
q i l o w q_i^{low} q i l o w
篩選 world。
再對重要 world:
q i h i g h . q_i^{high}. q i hi g h .
78. Coarse Result 不得冒充 High-Fidelity Result
Coarse Simulation ≠ High-Fidelity Validation . \boxed{
\text{Coarse Simulation}
\neq
\text{High-Fidelity Validation}.
} Coarse Simulation = High-Fidelity Validation .
79. Fidelity Escalation
q i l o w → q i m i d → q i h i g h . q_i^{low}
\rightarrow
q_i^{mid}
\rightarrow
q_i^{high}. q i l o w → q i mi d → q i hi g h .
80. Escalation Trigger
例如:
decision sensitivity;
risk;
disagreement;
uncertainty;
counterexample;
near-threshold result。
81. Four Resolution Fields
本文保留:
λ c o m p u t e ≠ λ o b s e r v e ≠ λ c a r r i e r ≠ λ r e n d e r . \boxed{
\lambda^{compute}
\neq
\lambda^{observe}
\neq
\lambda^{carrier}
\neq
\lambda^{render}.
} λ co m p u t e = λ o b ser v e = λ c a r r i er = λ r e n d er .
82. Compute Resolution
系統實際算多細。
83. Observe Resolution
observer 需要看多細。
84. Carrier Resolution
載體可承載多細。
85. Render Resolution
最終顯示多細。
86. 高精度 Compute 可以低精度 Observe
例如 solver 算到:
10 − 9 10^{-9} 1 0 − 9
但人只需要:
safe / unsafe . \text{safe / unsafe}. safe / unsafe .
87. 低精度 Compute 不可假裝高精度 Observe
如果 underlying compute 很粗,
UI 不能用華麗視覺假裝精確。
88. Materialization 與 Computation 分離
某 dependency 可以存在於 canonical graph,
但不必 active materialize。
89. Retrieval 可以替代 Recomputation
如果:
R e s u l t Result R es u l t
仍有效且 provenance compatible,
Router 可以選 retrieval。
90. Cache 不是 Truth
Cached Result ≠ Currently Valid Result . \boxed{
\text{Cached Result}
\neq
\text{Currently Valid Result}.
} Cached Result = Currently Valid Result .
需要 version / dependency check。
91. Shadow Execution
高風險新 route 可以先:
S h a d o w E x e c u t e . ShadowExecute. S ha d o w E x ec u t e .
92. Shadow Result 不 Commit
Shadow Result ≠ Canonical State . \boxed{
\text{Shadow Result}
\neq
\text{Canonical State}.
} Shadow Result = Canonical State .
93. Route Migration
若 backend 過載:
r a → r b . r_a
\rightarrow
r_b. r a → r b .
94. Backend Migration 不應改變 Semantic Contract
除非明示 approximate mode。
95. Heterogeneous Hardware 只是其中一層
CPU / GPU / FPGA / NPU heterogeneity 是:
R e x e c \mathfrak R^{exec} R e x ec
的異質性。
96. Semantic Heterogeneity 更上層
World family 還有:
algorithm heterogeneity;
transition-law heterogeneity;
domain heterogeneity;
precision heterogeneity;
evidence-mode heterogeneity。
97. External Engineering Analogy
2026 task-based data-flow work 已顯示同一應用可以在統一 runtime 下協調 CUDA、SYCL、Triton、OpenMP 與 vendor libraries。
這支持:
heterogeneous execution can be coordinated without forcing one backend . \boxed{
\text{heterogeneous execution can be coordinated without forcing one backend}.
} heterogeneous execution can be coordinated without forcing one backend .
98. 但本文層級更高
該類 runtime 主要處理 device / API heterogeneity。
WCO Computation Layer 還要選:
computational form;
world;
domain;
law;
precision;
evidence contract。
99. Computing-Aware Routing 類比
2025 CaRCS 類工作已把:
routing + scheduling \text{routing}
+
\text{scheduling} routing + scheduling
聯合優化。
本文接受這個工程方向。
但 WCO routing 更偏 semantic computational routing。
100. Real-Time Constraint
某些 world/domain 有 deadline:
T i j e x e c ≤ T i j d e a d l i n e . T_{ij}^{exec}
\leq
T_{ij}^{deadline}. T ij e x ec ≤ T ij d e a d l in e .
101. Deadline 會改變 Route
可能從高精度 solver 改為 bounded approximation。
102. Approximation 必須被標記
Approximate Result ≠ Exact Result . \boxed{
\text{Approximate Result}
\neq
\text{Exact Result}.
} Approximate Result = Exact Result .
103. Compute Debt
本文定義:
D c o m p = ( D a p p r o x , D s t a l e , D r o u t e , D s y n c , D b a c k e n d , D b u d g e t , D h i s t o r y ) . \boxed{
\mathbf D_{comp}
=
(
D_{approx},
D_{stale},
D_{route},
D_{sync},
D_{backend},
D_{budget},
D_{history}
).
} D co m p = ( D a pp r o x , D s t a l e , D r o u t e , D sy n c , D ba c k e n d , D b u d g e t , D hi s t or y ) .
104. Approximation Debt
數值/模型近似造成。
105. Staleness Debt
依賴過期 state 或 cache。
106. Routing Debt
選擇 suboptimal / mismatched computational form。
107. Synchronization Debt
asynchronous regions 暫時不一致。
108. Backend Debt
backend-specific limitation。
109. Budget Debt
因資源限制未完成理想計算。
110. History Debt
關鍵 route / version / seed 未完整保存。
111. Computation History
本文定義:
H t c o m p . H_t^{comp}. H t co m p .
至少保存:
route;
configuration;
backend;
versions;
seed;
inputs;
candidate deltas;
reconciliation;
commit receipt;
failure。
112. State Equality 不等 Computation-History Equality
S a = S b ⇏ H a c o m p = H b c o m p . \boxed{
S_a=S_b
\not\Rightarrow
H_a^{comp}=H_b^{comp}.
} S a = S b ⇒ H a co m p = H b co m p .
113. 為什麼歷史重要?
因為未來可達性可能受:
hidden cache;
learned state;
random seeds;
authority;
provenance;
irreversible external effects;
影響。
114. Noncommutative Computation
可能:
Φ a ∘ Φ b ≠ Φ b ∘ Φ a . \Phi_a\circ\Phi_b
\neq
\Phi_b\circ\Phi_a. Φ a ∘ Φ b = Φ b ∘ Φ a .
115. 因此終點相同仍可能不是同一執行
Endpoint Equality ≠ Execution Equivalence . \boxed{
\text{Endpoint Equality}
\neq
\text{Execution Equivalence}.
} Endpoint Equality = Execution Equivalence .
116. Computation Equivalence 需要 Task Specification
可定義:
H a c o m p ≡ τ H b c o m p H_a^{comp}
\equiv_{\tau}
H_b^{comp} H a co m p ≡ τ H b co m p
表示對 task τ \tau τ 所需 invariants 等價。
117. Global Reconciliation 也要看 History
不能只比較 final values。
118. Cross-World Coupled Computation
某些 worlds 故意共享:
random numbers;
model;
parameters;
以降低比較 variance。
119. Coupling 不等 Error
paired counterfactual simulation 可以合理共享 noise。
120. 但 Coupling 必須入 Ledger
否則會誤判 evidence independence。
121. Parallel Worlds 不等 Independent Worlds
Parallel Execution ≠ Epistemic Independence . \boxed{
\text{Parallel Execution}
\neq
\text{Epistemic Independence}.
} Parallel Execution = Epistemic Independence .
122. Parallelism 是 Execution Property
independence 是 evidence / model property。
123. AIVS 接口
當 active compute regions 很多時,
中央 AI 不應收全部 raw trace。
124. Hierarchical Relay
W o r k e r → R e l a y → G o v e r n o r . \boxed{
Worker
\rightarrow
Relay
\rightarrow
Governor.
} W or k er → R e l a y → G o v er n or .
125. Vertical Synchronization
只上送:
anomaly;
summary;
causal conflict;
threshold crossing;
commit request。
126. No Global Barrier by Default
Global Coherence ≠ Global Barrier . \boxed{
\text{Global Coherence}
\neq
\text{Global Barrier}.
} Global Coherence = Global Barrier .
127. Consistency Mode 應 Task-Relative
有些 domain 需要 strict consistency。
有些可 eventual reconciliation。
128. Strong Consistency Cost
C s y n c s t r o n g C_{sync}^{strong} C sy n c s t r o n g
可能很高。
129. Weak Consistency Risk
R s y n c w e a k R_{sync}^{weak} R sy n c w e ak
可能增加 temporary conflict。
130. Router 應做 Trade-off
σ ∗ = arg min σ ( C s y n c + λ R s y n c ) . \sigma^\ast
=
\operatorname*{arg\,min}_{\sigma}
\left(
C_{sync}+\lambda R_{sync}
\right). σ ∗ = σ arg min ( C sy n c + λ R sy n c ) .
131. Multi-Objective Routing
一般不應假設單一 cost。
可定義:
J ( κ ) = ( Q , L a t e n c y , E n e r g y , C o s t , R i s k , V e r i f i a b i l i t y , R e p r o d u c i b i l i t y ) . \mathbf J(\kappa)
=
(
Q,
Latency,
Energy,
Cost,
Risk,
Verifiability,
Reproducibility
). J ( κ ) = ( Q , L a t e n cy , E n er g y , C os t , R i s k , V er i f iabi l i t y , R e p r o d u c ibi l i t y ) .
132. Pareto Frontier
選擇:
κ ∈ P P a r e t o . \kappa
\in
\mathcal P_{\mathrm{Pareto}}. κ ∈ P Pareto .
133. Policy 才做 Scalarization
只有在 policy 明確時:
J w = ∑ i w i J i . J_w
=
\sum_iw_iJ_i. J w = i ∑ w i J i .
134. Weight 需要 Governance
不能由 renderer 或 executor 偷偷決定。
135. Risk-Aware Routing
高風險:
R i s k ↑ Risk\uparrow R i s k ↑
通常要求:
stronger validation;
higher precision;
more provenance;
stricter commit gate。
136. 但 Risk-Aware 不等一律最慢
低 latency 本身也可能是 safety requirement。
137. Value of Information
某 computation 是否值得繼續,可以看:
V o I ( κ ) . VoI(\kappa). V o I ( κ ) .
138. Stop 是正式動作
S t o p ∈ valid computation policy . \boxed{
\mathsf{Stop}
\in
\text{valid computation policy}.
} Stop ∈ valid computation policy .
139. 不計算也是計算治理的一部分
如果:
V o I < C c o m p u t e , VoI<C_{compute}, V o I < C co m p u t e ,
可以停止。
140. Replan
新 evidence 可以使 route:
κ t → κ t + 1 . \kappa_t
\rightarrow
\kappa_{t+1}. κ t → κ t + 1 .
141. Reframe
甚至 computational domain 本身可以改變:
D t → D t + 1 ′ . D_t
\rightarrow
D_{t+1}'. D t → D t + 1 ′ .
142. Global-Like AI 的計算智能
不只:
算得快。
更包括:
知道現在該算什麼、用什麼算、算到哪裡、何時停止、何時換方法。
143. Computational Routing Certificate
本文提出:
R o u t e C e r t = ⟨ T a s k , W o r l d , D o m a i n , C o n f i g u r a t i o n , A l t e r n a t i v e s , C o n s t r a i n t s , B u d g e t , R i s k , R e a s o n , V e r s i o n ⟩ . \boxed{
\mathsf{RouteCert}
=
\left\langle
Task,
World,
Domain,
Configuration,
Alternatives,
Constraints,
Budget,
Risk,
Reason,
Version
\right\rangle.
} RouteCert = ⟨ T a s k , W or l d , D o main , C o n f i g u r a t i o n , A l t er na t i v es , C o n s t r ain t s , B u d g e t , R i s k , R e a so n , V er s i o n ⟩ .
144. Candidate Certificate
C a n d C e r t = ( I n p u t V e r s i o n , R o u t e C e r t , E x e c u t o r , R e s u l t , D e b t , F a i l u r e S t a t e ) . \mathsf{CandCert}
=
(
InputVersion,
RouteCert,
Executor,
Result,
Debt,
FailureState
). CandCert = ( I n p u t V er s i o n , R o u t e C er t , E x ec u t or , R es u l t , D e b t , F ai l u r e S t a t e ) .
145. Reconciliation Certificate
R e c o n C e r t = ( C a n d i d a t e S e t , C o n s t r a i n t s , C o n f l i c t s , R e s o l u t i o n , A c c e p t e d , R e j e c t e d , D e f e r r e d ) . \mathsf{ReconCert}
=
(
CandidateSet,
Constraints,
Conflicts,
Resolution,
Accepted,
Rejected,
Deferred
). ReconCert = ( C an d i d a t e S e t , C o n s t r ain t s , C o n f l i c t s , R eso l u t i o n , A cce pt e d , R e j ec t e d , D e f er r e d ) .
146. Commit Receipt
C o m m i t R e c e i p t = ( W o r l d I d , B e f o r e V e r s i o n , A f t e r V e r s i o n , A c c e p t e d D e l t a , A u t h o r i t y , T i m e s t a m p , H i s t o r y R e f ) . \mathsf{CommitReceipt}
=
(
WorldId,
BeforeVersion,
AfterVersion,
AcceptedDelta,
Authority,
Timestamp,
HistoryRef
). CommitReceipt = ( W or l d I d , B e f or e V er s i o n , A f t er V er s i o n , A cce pt e d D e l t a , A u t h or i t y , T im es t am p , H i s t or y R e f ) .
147. Certificate-Carrying Computation
High-Risk Computation → Result + Certificate . \boxed{
\text{High-Risk Computation}
\rightarrow
\text{Result + Certificate}.
} High-Risk Computation → Result + Certificate .
148. Failure 也要結構化
F a i l u r e = ( S t a g e , R e a s o n , M i s s i n g C o n d i t i o n , R e c o v e r a b i l i t y , A f f e c t e d W o r l d s ) . \mathsf{Failure}
=
(
Stage,
Reason,
MissingCondition,
Recoverability,
AffectedWorlds
). Failure = ( S t a g e , R e a so n , M i ss in g C o n d i t i o n , R eco v er abi l i t y , A f f ec t e d W or l d s ) .
149. Failure 不等 Zero
不要用:
0 0 0
代表所有 failure。
150. Recovery Ladder
可以依序:
R e t r y → R e r o u t e → R e c o m p u t e → R o l l b a c k → C o m p e n s a t e → E s c a l a t e . Retry
\rightarrow
Reroute
\rightarrow
Recompute
\rightarrow
Rollback
\rightarrow
Compensate
\rightarrow
Escalate. R e t r y → R er o u t e → R eco m p u t e → R o l l ba c k → C o m p e n s a t e → E sc a l a t e .
151. Rollback 只對可回退 Computational State 成立
若已有 physical effect:
Compute Rollback ≠ Reality Rollback . \boxed{
\text{Compute Rollback}
\neq
\text{Reality Rollback}.
} Compute Rollback = Reality Rollback .
152. World-Family Computation Loop
W t G → R e s o l v e D t a c t i v e → R o u t e { κ i j , t } → E x e c u t e { Δ i j c a n d } → R e c o n c i l e { Δ a c c , Δ r e j , Δ d e f } → C o m m i t W t + 1 G → O b s e r v e E t + 1 . \boxed{
\begin{aligned}
\mathfrak W_t^G
&\xrightarrow{\mathsf{Resolve}}
\mathcal D_t^{active}
\\
&\xrightarrow{\mathsf{Route}}
\{\kappa_{ij,t}\}
\\
&\xrightarrow{\mathsf{Execute}}
\{\Delta_{ij}^{cand}\}
\\
&\xrightarrow{\mathsf{Reconcile}}
\{\Delta^{acc},\Delta^{rej},\Delta^{def}\}
\\
&\xrightarrow{\mathsf{Commit}}
\mathfrak W_{t+1}^{G}
\\
&\xrightarrow{\mathsf{Observe}}
E_{t+1}.
\end{aligned}
} W t G Resolve D t a c t i v e Route { κ ij , t } Execute { Δ ij c an d } Reconcile { Δ a cc , Δ r e j , Δ d e f } Commit W t + 1 G Observe E t + 1 .
153. Observe 在最後不代表只能最後看
執行中也可以有 monitoring observations。
Paper 04 會細化。
154. Compute / Observe 分離仍保持
Monitoring Observation ≠ Computation State itself . \boxed{
\text{Monitoring Observation}
\neq
\text{Computation State itself}.
} Monitoring Observation = Computation State itself .
155. Computation Layer 的 World-Family State
本文可寫:
C t W F = ⟨ W t G , D t , F t c o m p , L t , Λ t c o m p , R t e x e c , S t , C t G , A t c o m p , H t c o m p ⟩ . \boxed{
\mathfrak C_t^{WF}
=
\left\langle
\mathfrak W_t^G,
\mathcal D_t,
\mathfrak F_t^{comp},
\mathfrak L_t,
\Lambda_t^{comp},
\mathfrak R_t^{exec},
\mathfrak S_t,
\mathcal C_t^G,
\mathcal A_t^{comp},
H_t^{comp}
\right\rangle.
} C t W F = ⟨ W t G , D t , F t co m p , L t , Λ t co m p , R t e x ec , S t , C t G , A t co m p , H t co m p ⟩ .
156. Λ t c o m p \Lambda_t^{comp} Λ t co m p
compute resolution / materialization field。
157. A t c o m p \mathcal A_t^{comp} A t co m p
finite active computation support。
158. H t c o m p H_t^{comp} H t co m p
computation history。
159. WCO Paper 01 的 C t \mathfrak C_t C t 正式升級
Paper 01:
C t . \mathfrak C_t. C t .
Paper 03:
C t ⇝ C t W F . \boxed{
\mathfrak C_t
\rightsquigarrow
\mathfrak C_t^{WF}.
} C t ⇝ C t W F .
160. 計算層與世界層的接口
World Layer 提供:
identity;
mode;
state;
lineage;
authority;
evidence;
lifecycle。
Computation Layer 不得繞過這些 contract。
161. 計算層不能自行創造 Authority
Compute Capability ≠ Commit Authority . \boxed{
\text{Compute Capability}
\neq
\text{Commit Authority}.
} Compute Capability = Commit Authority .
162. 計算層不能自行升格 Evidence
Computed Result ≠ Verified Evidence . \boxed{
\text{Computed Result}
\neq
\text{Verified Evidence}.
} Computed Result = Verified Evidence .
163. 計算層不能把 Branch 當 Reality
S I M ≠ A L ≠ R . \boxed{
SIM
\neq
AL
\neq
\mathcal R.
} S I M = A L = R .
164. 計算層不能把 Approximation 藏起來
D a p p r o x must remain visible to later layers . \boxed{
D_{approx}
\text{ must remain visible to later layers}.
} D a pp r o x must remain visible to later layers .
165. Computation Layer Constitution
本文提出十條:
Globality is boundary-relative.
Heterogeneity is allowed.
Routing is broader than scheduling.
Local success requires global reconciliation.
Candidate is not commit.
Finite active support is mandatory.
Resolution layers are separable.
History is first-class.
Certificates travel with high-risk computation.
Computation cannot create its own authority.
166. MVP:四世界、五種計算形態
使用 Paper 02 的:
W A , W B , W C , W N . W_A,W_B,W_C,W_N. W A , W B , W C , W N .
167. Computational Forms
提供:
deterministic simulator;
Monte Carlo;
graph analysis;
symbolic constraints;
learned surrogate model。
168. Router
根據:
world mode;
task;
risk;
budget;
選:
f i . f_i. f i .
169. Multi-Fidelity
每個 world 先 low fidelity。
170. Escalation
只有 top-risk / top-disagreement worlds 升級 high fidelity。
171. Candidate Store
所有結果先不 commit。
172. Reconciliation
檢查:
shared resource;
contradictory invariants;
state version;
world mode;
evidence status。
173. Commit
更新 simulation worlds。
不觸發 reality action。
174. MVP Metrics
測:
quality;
compute;
latency;
energy;
route accuracy;
reroute rate;
reconciliation conflicts;
human intervention;
certificate completeness。
175. 實驗一:One Form Everywhere vs Routed Heterogeneity
Baseline:
f i = f 0 f_i=f_0 f i = f 0
for all domains。
比較 dynamic routing。
176. 實驗二:Fixed Fidelity vs Adaptive Fidelity
比較:
q i = q m a x q_i=q_{max} q i = q ma x
與 coarse-to-fine。
177. 實驗三:Local Commit vs Global Reconciliation
測沒有 reconciliation 時的 cross-world / cross-domain conflict。
178. 實驗四:Schedule-only vs Computational Routing
控制相同 hardware budget。
比較只決定 when/where 與同時決定 how。
179. 實驗五:History-Aware vs Endpoint-Only
建立相同 endpoint、不同 route history 的 cases。
測後續 recovery / audit。
180. 實驗六:Strong Barrier vs AIVS
比較:
global barrier;
hierarchical relay;
asynchronous reconciliation。
181. 實驗七:Cache Validity
故意使 upstream version 改變。
測 runtime 是否拒絕 stale cached result。
182. 實驗八:Route Failure Recovery
故意讓 backend 失效。
測:
R e t r y → R e r o u t e → R e c o m p u t e . Retry
\rightarrow
Reroute
\rightarrow
Recompute. R e t r y → R er o u t e → R eco m p u t e .
183. 實驗九:Risk-Aware Routing
高風險 world 與低風險 synthetic world 使用不同 validation level。
184. 實驗十:World-Family Scaling
逐步增加:
∣ W ∣ |\mathfrak W| ∣ W ∣
測 active support 是否維持 bounded。
185. 可反駁性
本文會被削弱,如果:
routed heterogeneity 在控制總 compute 後沒有穩定收益;
global reconciliation 在代表性 coupled tasks 中沒有降低 conflict;
finite active support 無法阻止 world-family scaling 的資源爆炸;
history preservation 對 recovery / audit 沒有價值;
route certificates 無法提高 reproducibility;
multi-fidelity routing 只增加 overhead;
simpler schedule-only runtime 在所有代表性 tasks 上同等有效。
186. 外部研究接口
2026 的 task-based data-flow methodology 已展示,一個應用可在 OmpSs-2 / OpenMP-style task graph 下協調 CUDA、SYCL、Triton、OpenMP offload 與 vendor libraries,並透過統一 threading/runtime 機制降低多 runtime 競爭。這證明 heterogeneous execution orchestration 已是現實工程問題。
2025 的 computing-aware routing / collaborative scheduling 工作也證明 routing 與 scheduling 可以被聯合建模,而非互相獨立。
2025 的 accelerator-based heterogeneous real-time scheduling survey 則顯示 deadline、energy、thermal constraints 與 heterogeneous device characteristics 本來就會改變 scheduler 的決策。
本文不宣稱取代上述研究。它在更高語義層追問:
如果除了 hardware heterogeneity,連 world、domain、computational form、transition law、precision、evidence mode 與 authority 都是可變項,AI runtime 應如何組合它們?
187. 本文不主張什麼
本文不主張:
24/72 是完備的計算分類;
存在單一普遍最優 routing policy;
所有 task 都需要 heterogeneous computation;
AI 應取代 OS scheduler;
AI 應取代 compiler;
AI 應取代 numerical solver;
GPU 一定優於 CPU;
learned model 一定優於 symbolic solver;
global reconciliation 必須是 global barrier;
approximate compute 一定不安全;
strong consistency 一定更好;
more compute 一定提高 quality;
more worlds 一定需要更多 active compute;
parallel execution 等於 independent evidence;
endpoint equality 等於 history equality;
world commit 等於 reality actuation;
candidate result 等於 verified evidence;
routing layer 可以自行取得 authority;
WF-GCC 已完成 production implementation;
本文取代 GCM、CDI/AIVS、PNCW、heterogeneous runtime literature 或 operating-system scheduling theory。
188. Computation Layer 核心非同一性
Computational Form ≠ Projection Form . \boxed{
\text{Computational Form}
\neq
\text{Projection Form}.
} Computational Form = Projection Form .
Scheduling ⊂ Computational Routing . \boxed{
\text{Scheduling}
\subset
\text{Computational Routing}.
} Scheduling ⊂ Computational Routing .
Computation ≠ Observation . \boxed{
\text{Computation}
\neq
\text{Observation}.
} Computation = Observation .
Global Dependency ≠ Full Materialization . \boxed{
\text{Global Dependency}
\neq
\text{Full Materialization}.
} Global Dependency = Full Materialization .
Local Success ⇏ Global Commit . \boxed{
\text{Local Success}
\not\Rightarrow
\text{Global Commit}.
} Local Success ⇒ Global Commit .
C a n d i d a t e ≠ C o m m i t ≠ R e a l i t y A c t u a t i o n . \boxed{
Candidate
\neq
Commit
\neq
RealityActuation.
} C an d i d a t e = C o mmi t = R e a l i t y A c t u a t i o n .
Parallel Execution ≠ Epistemic Independence . \boxed{
\text{Parallel Execution}
\neq
\text{Epistemic Independence}.
} Parallel Execution = Epistemic Independence .
State Equality ⇏ Computation-History Equality . \boxed{
\text{State Equality}
\not\Rightarrow
\text{Computation-History Equality}.
} State Equality ⇒ Computation-History Equality .
189. 核心母式一:計算配置
κ i j , t = ⟨ W i , D i j , f i j , ℓ i j , λ i j c o m p , r i j , τ i j , q i j , σ i j ⟩ . \boxed{
\kappa_{ij,t}
=
\left\langle
W_i,
D_{ij},
f_{ij},
\ell_{ij},
\lambda_{ij}^{comp},
r_{ij},
\tau_{ij},
q_{ij},
\sigma_{ij}
\right\rangle.
} κ ij , t = ⟨ W i , D ij , f ij , ℓ ij , λ ij co m p , r ij , τ ij , q ij , σ ij ⟩ .
190. 核心母式二:路由
R t W F C : ( W t G , D , T , B t , R i s k t , Γ t ) ↦ κ t . \boxed{
\mathcal R_t^{WFC}
:
(
\mathfrak W_t^G,
D,
T,
B_t,
Risk_t,
\Gamma_t
)
\mapsto
\kappa_t.
} R t W F C : ( W t G , D , T , B t , R i s k t , Γ t ) ↦ κ t .
191. 核心母式三:全域 Reconciliation
R e c o n c i l e G ( { Δ i j c a n d } , C t G , W t G , H t ) . \boxed{
\mathsf{Reconcile}_G
\left(
\{\Delta_{ij}^{cand}\},
\mathcal C_t^G,
\mathfrak W_t^G,
H_t
\right).
} Reconcile G ( { Δ ij c an d } , C t G , W t G , H t ) .
192. 核心母式四:WF-GCC
W F - G C C = ⟨ B o u n d a r y , D o m a i n s , C o n f i g u r a t i o n s , D e p e n d e n c i e s , B u d g e t s , C a n d i d a t e D e l t a s , R e c o n c i l i a t i o n , C o m m i t , H i s t o r y , C e r t i f i c a t e s ⟩ . \boxed{
\mathsf{WF\text{-}GCC}
=
\left\langle
Boundary,
Domains,
Configurations,
Dependencies,
Budgets,
CandidateDeltas,
Reconciliation,
Commit,
History,
Certificates
\right\rangle.
} WF - GCC = ⟨ B o u n d a r y , D o main s , C o n f i g u r a t i o n s , D e p e n d e n c i es , B u d g e t s , C an d i d a t eD e l t a s , R eco n c i l ia t i o n , C o mmi t , H i s t or y , C er t i f i c a t es ⟩ .
193. 核心母式五:有限活動計算
A t c o m p ⊆ U t c o m p , ∣ A t c o m p ∣ < ∞ . \boxed{
\mathcal A_t^{comp}
\subseteq
\mathcal U_t^{comp},
\qquad
|\mathcal A_t^{comp}|<\infty.
} A t co m p ⊆ U t co m p , ∣ A t co m p ∣ < ∞.
同時允許:
U t + 1 c o m p ⊃ U t c o m p . \boxed{
\mathcal U_{t+1}^{comp}
\supset
\mathcal U_t^{comp}.
} U t + 1 co m p ⊃ U t co m p .
194. 核心母式六:World-Family Computation Loop
W t G → R e s o l v e D t a c t i v e → R o u t e { κ i j , t } → E x e c u t e { Δ i j c a n d } → R e c o n c i l e { Δ a c c , Δ r e j , Δ d e f } → C o m m i t W t + 1 G . \boxed{
\begin{aligned}
\mathfrak W_t^G
&\xrightarrow{\mathsf{Resolve}}
\mathcal D_t^{active}
\\
&\xrightarrow{\mathsf{Route}}
\{\kappa_{ij,t}\}
\\
&\xrightarrow{\mathsf{Execute}}
\{\Delta_{ij}^{cand}\}
\\
&\xrightarrow{\mathsf{Reconcile}}
\{\Delta^{acc},\Delta^{rej},\Delta^{def}\}
\\
&\xrightarrow{\mathsf{Commit}}
\mathfrak W_{t+1}^{G}.
\end{aligned}
} W t G Resolve D t a c t i v e Route { κ ij , t } Execute { Δ ij c an d } Reconcile { Δ a cc , Δ r e j , Δ d e f } Commit W t + 1 G .
195. 結論:類全域 AI 不只是「算得更多」,而是「知道世界族該怎麼被算」
普通 compute scheduler 問:
哪個 task 先跑?
heterogeneous scheduler 再問:
跑在哪個 CPU / GPU / accelerator?
類全域 AI 的 Computation Layer 還必須問:
這是哪一個 world?
哪一個 domain?
該使用哪一種 computational form?
該使用哪種 transition law?
要算到什麼 resolution?
是 exact、approximate、retrieval 還是 simulation?
是否值得先用 coarse solver?
是否可以 parallel?
是否需要 strict synchronization?
結果只能作 candidate,還是可以 local commit?
這些 local candidates 是否和其他 worlds 的 constraint 衝突?
這條 route 的 history、backend、seed、precision 和 approximation 有沒有被保存?
因此本文把全域計算重新濃縮為:
Global Computation = Route Heterogeneously + Execute Locally + Reconcile Globally + Commit Explicitly . \boxed{
\text{Global Computation}
=
\text{Route Heterogeneously}
+
\text{Execute Locally}
+
\text{Reconcile Globally}
+
\text{Commit Explicitly}.
} Global Computation = Route Heterogeneously + Execute Locally + Reconcile Globally + Commit Explicitly .
它不是一個巨大、同質、全同步的計算。
它是一個在受治理世界族中,依 task、world、domain、risk 與 budget 持續重配置的計算組合過程。
因此:
Global-Like AI computation is not one giant calculation. \boxed{
\text{Global-Like AI computation is not one giant calculation.}
} Global-Like AI computation is not one giant calculation.
而是:
a governed routing-and-reconciliation process over heterogeneous computations distributed across a family of worlds . \boxed{
\text{a governed routing-and-reconciliation process
over heterogeneous computations distributed across a family of worlds}.
} a governed routing-and-reconciliation process over heterogeneous computations distributed across a family of worlds .
Paper 01 的:
C t \mathfrak C_t C t
因此在本文正式升級為:
C t W F . \boxed{
\mathfrak C_t^{WF}.
} C t W F .
下一篇將轉向 WCO 三重族的第三個核心:
Paper 04
觀察層:Global Observer 與 Observation Operator Family
也就是:
世界已經建立、計算也已經進行後,AI 到底應該看哪裡、看什麼、用什麼 observation operator,又如何開始自行發明自己的 computational way of seeing?
196. 下一篇接口
Paper 04 將正式處理:
observer identity;
observer-relative slice;
Global Observer;
observation family;
active observation;
passive observation;
internal / external observation;
cross-world observation;
observation accessibility;
observation budget;
observer-specific resolution;
observation operator selection;
AI-native observer emergence;
computational way of seeing;
observation debt;
observation vs projection。
參考文獻與內部前置研究
EveMissLab / Neo.K
Neo.K × Aletheia,《從計算 24/72 範式到全域計算方法論:總綱與交接索引》v0.1,2026。
Neo.K × Aletheia,《全域計算方法論:異質計算的全域一致組合》v0.1,2026。
Neo.K × Aletheia,《計算形態空間:從 24 範式與 72 格動力學到可路由計算配置》v0.1,2026。
Neo.K × Aletheia,《動態計算路由:多域、多範式與異質轉移律的 Runtime 組合》v0.1,2026。
Neo.K × Aletheia,《計算不等於觀察:全域演化、局部物化與解析度相對計算》v0.1,2026。
Neo.K × Aletheia,《有限活動實現與無界計算展開》v0.1,2026。
Neo.K × Aletheia,《終點不等於歷史:非交換計算序列、世界狀態與可追溯全域演化》v0.1,2026。
Neo.K × Aletheia,《CDI / AIVS Series》,2026。
Neo.K × Aletheia,《PNCW Paper 05:全域計算、局部顯現》v0.1,2026。
Neo.K × Aletheia,《WCO Series Paper 01》v0.1,2026。
Neo.K × Aletheia,《WCO Series Paper 02:Governed World Family》v0.1,2026。
Neo.K × Aletheia,《SWFR Paper 02》v0.1,2026。
External Research Interfaces
Boné, A., Aguirre, A., Álvarez, D., Martínez-Ferrer, P. J., & Beltran, V. (2026). A task-based data-flow methodology for programming heterogeneous systems with multiple accelerator APIs . Future Generation Computer Systems, 180, 108383.
Feng, L., Xie, R., Tang, Q., et al. (2025). CaRCS: Joint Optimization of Computing-Aware Routing and Collaborative Scheduling in Computing Power Networks . IEEE Network, 39(6), 270–278.
Zou, A., Xu, Y., Ni, Y., et al. (2025). A Survey of Real-time Scheduling on Accelerator-based Heterogeneous Architecture for Time Critical Applications . arXiv:2505.11970.
Bauer, M., Treichler, S., Slaughter, E., & Aiken, A. (2012). Legion: Expressing Locality and Independence with Logical Regions . SC '12.
Augonnet, C., Thibault, S., Namyst, R., & Wacrenier, P.-A. (2011). StarPU: A Unified Platform for Task Scheduling on Heterogeneous Multicore Architectures . Concurrency and Computation: Practice and Experience, 23(2), 187–198.
Paper 03 狀態:COMPLETE v0.1 下一篇:Paper 04 — 觀察層:Global Observer 與 Observation Operator Family Canonical source:UTF-8 Markdown;數學 delimiter 僅使用 $...$ 與 $$...$$。