模擬世界族執行時(Paper 02)
AI 對多重虛擬世界的非對偶、非對稱觀察與控制
Simulation World Family Runtime: Non-Dual, Asymmetric Observation and Control over Multiple Virtual Worlds
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定位: 載體投影與內視系列之工程延伸/PPOE Paper 02版本: v0.1日期: 2026-09-08研究定位: world model × game engine × simulation runtime × AI agent × branching computation × counterfactual reasoning × digital twin × authority architecture狀態: 統合架構與形式框架;承接既有 WDC/BWC、HDUS、PNCW 與 PPOE
摘要
Unity、Unreal Engine、Godot 等遊戲引擎早已實作一個重要但常被低估的工程模式:維持可執行 world state,讓 physics、rules、agents 與 event loop 推動世界演化,再透過 camera、renderer、minimap、collision view、navigation view、debug overlay 等不同 observation/projection path,將同一世界投影給不同觀察者。因此遊戲引擎不只是「3D 圖形程式」,而可以被理解為早期的:
World-State Execution + Projection Runtime . \boxed{
\text{World-State Execution}
+
\text{Projection Runtime}.
} World-State Execution + Projection Runtime .
本文進一步提出:未來 AI 不必只在單一虛擬世界裡觀察與行動,也不必把 world model 理解為單一 predictive state。AI 可以同時建立、執行、觀察、比較、嵌套、分支、修剪與治理一整個模擬世界族:
W = { W 0 , W 1 , … , W n } . \boxed{
\mathfrak W
=
\{W_0,W_1,\ldots,W_n\}.
} W = { W 0 , W 1 , … , W n } .
每個 world 具有自己的 state、transition law、time domain、agents、evidence state、authority boundary 與 projection family:
W i = ( S i , Φ i , T i , A i , E i , G i , P i ) . W_i
=
\left(
S_i,
\Phi_i,
T_i,
\mathcal A_i,
\mathcal E_i,
\mathcal G_i,
\mathcal P_i
\right). W i = ( S i , Φ i , T i , A i , E i , G i , P i ) .
本文稱此架構為 Simulation World Family Runtime(SWFR,模擬世界族執行時) 。
SWFR 的核心不是「多開幾個遊戲實例」,而是建立具型別、具方向、具 lineage、具 authority 與具 evidence semantics 的 world graph:
G W = ( V W , E W ) . \boxed{
G_{\mathfrak W}
=
(V_{\mathfrak W},E_{\mathfrak W}).
} G W = ( V W , E W ) .
world-to-world 邊具有明確型別:
e i j ∈ { O b s e r v e , P r o j e c t , F o r k , E m b e d , Q u e r y , C o u p l e , C o n t r o l , M e r g e , P r u n e , P r o m o t e , T r a n s p o r t E v i d e n c e } . e_{ij}
\in
\{
\mathsf{Observe},
\mathsf{Project},
\mathsf{Fork},
\mathsf{Embed},
\mathsf{Query},
\mathsf{Couple},
\mathsf{Control},
\mathsf{Merge},
\mathsf{Prune},
\mathsf{Promote},
\mathsf{TransportEvidence}
\}. e ij ∈ { Observe , Project , Fork , Embed , Query , Couple , Control , Merge , Prune , Promote , TransportEvidence } .
本文特別提出兩條統一原則:非對偶(non-duality)與 非對稱(asymmetry) 。
一般投影:
Π : W → P \Pi:
W
\rightarrow
P Π : W → P
通常是 many-to-one,因此可能:
W a ≠ W b W_a\neq W_b W a = W b
但:
Π ( W a ) = Π ( W b ) . \Pi(W_a)=\Pi(W_b). Π ( W a ) = Π ( W b ) .
所以:
Π − 1 need not exist . \boxed{
\Pi^{-1}
\text{ need not exist}.
} Π − 1 need not exist .
Observation 也不等於 control 的逆:
O b s e r v e ≠ C o n t r o l − 1 . \boxed{
\mathsf{Observe}
\neq
\mathsf{Control}^{-1}.
} Observe = Control − 1 .
另一方面,world-to-world observation、authority、causal coupling 與 evidence transport 可以天然非對稱:
O i j ≠ O j i , O_{ij}\neq O_{ji}, O ij = O j i ,
A i j ≠ A j i , A_{ij}\neq A_{ji}, A ij = A j i ,
K i j ≠ K j i , K_{ij}\neq K_{ji}, K ij = K j i ,
E i j ≠ E j i . E_{ij}\neq E_{ji}. E ij = E j i .
因此 master runtime 可以完整觀察、fork 或 pause 一個 child world,而 child world 的 local agent 不必知道 master 存在,更不需要反向控制 parent runtime。
本文明確承接 Neo.K 既有 WDC/BWC 與 HDUS,保留:
Future Candidate ≠ Runnable World ≠ Real-World Future , \boxed{
\text{Future Candidate}
\neq
\text{Runnable World}
\neq
\text{Real-World Future},
} Future Candidate = Runnable World = Real-World Future ,
Clone ≠ Fork ≠ Replay ≠ Counterfactual Branch , \boxed{
\text{Clone}
\neq
\text{Fork}
\neq
\text{Replay}
\neq
\text{Counterfactual Branch},
} Clone = Fork = Replay = Counterfactual Branch ,
World Generation ≠ World Governance , \boxed{
\text{World Generation}
\neq
\text{World Governance},
} World Generation = World Governance ,
Observation ≠ Authority ≠ Evaluation ≠ Governance , \boxed{
\text{Observation}
\neq
\text{Authority}
\neq
\text{Evaluation}
\neq
\text{Governance},
} Observation = Authority = Evaluation = Governance ,
以及:
World Count ≠ Independent Evidence Count ≠ Truth . \boxed{
\text{World Count}
\neq
\text{Independent Evidence Count}
\neq
\text{Truth}.
} World Count = Independent Evidence Count = Truth .
本文新增的是把上述 world-domain semantics 與「遊戲引擎作為投影計算機」「AI 並行觀察 simulation world family」「非對偶 projection」「非對稱 world graph」整合成一個 AI-native runtime。
最後,本文提出 World Family Scheduler(WFS) 。給定世界族、現實證據、任務、算力與風險:
W F S : ( W t , E t , τ , b , ρ ) → A W , \mathsf{WFS}
:
(\mathfrak W_t,E_t,\tau,b,\rho)
\rightarrow
\mathcal A_{\mathfrak W}, WFS : ( W t , E t , τ , b , ρ ) → A W ,
其中:
A W = { C o n t i n u e , F o r k , M e r g e , P r u n e , P a u s e , R e o b s e r v e , I n c r e a s e R e s o l u t i o n , P r o m o t e C a n d i d a t e } . \mathcal A_{\mathfrak W}
=
\{
\mathsf{Continue},
\mathsf{Fork},
\mathsf{Merge},
\mathsf{Prune},
\mathsf{Pause},
\mathsf{Reobserve},
\mathsf{IncreaseResolution},
\mathsf{PromoteCandidate}
\}. A W = { Continue , Fork , Merge , Prune , Pause , Reobserve , IncreaseResolution , PromoteCandidate } .
AI 的「想像」於是可以從 token 中的假設,真正 materialize 成多個 runnable worlds,再對它們施加不同 observation operators。
本文最終命題:
Future AI perception ≠ observation of one rendered world . \boxed{
\text{Future AI perception}
\neq
\text{observation of one rendered world}.
} Future AI perception = observation of one rendered world .
更一般地:
Future AI perception = structured observation over a family of actual-linked, simulated, counterfactual, replayed, synthetic, and nested worlds . \boxed{
\text{Future AI perception}
=
\text{structured observation over a family of actual-linked,
simulated, counterfactual, replayed, synthetic, and nested worlds}.
} Future AI perception = structured observation over a family of actual-linked, simulated, counterfactual, replayed, synthetic, and nested worlds .
關鍵詞: Simulation World Family Runtime、SWFR、world family、branching simulation、counterfactual world、nested simulation、asymmetric observation、authority matrix、world graph、game engine、world model
1. 遊戲引擎其實已經是一台投影計算機
一個 game engine 維持:
W t → Φ W t + 1 . W_t
\xrightarrow{\Phi}
W_{t+1}. W t Φ W t + 1 .
同一 W t W_t W t 可以有:
O c a m e r a ( W t ) , \mathcal O_{\mathrm{camera}}(W_t), O camera ( W t ) ,
O m a p ( W t ) , \mathcal O_{\mathrm{map}}(W_t), O map ( W t ) ,
O c o l l i s i o n ( W t ) , \mathcal O_{\mathrm{collision}}(W_t), O collision ( W t ) ,
O n a v ( W t ) . \mathcal O_{\mathrm{nav}}(W_t). O nav ( W t ) .
所以:
3D render is only one observation operator . \boxed{
\text{3D render}
\text{ is only one observation operator}.
} 3D render is only one observation operator .
2. Renderer 不等於 World
Renderer ≠ World State ≠ World Dynamics . \boxed{
\text{Renderer}
\neq
\text{World State}
\neq
\text{World Dynamics}.
} Renderer = World State = World Dynamics .
未來引擎的核心不應是「如何畫 3D」,而是:
Simulation Kernel + Projection Operator Family . \boxed{
\text{Simulation Kernel}
+
\text{Projection Operator Family}.
} Simulation Kernel + Projection Operator Family .
3. 同一世界可以有很多合法 View
P ( W ) = { P 3 D , P g r a p h , P c a u s a l , P e c o n o m i c , P r i s k , P s e m a n t i c , P A I } . \mathcal P(W)
=
\{
P_{\mathrm{3D}},
P_{\mathrm{graph}},
P_{\mathrm{causal}},
P_{\mathrm{economic}},
P_{\mathrm{risk}},
P_{\mathrm{semantic}},
P_{\mathrm{AI}}
\}. P ( W ) = { P 3D , P graph , P causal , P economic , P risk , P semantic , P AI } .
對 AI 而言,graph 或 structured state 可能比 3D viewport 更合理。
4. 從一個 World 到 World Family
本文定義:
W = { W 0 , … , W n } . \boxed{
\mathfrak W
=
\{W_0,\ldots,W_n\}.
} W = { W 0 , … , W n } .
每個 W i W_i W i 都是一個具有獨立 runtime semantics 的 world node。
5. World 的最小型別
W i = ( S i , Φ i , T i , A i , E i , G i , P i ) . W_i
=
\left(
S_i,
\Phi_i,
T_i,
\mathcal A_i,
\mathcal E_i,
\mathcal G_i,
\mathcal P_i
\right). W i = ( S i , Φ i , T i , A i , E i , G i , P i ) .
其中分別是 state、transition、local time、agents、evidence/events、governance/authority、projection family。
6. World Mode
定義:
μ ( W i ) ∈ { A c t u a l L i n k e d , D i g i t a l T w i n , S i m u l a t i o n , C o u n t e r f a c t u a l , R e p l a y , S y n t h e t i c , N e s t e d } . \mu(W_i)
\in
\{
\mathsf{ActualLinked},
\mathsf{DigitalTwin},
\mathsf{Simulation},
\mathsf{Counterfactual},
\mathsf{Replay},
\mathsf{Synthetic},
\mathsf{Nested}
\}. μ ( W i ) ∈ { ActualLinked , DigitalTwin , Simulation , Counterfactual , Replay , Synthetic , Nested } .
不同 mode 不得混淆。
7. Actual-Linked World 仍然不是 Reality
W a c t u a l − l i n k e d ≠ R . \boxed{
W_{\mathrm{actual-linked}}
\neq
\mathcal R.
} W actual − linked = R .
它只是與現實持續 evidence-coupled 的 reconstruction/runtime。
8. Simulation 與 Counterfactual
Simulation world 可以改 initial state、policy 或參數。
Counterfactual world 更明確包含:
do ( x = x ′ ) . \operatorname{do}(x=x'). do ( x = x ′ ) .
因此兩者可以重疊,但不必同義。
9. Replay 不是 Counterfactual
Replay 主要重新執行或查看既有歷史:
H 0 : t . H_{0:t}. H 0 : t .
所以承接舊 WDC:
Replay ≠ Counterfactual Branch . \boxed{
\text{Replay}
\neq
\text{Counterfactual Branch}.
} Replay = Counterfactual Branch .
10. Nested World
若 W i W_i W i 中的 agent 建立:
W j , W_j, W j ,
則:
W j ◃ W i . W_j\triangleleft W_i. W j ◃ W i .
但:
runtime nesting ≠ ontological containment . \boxed{
\text{runtime nesting}
\neq
\text{ontological containment}.
} runtime nesting = ontological containment .
11. World Identity 與 Snapshot 分離
承接 HDUS:
W o r l d S n a p s h o t ≠ W o r l d . \boxed{
WorldSnapshot
\neq
World.
} W or l d S na p s h o t = W or l d .
同樣:
W o r l d I d ≠ R e n d e r e d S c e n e I d . \boxed{
WorldId
\neq
RenderedSceneId.
} W or l d I d = R e n d er e d S ce n e I d .
12. World Family Graph
G W = ( V W , E W ) . G_{\mathfrak W}
=
(V_{\mathfrak W},E_{\mathfrak W}). G W = ( V W , E W ) .
節點是 world,邊必須有型別。
13. Typed World Edge
e i j = ( W i , W j , τ e , γ e ) . e_{ij}
=
(W_i,W_j,\tau_e,\gamma_e). e ij = ( W i , W j , τ e , γ e ) .
τ e \tau_e τ e 是 operation type, γ e \gamma_e γ e 是 contract。
14. World Edge Family
τ e ∈ { O b s e r v e , P r o j e c t , F o r k , E m b e d , Q u e r y , C o u p l e , C o n t r o l , M e r g e , P r u n e , P r o m o t e , T r a n s p o r t E v i d e n c e } . \tau_e
\in
\{
\mathsf{Observe},
\mathsf{Project},
\mathsf{Fork},
\mathsf{Embed},
\mathsf{Query},
\mathsf{Couple},
\mathsf{Control},
\mathsf{Merge},
\mathsf{Prune},
\mathsf{Promote},
\mathsf{TransportEvidence}
\}. τ e ∈ { Observe , Project , Fork , Embed , Query , Couple , Control , Merge , Prune , Promote , TransportEvidence } .
15. 非對偶一:Projection 通常不可逆
可能:
W a ≠ W b , W_a\neq W_b, W a = W b ,
但:
Π ( W a ) = Π ( W b ) . \Pi(W_a)=\Pi(W_b). Π ( W a ) = Π ( W b ) .
因此:
Π − 1 need not exist . \boxed{
\Pi^{-1}
\text{ need not exist}.
} Π − 1 need not exist .
從 projection 回推 world 是新的 inverse problem,不是把箭頭反過來。
16. 非對偶二:Observation 不是 World Recovery
Observation ≠ Complete World Reconstruction . \boxed{
\text{Observation}
\neq
\text{Complete World Reconstruction}.
} Observation = Complete World Reconstruction .
AI 看見一個 world slice 不等於擁有完整 world state。
17. 非對偶三:Observe 不是 Control 的逆
O b s e r v e ≠ C o n t r o l − 1 . \boxed{
\mathsf{Observe}
\neq
\mathsf{Control}^{-1}.
} Observe = Control − 1 .
某些 control 甚至是不可逆 action。
18. 非對稱 Observation Matrix
O i j ∈ [ 0 , 1 ] . O_{ij}
\in
[0,1]. O ij ∈ [ 0 , 1 ] .
一般:
O i j ≠ O j i . \boxed{
O_{ij}
\neq
O_{ji}.
} O ij = O j i .
master world 可以看 child,而 child 可以完全不知道 master。
19. Branch Blindness
local agent 不一定能觀察:
sibling worlds;
parent runtime;
evaluator;
governor。
這可以是實驗隔離需求,不是 bug。
20. Authority Matrix
A i j ∈ [ 0 , 1 ] . A_{ij}
\in
[0,1]. A ij ∈ [ 0 , 1 ] .
一般:
A i j ≠ A j i . \boxed{
A_{ij}
\neq
A_{ji}.
} A ij = A j i .
master 可能能 pause、fork、terminate child,但 child 沒有反向能力。
21. Observation 與 Authority 分離
承接 WDC:
Observation ≠ Authority ≠ Evaluation ≠ Governance . \boxed{
\text{Observation}
\neq
\text{Authority}
\neq
\text{Evaluation}
\neq
\text{Governance}.
} Observation = Authority = Evaluation = Governance .
所以:
O i j = 1 O_{ij}=1 O ij = 1
不推出:
A i j > 0. A_{ij}>0. A ij > 0.
22. Causal Coupling Matrix
K i j K_{ij} K ij
表示 W i W_i W i 是否能因果影響 W j W_j W j 。
一般:
K i j ≠ K j i . K_{ij}\neq K_{ji}. K ij = K j i .
這對 real world/simulation world 特別重要。
23. 現實與 Simulation 的典型非對稱
現實可以持續更新 digital twin:
K R , W > 0. K_{\mathcal R,W}>0. K R , W > 0.
但 simulation 不應自動改變現實:
K W , R = 0 K_{W,\mathcal R}=0 K W , R = 0
直到通過 actuation gate。
24. Evidence Transport Matrix
E i j E_{ij} E ij
表示 W i W_i W i 的 result 是否能成為 W j W_j W j 的 admissible evidence。
25. Simulation Result 不等 Real Evidence
World Execution Result ≠ Real-World Evidence . \boxed{
\text{World Execution Result}
\neq
\text{Real-World Evidence}.
} World Execution Result = Real-World Evidence .
必須建立 world-to-reality evidence transport contract。
26. World Count 不等 Evidence Count
World Count ≠ Independent Evidence Count ≠ Truth . \boxed{
\text{World Count}
\neq
\text{Independent Evidence Count}
\neq
\text{Truth}.
} World Count = Independent Evidence Count = Truth .
100 個 shared-backend worlds 可以共享同一 model-form error。
27. Cross-World Agreement 不等 Replication
Cross-World Agreement ≠ Independent Replication ≠ Real-World Validation . \boxed{
\text{Cross-World Agreement}
\neq
\text{Independent Replication}
\neq
\text{Real-World Validation}.
} Cross-World Agreement = Independent Replication = Real-World Validation .
28. AI 可以真的「看多個 Worlds」
令:
W t = { W t ( 1 ) , … , W t ( m ) } . \mathfrak W_t
=
\{W_t^{(1)},\ldots,W_t^{(m)}\}. W t = { W t ( 1 ) , … , W t ( m ) } .
AI 不必把每個 world render 成影片。
29. Per-World Observation Operator
P i = O τ i ( W i ) . P_i
=
\mathcal O_{\tau_i}
(W_i). P i = O τ i ( W i ) .
例如:
O r i s k , O c a u s a l , O e c o n o m i c , O g o a l . \mathcal O_{\mathrm{risk}},
\quad
\mathcal O_{\mathrm{causal}},
\quad
\mathcal O_{\mathrm{economic}},
\quad
\mathcal O_{\mathrm{goal}}. O risk , O causal , O economic , O goal .
30. AI-native World View
P i A I = ( G r a p h i , S t a t e i , V a l u e i , U n c e r t a i n t y i ) . P_i^{AI}
=
\left(
Graph_i,
State_i,
Value_i,
Uncertainty_i
\right). P i A I = ( G r a p h i , S t a t e i , V a l u e i , U n cer t ain t y i ) .
這可以完全沒有 3D 圖像。
31. Future AI Perception
P A I ( t ) = { O i , k ( W i ) } i , k . \boxed{
\mathcal P_{AI}(t)
=
\{
\mathcal O_{i,k}(W_i)
\}_{i,k}.
} P A I ( t ) = { O i , k ( W i ) } i , k .
AI perception 可以是 world × operator 的 observation family。
32. 不要求 Collapse 成單一 World
AI 可以維持:
Q t = { ( W i , q i ) } . \mathcal Q_t
=
\{(W_i,q_i)\}. Q t = {( W i , q i )} .
q i q_i q i 可以是 probability、rank、plausibility 或其他 task-specific weight。
本文不要求 Bayesian-only。
33. World Family Scheduler
W F S : ( W t , E t , τ , b , ρ ) → A W . \boxed{
\mathsf{WFS}
:
(\mathfrak W_t,E_t,\tau,b,\rho)
\rightarrow
\mathcal A_{\mathfrak W}.
} WFS : ( W t , E t , τ , b , ρ ) → A W .
34. Scheduler Action
A W = { C o n t i n u e , F o r k , M e r g e , P r u n e , P a u s e , R e o b s e r v e , I n c r e a s e R e s o l u t i o n , P r o m o t e C a n d i d a t e } . \mathcal A_{\mathfrak W}
=
\{
\mathsf{Continue},
\mathsf{Fork},
\mathsf{Merge},
\mathsf{Prune},
\mathsf{Pause},
\mathsf{Reobserve},
\mathsf{IncreaseResolution},
\mathsf{PromoteCandidate}
\}. A W = { Continue , Fork , Merge , Prune , Pause , Reobserve , IncreaseResolution , PromoteCandidate } .
35. Fork
當 uncertainty 或政策候選需要分開:
W i → W i ( a ) , W i ( b ) . W_i
\rightarrow
W_i^{(a)},W_i^{(b)}. W i → W i ( a ) , W i ( b ) .
如果 branch 可以獨立 authoritative evolution,應取得 distinct world identity。
36. Merge
W A + W B → W C . W_A+W_B
\rightarrow
W_C. W A + W B → W C .
但:
W o r l d M e r g e ≠ W A = W B = W C . \boxed{
WorldMerge
\neq
W_A=W_B=W_C.
} W or l d M er g e = W A = W B = W C .
W C W_C W C 是 merge descendant。
37. Prune
如果某 world:
Q ( W i ) < θ Q Q(W_i)<\theta_Q Q ( W i ) < θ Q
或成本:
C ( W i ) > θ C , C(W_i)>\theta_C, C ( W i ) > θ C ,
scheduler 可以 prune。
但:
P r u n e ≠ impossibility proof . \boxed{
\mathsf{Prune}
\neq
\text{impossibility proof}.
} Prune = impossibility proof .
38. Prune 不等 Delete
P r u n e ≠ D e l e t e . \boxed{
\mathsf{Prune}
\neq
\mathsf{Delete}.
} Prune = Delete .
budget prune 最好可 reopen。
39. Reobserve
對同一 world 改 operator:
O a → O b . \mathcal O_a
\rightarrow
\mathcal O_b. O a → O b .
這正好承接投影系列的 adversarial reprojection。
40. Increase Resolution
不是每個 world 都要最高 fidelity。
可以:
r i ≠ r j . r_i\neq r_j. r i = r j .
因此形成 multi-fidelity world family:
W = { ( W i , r i ) } . \mathfrak W
=
\{(W_i,r_i)\}. W = {( W i , r i )} .
41. Promote Candidate
simulation world 的結論可轉成 action proposal。
但:
PromoteCandidate ≠ CommitToReality . \boxed{
\text{PromoteCandidate}
\neq
\text{CommitToReality}.
} PromoteCandidate = CommitToReality .
42. World Family Utility
概念上:
J ( W i ) = U τ ( W i ) + V i n f o ( W i ) + V c o u n t e r e x a m p l e ( W i ) − C c o m p u t e ( W i ) − D m o d e l ( W i ) . J(W_i)
=
U_{\tau}(W_i)
+
V_{\mathrm{info}}(W_i)
+
V_{\mathrm{counterexample}}(W_i)
-
C_{\mathrm{compute}}(W_i)
-
D_{\mathrm{model}}(W_i). J ( W i ) = U τ ( W i ) + V info ( W i ) + V counterexample ( W i ) − C compute ( W i ) − D model ( W i ) .
43. 世界族不是越大越好
∣ W ∣ ↑ |\mathfrak W|\uparrow ∣ W ∣ ↑
會增加 coverage,也會增加:
compute;
memory;
coordination;
common-mode error;
false consensus。
44. World Family Budget
B W = ( B c o m p u t e , B m e m o r y , B l a t e n c y ) . B_{\mathfrak W}
=
(B_{\mathrm{compute}},B_{\mathrm{memory}},B_{\mathrm{latency}}). B W = ( B compute , B memory , B latency ) .
scheduler 分配:
∑ i b i ≤ B W . \sum_i b_i
\leq
B_{\mathfrak W}. i ∑ b i ≤ B W .
45. AI 的想像可以 Materialize 成 Runnable World
candidate thought → runnable world . \boxed{
\text{candidate thought}
\rightarrow
\text{runnable world}.
} candidate thought → runnable world .
這比純 token hypothesis 多出:
pause;
inspect;
query;
replay;
instrument;
fork。
46. Debug Observer 與 Local Agent 必須分開
外部 observer 可以看 hidden state。
但:
World-External Observer Access ≠ World-Local Agent Access . \boxed{
\text{World-External Observer Access}
\neq
\text{World-Local Agent Access}.
} World-External Observer Access = World-Local Agent Access .
47. 三種 Knowledge Domain
K i l o c a l ≠ K i g l o b a l ≠ K c r o s s . K_i^{local}
\neq
K_i^{global}
\neq
K^{cross}. K i l oc a l = K i g l o ba l = K cr oss .
分別是 local-agent knowledge、world-global knowledge、cross-world knowledge。
48. Master、Local Agent、Observer、Evaluator、Governor 分離
A m a s t e r ≠ A i l o c a l ≠ O i ≠ E i ≠ G . A^{master}
\neq
A_i^{local}
\neq
O_i
\neq
E_i
\neq
G. A ma s t er = A i l oc a l = O i = E i = G .
這是世界族治理的基本非坍縮。
49. Higher-Level 不自動擁有所有 Authority
execution nesting:
W j ◃ W i W_j\triangleleft W_i W j ◃ W i
不推出 higher level 在所有 domain 都天然全權。
authority 應為:
A i j ( d ) . A_{ij}^{(d)}. A ij ( d ) .
50. World Family Governor
G o v W \mathsf{Gov}_{\mathfrak W} Gov W
管理:
WorldId;
lineage;
lifecycle;
quota;
authority;
evidence transport;
merge/prune/promotion gate。
51. Governor 不等 Omniscient Observer
Governance Authority ≠ Complete Epistemic Access . \boxed{
\text{Governance Authority}
\neq
\text{Complete Epistemic Access}.
} Governance Authority = Complete Epistemic Access .
52. Nested Simulation Depth 必須有界
若:
W 0 ▹ W 1 ▹ ⋯ ▹ W d , W_0
\triangleright
W_1
\triangleright
\cdots
\triangleright
W_d, W 0 ▹ W 1 ▹ ⋯ ▹ W d ,
需有:
d ≤ d max . d\leq d_{\max}. d ≤ d m a x .
由 compute、latency、safety 與 governance 決定。
53. World Graph 不一定是 Tree
fork lineage 比較像 tree。
但 merge、observe、evidence transport、control 會形成一般 directed graph。
所以應分開:
G l i n e a g e G_{\mathrm{lineage}} G lineage
與:
G i n t e r a c t i o n . G_{\mathrm{interaction}}. G interaction .
54. Ancestor 不等 CanObserve
A n c e s t o r ( W i , W j ) Ancestor(W_i,W_j) A n ces t or ( W i , W j )
不表示:
C a n O b s e r v e ( W i , W j ) . CanObserve(W_i,W_j). C an O b ser v e ( W i , W j ) .
lineage edge 與 interaction edge 不可混用。
55. World Time 可以不同
T i ≠ T j . T_i\neq T_j. T i = T j .
因此 cross-world causality 不等 shared clock:
C r o s s W o r l d C a u s a l i t y ≠ S h a r e d W a l l C l o c k . \boxed{
CrossWorldCausality
\neq
SharedWallClock.
} C r oss W or l d C a u s a l i t y = S ha r e d W a l l C l oc k .
56. Time Mapping
需要:
T i → j . T_{i\rightarrow j}. T i → j .
但時間換算不改 World identity。
57. World Family Provenance
P r o v ( W i ) = ( W o r l d I d , A n c e s t o r , M o d e , M o d e l , S e e d , P o l i c y , E v i d e n c e , E n g i n e , T i m e , A u t h o r i t y ) . Prov(W_i)
=
\left(
WorldId,
Ancestor,
Mode,
Model,
Seed,
Policy,
Evidence,
Engine,
Time,
Authority
\right). P r o v ( W i ) = ( W or l d I d , A n ces t or , M o d e , M o d e l , S ee d , P o l i cy , E v i d e n ce , E n g in e , T im e , A u t h or i t y ) .
58. Fork Certificate
F o r k C e r t = ( A n c e s t o r , F o r k P o i n t , C h a n g e d V a r i a b l e s , A u t h o r i t y , P u r p o s e , M o d e ) . \mathsf{ForkCert}
=
\left(
Ancestor,
ForkPoint,
ChangedVariables,
Authority,
Purpose,
Mode
\right). ForkCert = ( A n ces t or , F or k P o in t , C han g e d V a r iab l es , A u t h or i t y , P u r p ose , M o d e ) .
59. Merge Certificate
M e r g e C e r t = ( P a r e n t s , C o n f l i c t S e t , M e r g e R u l e , P r e s e r v e d I n v a r i a n t s , A u t h o r i t y , D e b t ) . \mathsf{MergeCert}
=
\left(
Parents,
ConflictSet,
MergeRule,
PreservedInvariants,
Authority,
Debt
\right). MergeCert = ( P a r e n t s , C o n f l i c tS e t , M er g e R u l e , P r eser v e d I n v a r ian t s , A u t h or i t y , D e b t ) .
60. Prune Certificate
P r u n e C e r t = ( W o r l d I d , R e a s o n , E v i d e n c e , B u d g e t , R e o p e n P o l i c y ) . \mathsf{PruneCert}
=
\left(
WorldId,
Reason,
Evidence,
Budget,
ReopenPolicy
\right). PruneCert = ( W or l d I d , R e a so n , E v i d e n ce , B u d g e t , R eo p e n P o l i cy ) .
61. Cross-World Communication 必須 Explicit
定義 channel:
C i j w o r l d . C_{ij}^{world}. C ij w or l d .
可限制 bandwidth、event types、authority、latency、provenance。
因此:
world adjacency ⇏ ambient authority . \boxed{
\text{world adjacency}
\not\Rightarrow
\text{ambient authority}.
} world adjacency ⇒ ambient authority .
62. Simulation Sandbox
高風險 action 可以先在:
W s a n d b o x W_{\mathrm{sandbox}} W sandbox
中執行。
但:
Sandbox Success ≠ Real Success . \boxed{
\text{Sandbox Success}
\neq
\text{Real Success}.
} Sandbox Success = Real Success .
存在:
D s i m 2 r e a l . D_{\mathrm{sim2real}}. D sim2real .
63. Reality Re-Entry
真正 action 後,新的 real evidence:
E r e a l , t + 1 E_{real,t+1} E r e a l , t + 1
必須回流:
W t + 1 = U ( W t , E r e a l , t + 1 ) . \mathfrak W_{t+1}
=
\mathcal U(\mathfrak W_t,E_{real,t+1}). W t + 1 = U ( W t , E r e a l , t + 1 ) .
64. Expected Outcome 不等 Observed Outcome
承接 PNCW:
E x p e c t e d O u t c o m e ≠ O b s e r v e d O u t c o m e . \boxed{
ExpectedOutcome
\neq
ObservedOutcome.
} E x p ec t e d O u t co m e = O b ser v e d O u t co m e .
所以任何 promoted action 後都要重新 observe。
65. SWFR 與 Tree Search 不同
Tree search 主要是 state-action tree。
SWFR node 是具有 identity、time、authority、evidence、lineage、projection contract 的 runnable world。
所以:
S W F R ≠ T r e e S e a r c h . \boxed{
\mathsf{SWFR}
\neq
\mathsf{TreeSearch}.
} SWFR = TreeSearch .
66. SWFR 與 Model-Based RL
model-based RL 提供 learned dynamics 與 planning。
SWFR 新增治理焦點:
world identity;
lifecycle;
cross-world observation;
evidence;
authority;
projection semantics。
67. SWFR 與 Digital Twin
digital twin 常是:
R ↔ W D T . \mathcal R
\leftrightarrow
W_{\mathrm{DT}}. R ↔ W DT .
SWFR 再生成:
W D T → { W s c e n a r i o ( i ) } . W_{\mathrm{DT}}
\rightarrow
\{W_{\mathrm{scenario}}^{(i)}\}. W DT → { W scenario ( i ) } .
形成 scenario family。
68. SWFR 與 Generative Video World Model
video world model 可是 backend。
但:
Generated Video ≠ Complete Runnable World Semantics . \boxed{
\text{Generated Video}
\neq
\text{Complete Runnable World Semantics}.
} Generated Video = Complete Runnable World Semantics .
SWFR 不把 world 強制降成 pixels。
69. Symbolic World 也可以 Runnable
例如:
PDDL;
executable simulator;
causal state machine;
graph transition system。
所以 world backend 可以非視覺。
70. Hybrid World Backend
Φ i = Φ p h y s i c s + Φ L L M + Φ a g e n t + Φ d a t a . \Phi_i
=
\Phi_{\mathrm{physics}}
+
\Phi_{\mathrm{LLM}}
+
\Phi_{\mathrm{agent}}
+
\Phi_{\mathrm{data}}. Φ i = Φ physics + Φ LLM + Φ agent + Φ data .
不同 backend 衝突時需要 arbitration,而不是自動平均。
71. Engine Independence
Unity、Unreal、Godot 都可成為:
B a c k e n d ( W i ) . Backend(W_i). B a c k e n d ( W i ) .
SWFR 不綁任一 engine。
72. Headless World 是一等公民
R e n d e r e r = N u l l Renderer=Null R e n d er er = N u l l
世界仍然 runnable。
這對 AI 同時運行大量 branches 很重要。
73. Game Engine 的下一代拆法
Simulation Kernel + World Identity + Branching Runtime + Observation Operator Registry + Authority Graph + Evidence Graph + World Family Scheduler . \boxed{
\text{Simulation Kernel}
+
\text{World Identity}
+
\text{Branching Runtime}
+
\text{Observation Operator Registry}
+
\text{Authority Graph}
+
\text{Evidence Graph}
+
\text{World Family Scheduler}.
} Simulation Kernel + World Identity + Branching Runtime + Observation Operator Registry + Authority Graph + Evidence Graph + World Family Scheduler .
74. Observation Operator Plugin
可以定義:
O b s P l u g i n = ⟨ I n p u t W o r l d S c h e m a , O u t p u t S c h e m a , S c o p e , C o s t , D e b t , P e r m i s s i o n ⟩ . \mathsf{ObsPlugin}
=
\left\langle
InputWorldSchema,
OutputSchema,
Scope,
Cost,
Debt,
Permission
\right\rangle. ObsPlugin = ⟨ I n p u t W or l d S c h e ma , O u tp u tS c h e ma , S co p e , C os t , D e b t , P er mi ss i o n ⟩ .
75. SWFR 最小 Runtime
S W F R = ( W , G l i n e a g e , G i n t e r a c t i o n , O , A , K , E , W F S , G o v W , P r o j ) . \boxed{
\mathsf{SWFR}
=
\left(
\mathfrak W,
G_{\mathrm{lineage}},
G_{\mathrm{interaction}},
\mathbf O,
\mathbf A,
\mathbf K,
\mathbf E,
\mathsf{WFS},
\mathsf{Gov}_{\mathfrak W},
\mathsf{Proj}
\right).
} SWFR = ( W , G lineage , G interaction , O , A , K , E , WFS , Gov W , Proj ) .
76. SWFR 最小 Loop
E t → W t → O b s e r v e → E v a l u a t e → S c h e d u l e → W t + 1 . \boxed{
E_t
\rightarrow
\mathfrak W_t
\rightarrow
\mathsf{Observe}
\rightarrow
\mathsf{Evaluate}
\rightarrow
\mathsf{Schedule}
\rightarrow
\mathfrak W_{t+1}.
} E t → W t → Observe → Evaluate → Schedule → W t + 1 .
如果涉及現實 action,再增加:
P r o p o s a l → A u t h o r i t y → A c t u a t i o n → E r e a l , t + 1 . \mathsf{Proposal}
\rightarrow
\mathsf{Authority}
\rightarrow
\mathsf{Actuation}
\rightarrow
E_{real,t+1}. Proposal → Authority → Actuation → E r e a l , t + 1 .
77. SWFR 與 PPOE 接合
Paper 01 回答:
W → P → Φ → O b s e r v e r . W
\rightarrow
P
\rightarrow
\Phi
\rightarrow
Observer. W → P → Φ → O b ser v er .
Paper 02 回答:
W → { P i } → O b s e r v e r / A I . \mathfrak W
\rightarrow
\{P_i\}
\rightarrow
Observer/AI. W → { P i } → O b ser v er / A I .
所以:
W → observation selection P ∗ → carrier P ^ . \boxed{
\mathfrak W
\xrightarrow{\text{observation selection}}
P^\ast
\xrightarrow{\text{carrier}}
\widehat P.
} W observation selection P ∗ carrier P .
78. AI 不需要物理 Materialize 所有 Worlds
對 AI-native simulation:
P i A I P_i^{AI} P i A I
可以只存在於計算載體。
只有人類 inspection 時才:
W i → Π H P i H . W_i
\xrightarrow{\Pi_H}
P_i^H. W i Π H P i H .
79. World Mode 必須始終可見
每個結果必須知道它是:
O b s e r v e d , P r e d i c t e d , S i m u l a t e d , C o u n t e r f a c t u a l , R e p l a y , S y n t h e t i c . Observed,
Predicted,
Simulated,
Counterfactual,
Replay,
Synthetic. O b ser v e d , P r e d i c t e d , S im u l a t e d , C o u n t er f a c t u a l , R e pl a y , S y n t h e t i c .
否則會發生 world-modality confusion。
80. AI 也需要 World Source Monitoring
counterfactual event 不能混進 real evidence store。
這就是 B04 source monitoring 的 world-runtime 版本。
81. World Audit Packet
W A u d i t = ⟨ W o r l d I d , M o d e , A n c e s t o r , M o d e l , E v i d e n c e , O b s e r v a t i o n , A u t h o r i t y , T i m e , D e b t , V a l i d a t i o n ⟩ . \mathsf{WAudit}
=
\left\langle
WorldId,
Mode,
Ancestor,
Model,
Evidence,
Observation,
Authority,
Time,
Debt,
Validation
\right\rangle. WAudit = ⟨ W or l d I d , M o d e , A n ces t or , M o d e l , E v i d e n ce , O b ser v a t i o n , A u t h or i t y , T im e , D e b t , V a l i d a t i o n ⟩ .
82. 高風險第一原則
more simulated worlds ≠ more truth . \boxed{
\text{more simulated worlds}
\neq
\text{more truth}.
} more simulated worlds = more truth .
83. 第二原則
better-looking world ≠ better-calibrated world . \boxed{
\text{better-looking world}
\neq
\text{better-calibrated world}.
} better-looking world = better-calibrated world .
84. 第三原則
world-local success ≠ real-world action authority . \boxed{
\text{world-local success}
\neq
\text{real-world action authority}.
} world-local success = real-world action authority .
85. 第四原則
cross-world access ≠ cross-world control . \boxed{
\text{cross-world access}
\neq
\text{cross-world control}.
} cross-world access = cross-world control .
86. 第五原則
nested level ≠ epistemic superiority . \boxed{
\text{nested level}
\neq
\text{epistemic superiority}.
} nested level = epistemic superiority .
87. Counterexample World
世界族不能只做 majority vote。
一個揭露 catastrophic failure 的 world 可能比大量「成功」world 更有決策價值。
可定義:
V c e ( W i ) . V_{\mathrm{ce}}(W_i). V ce ( W i ) .
88. Diversity 不是 Independence
世界族可有:
D W = ( D p a r a m , D m o d e l , D p o l i c y , D s e e d , D e v i d e n c e ) . \mathbf D_{\mathfrak W}
=
(
D_{\mathrm{param}},
D_{\mathrm{model}},
D_{\mathrm{policy}},
D_{\mathrm{seed}},
D_{\mathrm{evidence}}
). D W = ( D param , D model , D policy , D seed , D evidence ) .
不同 seed 不等於不同 model。
89. Common-Mode Error
如果所有 worlds 使用同 backend:
M b , M_b, M b ,
就可能共享:
D m o d e l . D_{\mathrm{model}}. D model .
所以真正 audit 應看:
N i n d e p e n d e n t a s s u m p t i o n s N_{\mathrm{independent\ assumptions}} N independent assumptions
而不是只看 N w o r l d N_{\mathrm{world}} N world 。
90. 現有研究接口
2025 的 SimuRA 已讓 agent 利用 world model 模擬候選 action 後果再規劃;model-based web-agent 研究也明確提出在不可逆 web action 前先用 world model deliberation。
2025 的 FLIP、AdaWorld 等工作把 learned world models 用於 long-horizon interactive planning。
2025 Agent2World 類研究則生成 symbolic/executable world models,再透過 simulation-based testing 驗證。
2026 Gamma-World 已開始處理多個 independently controllable agents 與 shared interactive world 的 real-time rollout。
2025–2026 digital-twin + agent-based simulation 工作則已實作 real-time state ingestion、scenario instantiation、alternative configuration comparison 與 decision support。
這些工作提供了 SWFR 的現實技術接口,但本文新增的是 world-family identity、typed relation、asymmetry、non-duality、authority、evidence transport 與 scheduler 的統一。
91. MVP
第一版甚至不必自己寫 game engine。
可以使用 Godot、Unity 或 Unreal 任一現成 backend,加上:
world_registry
world_heads
snapshot_store
lineage_edges
interaction_edges
observer_plugins
authority_matrix
evidence_matrix
world_family_scheduler
world_audit
92. MVP 世界族
建立:
W 0 → W 1 , W 2 , W 3 , W 4 . W_0
\rightarrow
W_1,W_2,W_3,W_4. W 0 → W 1 , W 2 , W 3 , W 4 .
四個 branches 分別:
P o l i c y A , P o l i c y B , P o l i c y C , N o A c t i o n . Policy_A,
Policy_B,
Policy_C,
NoAction. P o l i c y A , P o l i c y B , P o l i c y C , N o A c t i o n .
全部 headless run。
93. AI 不看影片
AI 可以直接讀:
O r i s k ( W i ) , \mathcal O_{\mathrm{risk}}(W_i), O risk ( W i ) ,
O c o s t ( W i ) , \mathcal O_{\mathrm{cost}}(W_i), O cost ( W i ) ,
O c a u s a l ( W i ) . \mathcal O_{\mathrm{causal}}(W_i). O causal ( W i ) .
人需要時才 render 3D。
94. Nested MVP
讓 W 1 W_1 W 1 中 local agent 建立:
W 1 , 1 . W_{1,1}. W 1 , 1 .
測試:
parent observation;
child blindness;
authority firewall;
cross-world provenance。
95. Evidence MVP
simulation result:
e i s i m e_i^{sim} e i s im
只能進:
S i m E v i d e n c e S t o r e , SimEvidenceStore, S im E v i d e n ce S t or e ,
不能自動進:
R e a l E v i d e n c e S t o r e . RealEvidenceStore. R e a l E v i d e n ce S t or e .
96. Scheduler MVP
先用 heuristic:
S c o r e i = α Q i + β V i + γ V c e , i − δ C i . Score_i
=
\alpha Q_i
+
\beta V_i
+
\gamma V_{\mathrm{ce},i}
-
\delta C_i. S cor e i = α Q i + β V i + γ V ce , i − δ C i .
後續再考慮 learned scheduler。
97. Scheduler 也不能 Ambient Authority
WFS 可以建議 prune、pause、fork。
高風險 destructive operation 仍應經 governor/authority gate。
98. 可實驗研究
本文提出至少八類實驗:
single-world vs world-family planning;
asymmetric observation;
authority separation;
shared-backend evidence illusion;
catastrophic counterexample world;
adaptive pruning;
multi-fidelity world family;
nested simulation depth。
99. 本文不主張什麼
本文不主張 simulated worlds 是平行宇宙,也不賦予它們自動本體地位。
本文不主張 world 數越多越接近真理,不主張 simulation consensus 等於 reality validation,也不主張 higher-level agent 自動擁有 lower-level 全權。
本文不把 SWFR 等同 tree search、model-based RL、digital twin、game engine、video world model 或 LLM world model。
它是一個統合 runtime 抽象。
100. 核心命題總結
Game Engine ⊃ World Execution + Projection Operators . \boxed{
\text{Game Engine}
\supset
\text{World Execution}
+
\text{Projection Operators}.
} Game Engine ⊃ World Execution + Projection Operators .
W = { W 0 , W 1 , … , W n } . \boxed{
\mathfrak W
=
\{W_0,W_1,\ldots,W_n\}.
} W = { W 0 , W 1 , … , W n } .
G W = ( V W , E W ) . \boxed{
G_{\mathfrak W}
=
(V_{\mathfrak W},E_{\mathfrak W}).
} G W = ( V W , E W ) .
Π − 1 need not exist . \boxed{
\Pi^{-1}
\text{ need not exist}.
} Π − 1 need not exist .
O b s e r v e ≠ C o n t r o l − 1 . \boxed{
\mathsf{Observe}
\neq
\mathsf{Control}^{-1}.
} Observe = Control − 1 .
O i j ≠ O j i , A i j ≠ A j i . \boxed{
O_{ij}\neq O_{ji},
\qquad
A_{ij}\neq A_{ji}.
} O ij = O j i , A ij = A j i .
K i j ≠ K j i , E i j ≠ E j i . \boxed{
K_{ij}\neq K_{ji},
\qquad
E_{ij}\neq E_{ji}.
} K ij = K j i , E ij = E j i .
World Count ≠ Independent Evidence Count ≠ Truth . \boxed{
\text{World Count}
\neq
\text{Independent Evidence Count}
\neq
\text{Truth}.
} World Count = Independent Evidence Count = Truth .
P A I ( t ) = { O i , k ( W i ) } i , k . \boxed{
\mathcal P_{AI}(t)
=
\{
\mathcal O_{i,k}(W_i)
\}_{i,k}.
} P A I ( t ) = { O i , k ( W i ) } i , k .
101. 結論
人類通常是在一個世界裡想像很多可能。
傳統 AI 也常把「如果怎樣」寫成一段 token sequence。
未來 AI 可以再往前一步:
hypothesis → runnable world . \boxed{
\text{hypothesis}
\rightarrow
\text{runnable world}.
} hypothesis → runnable world .
它可以真正建立:
W 1 , W 2 , … , W n , W_1,W_2,\ldots,W_n, W 1 , W 2 , … , W n ,
讓每個 world 擁有自己的 state、clock、agents、causal evolution、projection operators、authority boundary、evidence semantics 與 lineage。
AI 的 perception 也不必是一個 viewport。
它可以同時問:
O r i s k , O c a u s a l , O g o a l , O c o u n t e r e x a m p l e \mathcal O_{\mathrm{risk}},
\quad
\mathcal O_{\mathrm{causal}},
\quad
\mathcal O_{\mathrm{goal}},
\quad
\mathcal O_{\mathrm{counterexample}} O risk , O causal , O goal , O counterexample
作用在整個:
W t . \mathfrak W_t. W t .
因此真正 AI-native 的「多重世界觀察」不是科幻式觀看平行宇宙。
它是一個很明確的計算機架構:
one AI → many runnable worlds → many typed observation operators . \boxed{
\text{one AI}
\rightarrow
\text{many runnable worlds}
\rightarrow
\text{many typed observation operators}.
} one AI → many runnable worlds → many typed observation operators .
而且 world relations 天然不必對偶、不必對稱。
AI 能觀察 child world,不代表 child 能觀察 AI。
AI 能 pause simulation,不代表 simulation agent 能 pause parent。
兩個不同 worlds 甚至可以在同一 projection 下完全相同,因此 projection 不能唯一恢復 world。
這就是:
non-dual + asymmetric + typed world-family computation . \boxed{
\text{non-dual}
+
\text{asymmetric}
+
\text{typed world-family computation}.
} non-dual + asymmetric + typed world-family computation .
遊戲引擎提供了工程起點。
下一代則應從:
「一個虛擬世界給玩家玩」
升級成:
「一個 AI 可以生成、運行、觀察、嵌套、比較、修剪與治理一整個模擬世界族。」
102. 兩篇工程延伸總接口
Paper 01:
M → P ∗ → u → Φ → P ^ . \boxed{
M
\rightarrow
P^\ast
\rightarrow
u
\rightarrow
\Phi
\rightarrow
\widehat P.
} M → P ∗ → u → Φ → P .
Paper 02:
W = { W 0 , … , W n } , P A I = { O i , k ( W i ) } . \boxed{
\mathfrak W
=
\{W_0,\ldots,W_n\},
\qquad
\mathcal P_{AI}
=
\{\mathcal O_{i,k}(W_i)\}.
} W = { W 0 , … , W n } , P A I = { O i , k ( W i )} .
兩篇合起來:
World Family → Observation Operator → Projection Compiler → Carrier → Observer . \boxed{
\text{World Family}
\rightarrow
\text{Observation Operator}
\rightarrow
\text{Projection Compiler}
\rightarrow
\text{Carrier}
\rightarrow
\text{Observer}.
} World Family → Observation Operator → Projection Compiler → Carrier → Observer .
參考文獻與內部依賴
Neo.K (2026). 《PPOE Paper 01:物理投影—觀察工程》.
Neo.K (2026). WDC-01 / Branching World Computation foundations .
Neo.K (2026). WDC-02: Clone, Fork, Replay, and Counterfactual Branch semantics .
Neo.K (2026). WDC-03: World Generation and Governance separation .
Neo.K (2026). WDC-04: Nested Agents and Observer Separation .
Neo.K (2026). WDC-05: Cross-World Evidence .
Neo.K (2026). HDUS Multi-World Recursive World Runtime .
Neo.K (2026). PNCW Paper 08: Projection-Native Perception–Action Loop .
Neo.K (2026). HDUS Architecture Constitution v0.1 .
Deng, M., Hou, J., Shen, Y., et al. (2025). SimuRA: Towards General Goal-Oriented Agent via Simulative Reasoning Architecture with LLM-Based World Model . arXiv:2507.23773.
Gu, Y., Zhang, K., Ning, Y., et al. (2025). Is Your LLM Secretly a World Model of the Internet? Model-Based Planning for Web Agents . Transactions on Machine Learning Research.
Gao, C., Zhang, H., Xu, Z., et al. (2025). FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model . ICLR 2025.
Gao, S., Zhou, S., Du, Y., Zhang, J., & Gan, C. (2025). AdaWorld: Learning Adaptable World Models with Latent Actions . ICML 2025.
Hu, M., Xia, B., Wu, Y., et al. (2025). Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback . arXiv:2512.22336.
Liu, F., He, K., Shen, T., et al. (2026). Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players . arXiv:2605.28816.
Yeung, T., Martinez Ribon, J. G. T., Schlenger, J., et al. (2025). Integrating digital twin and agent-based simulation to support adaptive production system design in building projects . Automation in Construction, 180, 106550.
Paper 02 狀態:COMPLETE v0.1 兩篇工程延伸狀態:COMPLETE Canonical source:UTF-8 Markdown;數學 delimiter 僅使用 $...$ 與 $$...$$。