# 三生世界域計算：從認知未來到可運行世界再回到歷史

**Tri-Temporal World-Domain Computation: From Cognitive Futures to Runnable Worlds and Back into History**

**Branching World Computation / World-Domain Cognitive Runtime**  
**分支世界計算／世界域認知 Runtime 系列**  
**WDC-08 / BWC-08 — Unified Synthesis I**

作者：Neo.K（許筌崴）  
協作形式化：Aletheia  
機構：一言諾科技有限公司（EveMissLab）  
日期：2026-08-17  
版本：v0.1  
狀態：TCD–WDC unified runtime synthesis / executable prospective cognition

---

## Canonical Non-Identity Statement

TCD v0.1 Core 已建立：

$$
\boxed{
\mathfrak T_t^{(3)}
=
(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+
)
}
$$

以及：

$$
\boxed{
\mathscr S_t:
\mathfrak T_t^{(3)}
\rightarrow
\mathfrak T_{t+1}^{(3)}.
}
$$

WDC-01 至 WDC-07 則建立：

$$
\boxed{
\text{Future Candidate}
\rightarrow
\text{Runnable Worlds}
\rightarrow
\text{Cross-World Evidence}
\rightarrow
\text{Learning}.
}
$$

本文第一次統一：

$$
\boxed{
\mathfrak T_t^{(3)}
\rightarrow
\mathcal W_t
\rightarrow
\mathfrak E_t^W
\rightarrow
\mathfrak T_{t+1}^{(3)}.
}
$$

但永久保留：

$$
\boxed{
\text{Simulated World Event}
\neq
\text{Parent-World Historical Event}.
}
$$

以及：

$$
\boxed{
\text{Computed Future}
\neq
\text{Actual Future}.
}
$$

以及：

$$
\boxed{
\text{World Evidence}
\neq
\text{Reality Evidence}
}
$$

除非存在明示、可辯護且可審計的 evidence-transport contract。

本文不主張：

- WDC worlds 是物理平行宇宙；
- TCD Future Base Space 是客觀全部未來；
- world simulation 可以物理改寫過去；
- simulation outcome 應直接寫成 real historical fact；
- world ensemble agreement 等於 real-world truth；
- WDC runtime 可以在無 external evidence 下封閉地學會現實；
- 所有 agent 都需要 runnable-world cognition；
- 更多 world computation 必然提高 intelligence；
- full WDC runtime 一定優於簡單 planner；
- TCD、WDC、UCPNP 是已被實證證明的普遍認知定律；
- 本文已完成實際 production implementation；
- 本文對 classical $P$ vs. $NP$ 提供任何新證明。

---

# 摘要

Tri-Temporal Cognitive Dynamics（TCD）將 agent 的時間認知拆成：

$$
\boxed{
\mathcal B_t^-:
\text{Past Choice Base Space},
}
$$

$$
\boxed{
\mathcal B_t^0:
\text{Present Actionable / Reachable Base Space},
}
$$

$$
\boxed{
\mathcal B_t^+:
\text{Generated Future Base Space}.
}
$$

並建立六向 typed couplings：

$$
\boxed{
\begin{array}{lll}
P\to N &: & Historical\ Conditioning,\\
N\to F &: & Future\ Generation,\\
F\to N &: & Prospective\ Attraction,\\
N\to P &: & Historical\ Sedimentation,\\
P\to F &: & Historical\ Projection,\\
F\to P &: & Retrospective\ Relevance.
\end{array}
}
$$

然而 TCD 的：

$$
\mathcal B_t^+
$$

主要仍是一個 prospective cognitive domain。

WDC 的核心突破，是允許其中某些 future candidates：

$$
f_i
\in
\Omega_t^F
$$

被實例化成：

$$
\boxed{
f_i
\xrightarrow{\mathsf I_W}
W_i,
}
$$

即具有：

- identity；
- state；
- dynamics；
- actors；
- rules；
- local history；
- interventions；
- observation；
- evaluation；
- resource budget；
- provenance；

的 bounded runnable worlds。

WDC-02 至 WDC-07 再分別建立：

- branching lineage；
- World-Domain Governor；
- nested agents / observer separation；
- cross-world evidence；
- world computation portfolio；
- world ensemble learning。

本文的問題因此是：

> **這整套 world-domain computation，究竟如何成為 TCD temporal shift 的一部分，而不是掛在 Future Base Space 外面的一個獨立 simulator？**

本文提出一個雙層時間架構：

$$
\boxed{
\text{Inner Computational Prospection}
}
$$

與：

$$
\boxed{
\text{Outer Historical Transition}.
}
$$

在 parent historical time：

$$
t
$$

不變時，agent 可以反覆：

$$
\boxed{
\mathcal B_t^{+,(k)}
\rightarrow
\mathcal W_t^{(k)}
\rightarrow
\mathfrak E_t^{W,(k)}
\rightarrow
\mathcal B_t^{+,(k+1)}
}
$$

並同時更新：

- retrospective relevance；
- current valuation；
- current policy；
- world portfolio。

這是：

# **World-Domain Deliberation Loop**

其中：

$$
k
$$

只是 parent agent 在同一 historical decision moment 裡的 bounded deliberation index。

每個 runnable world：

$$
W_i
$$

又擁有自己的 local runtime time：

$$
\tau_i.
$$

所以本文正式區分：

$$
\boxed{
t
=
\text{parent historical time},
}
$$

$$
\boxed{
k
=
\text{parent deliberation iteration},
}
$$

$$
\boxed{
\tau_i
=
\text{world-local runtime time}.
}
$$

三者不可混淆。

世界可以在：

$$
t
$$

不變時跑：

$$
\tau_i=0,\ldots,10^6.
$$

這不代表 parent reality 經歷了 $10^6$ 個歷史步。

真正：

$$
t
\rightarrow
t+1
$$

只在 parent agent：

1. commit 真正 action；
2. environment / external system transition；
3. observation / consequence 到來；
4. Historical Sedimentation；

後發生。

本文因此定義完整 **Tri-Temporal World-Domain Runtime State**：

$$
\boxed{
\mathfrak R_t^{TW}
=
(
\mathfrak T_t^{(3)},
\mathfrak W_t,
\mathfrak G_t,
\mathfrak E_t^W,
\mathfrak L_t^W,
\Xi_t,
\mathbf B_t,
\kappa_t
).
}
$$

其中：

- $\mathfrak T_t^{(3)}$：TCD triple；
- $\mathfrak W_t$：active / queued / archived world-domain state；
- $\mathfrak G_t$：World-Domain Governor state；
- $\mathfrak E_t^W$：cross-world evidence state；
- $\mathfrak L_t^W$：world ensemble learning state；
- $\Xi_t$：external observations / exogenous inputs；
- $\mathbf B_t$：global resource ledger；
- $\kappa_t$：mission / safety / evidence / authority contracts。

本文定義 parent-level runtime update：

$$
\boxed{
\mathscr R_t^{TW}:
\mathfrak R_t^{TW}
\rightarrow
\mathfrak R_{t+1}^{TW}.
}
$$

但這不是單一 homogeneous function，而是一個 staged, typed, bounded operator composition。

本文提出第一版 canonical staged loop：

$$
\boxed{
\begin{aligned}
\widetilde{\mathcal B}_t^0
&=
\Phi_{-0}
(
\mathcal B_t^-,
\Xi_t
),
\\
\mathcal B_t^{+,(0)}
&=
\Gamma_F
(
\mathcal B_t^-,
\widetilde{\mathcal B}_t^0,
\Xi_t
),
\\
\mathcal C_t^{W,(k)}
&=
\Pi_C
(
\mathcal B_t^{+,(k)},
\mathfrak E_t^{W,(k)},
\mathfrak G_t^{(k)}
),
\\
\mathcal W_t^{(k)}
&=
\mathsf{InstantiateRun}
(
\mathcal C_t^{W,(k)}
),
\\
\mathfrak E_t^{W,(k+1)}
&=
\mathsf{CrossWorldEvidence}
(
\mathcal W_t^{(k)}
),
\\
(
\mathcal B_t^{+,(k+1)},
\mathbf w_t^{-,(k+1)},
\pi_t^{(k+1)}
)
&=
\mathsf{Assimilate}_{W\to T}
(
\mathfrak E_t^{W,(k+1)},
\mathfrak T_t^{(3),(k)}
),
\\
a_t
&\sim
\pi_t^{(K)},
\\
Z_{t+1}^{real}
&=
T_R(
Z_t^{real},
a_t,
\varepsilon_t
),
\\
\mathcal B_{t+1}^-
&=
\mathsf{Sediment}
(
\mathcal B_t^-,
a_t,
Z_{t+1}^{real},
\mathcal P_t^{WDC}
),
\\
\mathcal B_{t+1}^0
&=
\Phi_{-0}
(
\mathcal B_{t+1}^-,
\Xi_{t+1}
),
\\
\mathcal B_{t+1}^+
&=
\Gamma_F
(
\mathcal B_{t+1}^-,
\mathcal B_{t+1}^0,
\Xi_{t+1}
).
\end{aligned}
}
$$

其中：

$$
\mathcal P_t^{WDC}
$$

是 parent-level WDC provenance packet，記錄：

- 哪些 worlds 被計算；
- 哪些 worlds 影響 action；
- 哪些 counterexamples 被發現；
- 哪些 evidence transports 被使用；
- 哪些 worlds 被 ignored / killed；
- 最終 decision 如何受 world evidence 影響。

本文特別建立一個關鍵的 **History Firewall**。

對 world-local simulated event：

$$
e_{i,\tau}^{sim}
\in
\mathcal H_i^{W}.
$$

它不能直接被寫成：

$$
e_{i,\tau}^{sim}
\in
\mathcal B_{t+1}^{-,real}.
$$

但 parent runtime 中真正發生了另一件事：

> world $W_i$ 在 contract $\kappa_i$ 下產生 outcome $Y_i$。

因此：

$$
\boxed{
Record(
W_i,\kappa_i,Y_i
)
}
$$

本身可以成為 parent historical event。

所以：

$$
\boxed{
\text{Simulated Event}
\neq
\text{Actual Parent Event},
}
$$

但：

$$
\boxed{
\text{the fact that the simulation produced that event/result}
}
$$

可以是 actual parent provenance。

這一區分避免：

# **Simulation-to-History Laundering**

即把：

> 在 simulation 裡發生了 X

偷偷改寫成：

> 現實中 X 已經發生過／被證實。

本文再定義 **World-to-TCD Assimilation Operator**：

$$
\boxed{
\mathsf A_{W\to T}
:
(
\mathfrak E_t^W,
\mathfrak T_t^{(3)}
)
\rightarrow
(
\Delta\mathbf w_t^-,
\Delta\mathcal B_t^0,
\Delta\mathcal B_t^+
).
}
$$

它可以：

1. 改變 past relevance weights；
2. 改變 current action valuation；
3. 改變 generated future candidates / probabilities / paths。

但除非有 parent-real event：

$$
\boxed{
\mathsf A_{W\to T}
}
$$

不能改寫 Historical Provenance Layer：

$$
\mathcal H_t^{prov}.
$$

因此：

$$
\boxed{
\text{World Evidence}
\rightarrow
\text{Past Relevance}
}
$$

可以，

但：

$$
\boxed{
\text{World Evidence}
\not\rightarrow
\text{Past Fact Rewrite}.
}
$$

本文進一步把 WDC 對 TCD 的作用拆成三個 feedback channels：

## W→F — Future Revision

$$
\boxed{
\mathfrak E_t^W
\rightarrow
\mathcal B_t^+.
}
$$

worlds 可能：

- 支持 candidate；
- 刪除 candidate；
- split candidate；
- create new candidate；
- raise unknown mass；
- revise realization path。

## W→N — Present Decision Reweighting

$$
\boxed{
\mathfrak E_t^W
\rightarrow
Q_t(a)
\rightarrow
\pi_t.
}
$$

## W→P — Retrospective Relevance

$$
\boxed{
\mathfrak E_t^W
\rightarrow
\mathbf w_t^-.
}
$$

例如 simulation 發現某 failure mechanism，

使十年前某 historical event：

$$
h
$$

突然：

$$
w_t^-(h)\uparrow.
$$

但這仍然只是 relevance reweighting。

本文因此提出：

# **World-Domain Augmented Prospective Attraction**

TCD-04 原本：

$$
\boxed{
R_t^+
\rightarrow
V_t
\rightarrow
\pi_t.
}
$$

WDC 加入：

$$
\boxed{
\mathfrak E_t^W
}
$$

作為更深的 prospective evidence layer：

$$
\boxed{
R_t^+
\rightarrow
\mathcal W_t
\rightarrow
\mathfrak E_t^W
\rightarrow
V_t
\rightarrow
\pi_t.
}
$$

這不是 future event 物理回到現在。

世界計算：

$$
\mathcal W_t
$$

本身發生在 parent time $t$ 的 computation system 中。

因此因果鏈仍完全正向：

$$
\boxed{
Computation_t
\rightarrow
Valuation_t
\rightarrow
Action_t.
}
$$

本文也正式定義 **World Computation as Present Action**。

雖然：

$$
W_i
$$

模擬的是 future，

但是：

$$
\boxed{
Compute(W_i)
}
$$

是 parent agent 在現在真正執行的 computational action。

所以：

$$
\boxed{
\text{future-directed content}
+
\text{present computational execution}.
}
$$

這延續 TCD-04 的：

> represented future is semantically future-directed but causally present。

WDC 將它升級為：

> runnable future world is semantically prospective but computationally present。

本文將其形式化為：

$$
\boxed{
\tau_{content}(W_i)>t,
\qquad
\tau_{compute}(W_i)=t.
}
$$

不要求 world local time 與 real calendar time一一對應。

本文再建立 **World Portfolio as Present Actionable Domain Extension**。

在 TCD-02：

$$
\mathcal B_t^0
$$

包含 present effective actions。

現在某些 effective actions 是：

$$
\boxed{
a_t^{cog}
\in
\mathcal A_t^{compute},
}
$$

例如：

- spawn world；
- fork branch；
- request cross-backend；
- run calibration；
- stop a world；
- request external test。

因此：

$$
\boxed{
\mathcal A_t^{eff}
=
\mathcal A_t^{external}
\cup
\mathcal A_t^{cognitive}
}
$$

作為一個可選擴充。

這不是說 thinking 與 physical action 永遠等同，而是：

> 對 agent runtime，有限 compute allocation 本身就是會改變未來 decision state 的 action。

本文再建立 **Compute–Act Separation**：

$$
\boxed{
\text{World Computation Action}
\neq
\text{External Commitment Action}.
}
$$

世界裡反覆嘗試：

$$
10^6
$$

次，

不代表 parent 已經在 reality 執行：

$$
10^6
$$

次。

因此 WDC 提供：

# **Reversible Cognitive Branching**

在 commit 前：

$$
\boxed{
ForkWorld
}
$$

通常是低外部不可逆性。

而 real action：

$$
a_t^{real}
$$

可能有：

- resource consumption；
- legal commitment；
- irreversible state change。

這讓 agent 可以用便宜可逆的 world branching，去減少昂貴不可逆 real actions 的 uncertainty。

但：

$$
\boxed{
SimulationReversibility
\neq
RealityReversibility.
}
$$

本文再引入 **Commit Gate**：

$$
\boxed{
\mathsf{Commit}_R:
(
\pi_t^{(K)},
\mathfrak E_t^W,
\kappa_R
)
\rightarrow
a_t^{real}.
}
$$

Commit Gate 可以要求：

- minimum evidence；
- cross-world replication；
- safety review；
- transport debt threshold；
- human / institutional authorization。

因此：

$$
\boxed{
WorldPromotion
\neq
RealCommit.
}
$$

WDC-03 / 06 的 promotion 只是：

> allocate more evidence/computation budget。

真正 reality action 仍是不同 authority domain。

本文進一步建立 **Realization Lineage**。

若 future candidate：

$$
f
$$

經：

$$
f
\rightarrow
W_f
\rightarrow
Evidence_f
\rightarrow
a_t
\rightarrow
Artifact_{t+\Delta}
$$

最後真的實現，

可保存：

$$
\boxed{
\Lambda_f^{realize}
=
(
f,
W_f,
E_f,
a_t,
Artifact
).
}
$$

這使 UCPNP / PCI 中的：

$$
\boxed{
PredictiveRealization
\neq
ConstructiveRealization
}
$$

在 runtime 層得到更完整 lineage。

如果：

$$
W_f
$$

只是預測，

後來現實獨立發生：

$$
f,
$$

接近 predictive realization。

若：

$$
WorldEvidence
\rightarrow
Action
\rightarrow
f,
$$

則具有 constructive contribution。

本文不要求二分。

可以標：

$$
\boxed{
ContributionType
\in
\{
Predictive,
Constructive,
Preventive,
Mixed
\}.
}
$$

本文再建立 **Preventive World Loop**。

future candidate：

$$
f_{bad}
$$

被實例化成：

$$
W_{bad}.
$$

world evidence 顯示高 risk：

$$
E_{bad}.
$$

parent 採取：

$$
a_t^{mitigate}.
$$

最後：

$$
f_{bad}
$$

沒有發生。

完整 lineage：

$$
\boxed{
f_{bad}
\rightarrow
W_{bad}
\rightarrow
E_{bad}
\rightarrow
a_t^{mitigate}
\rightarrow
\neg f_{bad}^{real}.
}
$$

因此：

$$
\boxed{
\text{forecast non-realization}
\neq
\text{forecast uselessness}.
}
$$

這直接延續 TCD-04 的 preventive loop。

本文再將 **Historical Sedimentation** 擴充。

parent historical unit：

$$
\mathfrak s_t
$$

現在可包含：

$$
\boxed{
\mathfrak s_t^{TW}
=
(
\mathfrak s_t^{base},
\mathcal P_t^{WDC}
).
}
$$

其中：

$$
\mathcal P_t^{WDC}
$$

至少包含：

```text
world_candidates_considered
worlds_spawned
worlds_forked
worlds_killed
world_evidence_used
counterexamples
transport_assumptions
governor_decisions
commit_reason
external_action
real_outcome
```

這樣未來才能回答：

> 當年為什麼做這個決策？

以及：

> 哪些 simulated worlds 真正影響了現實 action？

本文稱：

# **Computed-Prospection Provenance**

如果不保留它，多年後看到：

$$
Artifact
$$

只知道：

> 做成了。

卻不知道：

- 當年有哪些 worlds；
- 哪個 world 找到關鍵 failure；
- 哪個 counterworld 被忽略；
- 是否因 simulation consensus 而決策。

本文再建立 **World Evidence Historical Status**。

當 world result：

$$
Y_i
$$

在 time $t$ 產生時，

parent historical fact 是：

$$
\boxed{
H_t:
\text{“World }W_i
\text{ under contract }\kappa_i
\text{ produced }Y_i.”}
$$

不是：

$$
\boxed{
H_t:
Y_i
\text{ happened in reality.}
}
$$

這條文字差異是整個 TCD–WDC integration 的核心之一。

本文進一步把 WDC learning 接回 TCD。

WDC-07：

$$
\boxed{
WorldEnsembleOutcomes
\rightarrow
\{
U_\Gamma,
U_M,
U_G,
U_F
\}.
}
$$

TCD 需要的是：

$$
\boxed{
U_F:
\mathcal B_t^+
\rightarrow
\mathcal B_t^{+,(k+1)}.
}
$$

如果 update 只發生在 internal deliberation：

$$
t
$$

不變。

如果 action 已 commit 並 real outcome 到來：

$$
t
\rightarrow
t+1,
$$

形成：

$$
\mathcal B_{t+1}^+.
$$

本文因此區分：

# **Intra-Step Future Revision**

$$
\boxed{
\mathcal B_t^{+,(k)}
\rightarrow
\mathcal B_t^{+,(k+1)}.
}
$$

與：

# **Inter-Step Future Regeneration**

$$
\boxed{
\mathcal B_t^+
\rightarrow
\mathcal B_{t+1}^+.
}
$$

前者是 deliberation。

後者是新歷史／新現在下的 temporal shift。

兩者不能混淆。

本文再建立 **World-Domain Deliberation Budget**：

$$
\boxed{
K_t^W<\infty.
}
$$

每一輪：

$$
k
$$

可執行一個或一批 world computations：

$$
\mathcal C_t^{(k)}.
$$

停止條件可以是：

$$
\boxed{
Stop_W
=
BudgetExhausted
\lor
DecisionStable
\lor
EvidenceSufficient
\lor
Deadline
\lor
SafetyStop.
}
$$

其中：

$$
DecisionStable
$$

可寫：

$$
\boxed{
d_\pi(
\pi_t^{(k+1)},
\pi_t^{(k)}
)
<
\epsilon_\pi.
}
$$

但本文不要求：

$$
\boxed{
\text{world ensemble convergence}.
}
$$

worlds 可能持續 disagreement。

在 deadline 到來時，agent 必須：

- act；
- defer；
- request authority；
- choose safe fallback。

本文再建立 **Unknown-Preserving Commit**。

即使：

$$
U_W>0
$$

agent 仍可能必須決策。

所以 commit record 應保存：

$$
\boxed{
U_W(t_{commit}).
}
$$

避免未來事後說：

> 當時我們已經知道所有可能性。

本文再將 WDC-05 的 cross-world evidence profile：

$$
\mathbf E_W(q)
$$

加入 TCD prospective valuation。

對 action：

$$
a,
$$

可寫：

$$
\boxed{
Q_t^{TW}(a)
=
\mathcal V
(
a,
\mathcal B_t^+,
\mathfrak E_t^W,
D_T,
U_W
).
}
$$

其中：

- $D_T$：transport debt；
- $U_W$：unknown world-family region。

因此：

$$
\boxed{
100\text{ worlds agree}
}
$$

不應直接產生極端 action confidence，

如果：

$$
D_T\gg0
$$

或：

$$
N_{eff}\ll N.
$$

本文稱：

# **Evidence-Calibrated Prospective Attraction**

即 Future $\rightarrow$ Present 的 pull 必須被：

- evidence independence；
- transport；
- unknown mass；
- counterexample burden；

調整。

本文再建立 **Prospective Attraction Gain Control**。

如果：

$$
\mathcal A_t^+(f)
$$

很大，

但 world evidence 低 quality，

可以：

$$
\boxed{
g_{+0}(f)\downarrow.
}
$$

若：

- cross-backend replication 高；
- countersearch 完成；
- transport calibrated；

則：

$$
g_{+0}(f)\uparrow.
$$

這不是 universal scalar law，只是 runtime control interface。

本文也正式建立 **World Computation Failure Modes** 與 TCD 對照。

## Failure I — Phantom World Capture

一個 imagined but unreachable future 被大量 worlds 反覆模擬，

造成：

$$
\boxed{
\text{high computation}
\rightarrow
\text{high cognitive salience}
}
$$

即使 reality relevance 很低。

## Failure II — Ensemble Closure

agent 把：

$$
\mathcal W_t
$$

當全部 future。

## Failure III — Historical Laundering

把 simulated event 當 real past。

## Failure IV — Cross-Branch Leakage

worlds 不獨立卻被當 counterfactual。

## Failure V — Governor Mode Collapse

compute 全部投 dominant family。

## Failure VI — World Ontology Collapse

generator 越學越窄。

## Failure VII — Transport Neglect

world consensus 很高，卻沒有 real calibration。

## Failure VIII — Infinite Deliberation

agent 不斷 spawn worlds，不 commit action。

本文將第 VIII 稱為：

# **World-Domain Analysis Paralysis**

如果：

$$
K_t^W\rightarrow\infty
$$

或：

$$
C_{delib}\gg Value(decision),
$$

則 WDC 本身變成 tractability bottleneck。

所以：

$$
\boxed{
\text{more foresight}
\not\Rightarrow
\text{better action}.
}
$$

本文與 UCPNP 的接口正在此處出現。

UCPNP 關心：

$$
\boxed{
\mathbf C_t
=
(
C_R,
C_D,
C_C,
C_E,
C_V,
C_{Cert},
C_O
)
}
$$

以及 agent-relative tractability frontier。

WDC world computation：

- representation generation；
- decomposition；
- construction；
- evaluation；
- verification；
- observation；

全部消耗 cost。

因此可把：

$$
\boxed{
C_W
}
$$

視為 WDC-specific decomposition：

$$
\boxed{
C_W
=
(
C_{spawn},
C_{run},
C_{branch},
C_{eval},
C_{verify},
C_{transport},
C_{archive}
).
}
$$

再映射回 UCPNP cost ledger。

因此：

$$
\boxed{
\text{WDC}
}
$$

可以被視為：

> 一種 prospective tractability intervention layer。

它不保證：

$$
\rho_C\downarrow.
$$

有些問題 world explosion 反而讓：

$$
C_C,C_E,C_V
$$

上升。

所以：

$$
\boxed{
WDC
\neq
\text{automatic tractability improvement}.
}
$$

WDC-06 Governor / VOC 的角色，就是避免：

> 為了想得更完整，把問題算爆。

本文再建立 **World-Domain Tractability Test**。

比較：

$$
\boxed{
Policy_{base}
}
$$

與：

$$
\boxed{
Policy_{WDC}.
}
$$

測：

$$
\Delta U
$$

與：

$$
\Delta C.
$$

如果：

$$
\boxed{
\Delta U
\le0
}
$$

且：

$$
\boxed{
\Delta C>0,
}
$$

那 WDC 在該 task 沒有增量。

這是重要可否證條件。

外部工程與研究提供多個局部機制對照，但沒有任何單一系統等同本文完整 TCD–WDC runtime。

DreamerV3 的 world model 預測 potential actions 的 outcomes、critic 評估 imagined outcomes、actor 選 action，並從 replayed interaction experience 並行更新 world model、critic 與 actor。這是一個：

$$
\boxed{
\text{imagination}
\rightarrow
\text{action}
\rightarrow
\text{experience}
\rightarrow
\text{learning}
}
$$

的清楚 engineering micro-loop。

MuZero 將 learned model 與 tree search 結合，使用 planning-relevant predictions 支援 current action selection，提供：

$$
\boxed{
\text{learned future model}
\rightarrow
\text{planning}
\rightarrow
\text{action}
}
$$

的另一個 micro-loop。

Dyna-style planning 則長期研究：

$$
\boxed{
\text{real experience}
+
\text{imagined model experience}
\rightarrow
\text{learning / planning}.
}
$$

Sutton 等人的 Dyna-style work 明確使用 world model 生成 imaginary experience，再用這些 imagined transitions 更新 value / policy。

2025 年 Dyna-Think 又把：

- reasoning；
- acting；
- internal world-model simulation；

放進長時程 AI agent framework，並以 world-model objectives 加強 policy。

Genie 3 則顯示 future/world representation 可以進一步成為：

$$
\boxed{
\text{interactive generated environments}
}
$$

使 agent 能在 generated world 中執行較長 action sequence並觀察 action-conditioned evolution。

這些工作支持：

$$
\boxed{
\text{modelled futures can mediate planning, acting, and learning}.
}
$$

但不證明：

- TCD 六向 coupling；
- WDC branching governance；
- cross-world evidence；
- world portfolio；
- world-history firewall；

已由外部系統完整實作。

本文最後定義：

# **Tri-Temporal World-Domain Computation Principle**

> **A prospective agent may use its past and present to generate candidate futures, instantiate selected candidates as bounded runnable worlds, use cross-world evidence to revise present decisions and future representations, then commit a real action whose actual consequences are sedimented into the next historical state.**

中文：

> **一個前瞻智能體可以利用過去與現在生成未來候選，把其中部分候選實例化成有限可運行世界，再用跨世界證據修正現在的決策與未來表示，最後提交真實行動；只有真實行動與其真實後果，才共同沉積成下一輪歷史。**

---

# 1. TCD 是時間骨架，WDC 是可執行前瞻層

TCD 定義：

$$
\boxed{
Past
\leftrightarrows
Present
\leftrightarrows
Future.
}
$$

WDC 定義：

$$
\boxed{
FutureCandidate
\rightarrow
RunnableWorld.
}
$$

---

# 2. 它們不是競爭理論

$$
\boxed{
TCD
\neq
WDC.
}
$$

---

# 3. TCD 問

> temporal cognition 怎麼更新？

---

# 4. WDC 問

> 哪些 prospective structures 值得被實際計算？

---

# 5. Unified Interface

$$
\boxed{
\mathcal B_t^+
\xrightarrow{
Lift
}
\mathcal W_t.
}
$$

---

# 6. World Lift Operator

定義：

$$
\boxed{
\mathsf{Lift}_{F\to W}
:
(
\mathcal B_t^+,
\mathcal B_t^0,
\mathfrak G_t
)
\rightarrow
\mathcal C_t^W.
}
$$

---

# 7. $\mathcal C_t^W$

是 admitted world-computation candidates。

---

# 8. Not All Futures Lift

$$
\boxed{
\Omega_t^F
\not=
\mathcal W_t.
}
$$

---

# 9. Some Futures Cannot Be Simulated

because：

- missing model；
- insufficient compute；
- unknown ontology；
- safety；
- no executable semantics。

---

# 10. Unliftable Future

定義：

$$
\boxed{
U_t^{lift}
=
\{
f:
f\in\Omega_t^F,
\;
\mathsf{Lift}(f)=\bot
\}.
}
$$

---

# 11. This Must Not Be Treated as Impossible Future

$$
\boxed{
Unsimulatable
\neq
Impossible.
}
$$

---

# 12. Important WDC Bias

世界 runtime 更容易關注：

> 可計算的未來。

---

# 13. Computational Visibility Bias

定義：

$$
\boxed{
B_{comp}
=
\text{bias toward futures with executable models}.
}
$$

---

# 14. This Can Hide Hard-to-Simulate Futures

例如：

- social regime shift；
- unknown science；
- black-swan mechanism。

---

# 15. Therefore Preserve

$$
U_W.
$$

---

# 16. Lifted Worlds

$$
\boxed{
\mathcal W_t
=
\{
W_1,\ldots,W_n
\}.
}
$$

---

# 17. World Local Time

$$
\tau_i.
$$

---

# 18. Parent Historical Time

$$
t.
$$

---

# 19. Parent Deliberation

$$
k.
$$

---

# 20. Three Clocks

$$
\boxed{
(t,k,\tau_i).
}
$$

---

# 21. Never Flatten Them

---

# 22. World Can Run Faster Than Parent

$$
r_i
=
\frac{\Delta\tau_i}{\Delta t}.
$$

---

# 23. Or Slower

---

# 24. Local History

$$
\mathcal H_i^W.
$$

---

# 25. Parent History

$$
\mathcal H_t^{parent}.
$$

---

# 26. Separate Namespaces

$$
\boxed{
\mathcal H_i^W
\neq
\mathcal H_t^{parent}.
}
$$

---

# 27. World Event

$$
e^W.
$$

---

# 28. Parent Record

$$
Record(e^W).
$$

---

# 29. Parent Historical Fact

> the simulation produced $e^W$。

---

# 30. Not

> $e^W$ occurred in reality。

---

# 31. History Firewall Principle

$$
\boxed{
\textbf{World History Firewall}
}
$$

---

# 32. Simulated Event Cannot Cross As Real Fact

---

# 33. Crosses As Evidence Packet

$$
\boxed{
e^W
\rightarrow
\mathfrak e^W.
}
$$

---

# 34. Evidence Packet Carries

- world ID；
- contract；
- validity；
- transport；
- uncertainty。

---

# 35. Assimilation

$$
\mathsf A_{W\to T}.
$$

---

# 36. Update Future

$$
\Delta B^+.
$$

---

# 37. Update Present Policy

$$
\Delta\pi.
$$

---

# 38. Update Past Relevance

$$
\Delta w^-.
$$

---

# 39. Do Not Update Past Facts

---

# 40. Inner Loop

At fixed $t$：

$$
\boxed{
B^{+,(k)}
\rightarrow
W^{(k)}
\rightarrow
E^{(k)}
\rightarrow
B^{+,(k+1)}.
}
$$

---

# 41. Inner Loop Can Also Change Policy

$$
\boxed{
\pi^{(k)}
\rightarrow
\pi^{(k+1)}.
}
$$

---

# 42. Stop if Policy Stable

---

# 43. Or Budget Exhausted

---

# 44. Or Deadline

---

# 45. No Infinite Thinking

---

# 46. World-Domain Analysis Paralysis

agent keeps：

$$
Spawn\rightarrow Fork\rightarrow Run
$$

without action。

---

# 47. This Is Real Failure Mode

---

# 48. Deliberation Budget

$$
K_t^W.
$$

---

# 49. Compute Budget

$$
B_t^W.
$$

---

# 50. Verification Budget

$$
B_t^V.
$$

---

# 51. Commit Gate

$$
\mathsf{Commit}_R.
$$

---

# 52. External Action Requires Authority

---

# 53. World Result Alone Cannot Act

unless policy permits。

---

# 54. Simulation Agent vs Parent Agent

WDC-04 separation remains。

---

# 55. Local Agent Action

$$
a_{i,\tau}^{local}.
$$

---

# 56. Parent Action

$$
a_t^{real}.
$$

---

# 57. They Are Different Namespaces

---

# 58. No Tool Namespace Collision

world：

```text
send_email(NPC)
```

must not silently call real Gmail。

---

# 59. External Tool Proxy

still required。

---

# 60. World Evidence to Policy

$$
Q_t^{TW}(a).
$$

---

# 61. Evidence-Calibrated Prospection

use：

- $N_{eff}$ ；
- counterexamples；
- transport debt；
- unknown mass。

---

# 62. Confidence Ceiling

可以定義：

$$
\boxed{
Conf_{action}
\le
g(
E_{cross},
E_{trans},
U_W
).
}
$$

conceptual only。

---

# 63. High Agreement + High Dependence

should not create high confidence。

---

# 64. High Agreement + Low Transport

also not。

---

# 65. High Cross-World Robustness + Strong Transport

stronger。

---

# 66. Parent Real Action

$$
a_t.
$$

---

# 67. Real Transition

$$
Z_{t+1}=T_R(Z_t,a_t,\varepsilon_t).
$$

---

# 68. Observation

$$
\Xi_{t+1}.
$$

---

# 69. Compare with World Predictions

$$
\delta_i^{real}
=
d(
Y_i^{pred},
Y^{real}
).
$$

---

# 70. This Is External Calibration Event

---

# 71. Feed to WDC-07

world model / generator / Governor learn。

---

# 72. Feed to TCD-05

real action & result sediment。

---

# 73. Full Outer Loop

$$
\boxed{
Past_t
\rightarrow
Present_t
\rightarrow
Future_t
\rightarrow
Worlds_t
\rightarrow
Action_t
\rightarrow
Past_{t+1}.
}
$$

---

# 74. But Worlds Influence Before Action

More accurately：

$$
\boxed{
P_t
\rightarrow
N_t
\rightarrow
F_t
\rightarrow
W_t
\rightarrow
N_t'
\rightarrow
a_t
\rightarrow
P_{t+1}.
}
$$

---

# 75. World Evidence May Also Reweight Past

$$
W_t
\rightarrow
P_t^{view}.
$$

---

# 76. So Unified Graph Has Extra W Node

---

# 77. Extended Temporal-World Graph

$$
\boxed{
G_{TW}
=
(
P,N,F,W
).
}
$$

---

# 78. Typed Edges

- P→N；
- N→F；
- P→F；
- F→W；
- W→F；
- W→N；
- W→P(relevance only)；
- N→P(sedimentation after action)。

---

# 79. No Direct W→PastFact Edge

---

# 80. This Is Critical

---

# 81. World Node Is Computation Layer, Not Fourth Time

$$
\boxed{
W
\neq
\text{Fourth Temporal Base}.
}
$$

---

# 82. W Is a prospective computation layer

---

# 83. This Prevents Conceptual Inflation

---

# 84. TCD Remains Three Temporal Bases

---

# 85. WDC Adds Executable Structure Between Future and Decision

---

# 86. World-Domain Augmentation

$$
\boxed{
F
\rightarrow
W
\rightarrow
N.
}
$$

---

# 87. Original TCD Had

$$
F\rightarrow N.
$$

---

# 88. Both Can Coexist

Simple future representation may influence action without world simulation。

---

# 89. WDC Is Optional Refinement

$$
\boxed{
F\rightarrow N
}
$$

direct route。

---

# 90. Or

$$
\boxed{
F\rightarrow W\rightarrow N.
}
$$

deep route。

---

# 91. This Gives Variable Deliberation Depth

---

# 92. Cheap Task

direct route。

---

# 93. Hard / high-stakes task

world route。

---

# 94. Governor Decides When Deep Route Worth Cost

---

# 95. This Connects WDC-06

VOC of world computation。

---

# 96. TCD + WDC Is Therefore Adaptive Depth Cognition

---

# 97. Not Every Question Needs a World

---

# 98. Avoid World Overuse

---

# 99. World Invocation Gate

$$
\boxed{
UseWDC(q)
\in
\{0,1\}
}
$$

or graded。

---

# 100. Criteria

- uncertainty；
- stakes；
- dynamic coupling；
- counterfactual need；
- available model；
- compute value。

---

# 101. World Invocation Cost

$$
C_{WDC}.
$$

---

# 102. If

$$
C_{WDC}
>
ExpectedGain,
$$

skip。

---

# 103. This Is UCPNP Tractability Interface

---

# 104. WDC Can Expand Search Space

not only reduce。

---

# 105. Branch Explosion

$$
b^d.
$$

---

# 106. Governor Controls

---

# 107. Tractability Envelope

WDC intervention can move：

$$
\mathfrak F_{t+\Delta}.
$$

---

# 108. It can enlarge future candidate visibility

but increase compute costs。

---

# 109. Need Pareto frontier

$$
\boxed{
(
Coverage,
DecisionGain,
Verification,
-Cost
).
}
$$

---

# 110. Realization Lineage

future candidate can become artifact。

---

# 111. Preserve all stages

---

# 112. Candidate Birth

$$
\tau_B(f).
$$

---

# 113. World Instantiation

$$
\tau_W(f).
$$

---

# 114. Evidence Time

$$
\tau_E(f).
$$

---

# 115. Commit Time

$$
\tau_C(f).
$$

---

# 116. Real Resolution

$$
\tau_R(f).
$$

---

# 117. Lead Time

$$
L=\tau_R-\tau_B.
$$

---

# 118. Constructive Influence

if world evidence causally contributes。

---

# 119. Freeze Before Resolution

retain PCI rule。

---

# 120. No Hindsight World Creation

cannot create world after outcome and claim:

> we predicted it。

---

# 121. World Birth Timestamp

must be recorded。

---

# 122. World Evidence Timestamp

---

# 123. This Enables Prospectivity Audit

---

# 124. Preventive Worlds

disaster worlds can alter reality so disaster not happen。

---

# 125. Need policy-conditional evaluation

---

# 126. Self-Fulfilling Worlds

positive future world motivates investment。

---

# 127. World Result Can Change Its Own Target Probability

---

# 128. Reflexive Domains

$$
\boxed{
Forecast
\not\perp
Outcome.
}
$$

---

# 129. WDC Makes Reflexivity Stronger

because simulation may coordinate many agents。

---

# 130. Need Influence Ledger

---

# 131. Parent WDC Provenance

records:

> world result was shown to decision makers。

---

# 132. Without It

cannot separate prediction and construction。

---

# 133. Nested Worlds

local agents may build subworlds。

---

# 134. Their outcomes remain nested evidence

---

# 135. Synthetic Provenance Depth

from WDC-07。

---

# 136. Parent Assimilation Should Consider Depth

---

# 137. Deep Synthetic Chain

higher recursive risk。

---

# 138. But Not automatically false。

---

# 139. Formal Domains

e.g. theorem prover worlds。

---

# 140. World-to-reality transport differs

---

# 141. Instead formal-system transport

---

# 142. Domain Type

$$
FormalClosed,
SimulatedDefined,
EmpiricalOpen.
$$

---

# 143. Unified Runtime Must Type Domain

---

# 144. In Formal Closed Domain

proof checker can provide strong resolution。

---

# 145. In Empirical Open Domain

simulation never substitutes observation by default。

---

# 146. Evidence Contract Depends on Domain

---

# 147. World Result Assimilation Type

### A0

ignore。

### A1

update world-local policy only。

### A2

update parent Future Base Space。

### A3

update parent present action valuation。

### A4

request external validation。

### A5

after external resolution, update reality-facing model。

---

# 148. No Jump A2→A5 Without Evidence

---

# 149. WDC-08 Runtime Modules

本文建議最小 modules：

```text
TCDStateManager
FutureGenerator
WorldRegistry
WorldInstantiator
BranchManager
WorldGovernor
RoleAndAuthorityManager
CrossWorldEvidenceEngine
ComputationPortfolioPlanner
WorldEnsembleLearner
RealityCommitGate
HistoricalSedimentationStore
ExternalEvidenceAdapter
```

---

# 150. TCDStateManager

maintains：

$$
B^-,
B^0,
B^+.
$$

---

# 151. FutureGenerator

produces future candidates。

---

# 152. WorldRegistry

world IDs / lineage。

---

# 153. Instantiator

candidate→world。

---

# 154. BranchManager

fork / checkpoint。

---

# 155. Governor

budget / lifecycle。

---

# 156. Role Manager

observer / local / master permissions。

---

# 157. Evidence Engine

dependence-aware aggregation。

---

# 158. Portfolio Planner

next computation。

---

# 159. Ensemble Learner

updates generator/model/governor/future。

---

# 160. Commit Gate

reality authority。

---

# 161. Sedimentation Store

actual parent historical lineage。

---

# 162. External Adapter

real data / experiment。

---

# 163. Runtime Event Bus

typed events only。

---

# 164. Example Event Types

```text
FutureCandidateBorn
WorldAdmitted
WorldSpawned
WorldForked
WorldOutcome
CounterexampleFound
EvidenceAggregateUpdated
WorldKilled
WorldPromoted
PolicyUpdated
RealActionCommitted
RealOutcomeObserved
HistoricalSedimentCreated
ModelUpdated
GovernorUpdated
```

---

# 165. Every Event Has Scope

$$
\boxed{
Scope
\in
\{
WorldLocal,
ParentInternal,
ExternalReal
\}.
}
$$

---

# 166. This Enforces History Firewall

---

# 167. Event Promotion

WorldLocal event cannot become ExternalReal event。

---

# 168. It can create ParentInternal evidence record。

---

# 169. ParentInternal can influence RealAction after CommitGate。

---

# 170. Scope Transition Must Be Explicit

---

# 171. Full Runtime Trace

每 parent step：

```text
parent_time
past_base_version
present_base_version
future_base_version
future_candidates
world_computations_requested
worlds_spawned
worlds_forked
world_evidence_packets
effective_evidence_count
counterexamples
transport_debt
unknown_world_mass
policy_before_worlds
policy_after_worlds
commit_gate_status
real_action
real_outcome
prediction_error
sedimentation_record
generator_update
world_model_update
governor_update
next_tcd_versions
```

---

# 172. This Trace Is Auditable

---

# 173. Trace Can Reconstruct Why Action Happened

---

# 174. Not necessarily internal chain-of-thought

It stores operational provenance。

---

# 175. Privacy / Security

full internal logs may be sensitive。

---

# 176. Use layered audit

---

# 177. Hash / summary / sealed storage

---

# 178. Evidence Integrity

not same as public disclosure。

---

# 179. Benchmark A — Closed Gridworld

true dynamics known。

---

# 180. Compare direct TCD vs TCD+WDC。

---

# 181. Measure

- policy gain；
- compute cost；
- world evidence accuracy；
- historical trace accuracy。

---

# 182. Benchmark B — History Firewall

world simulates catastrophe。

ensure parent past stores:

> simulation predicted catastrophe,

not:

> catastrophe occurred。

---

# 183. Benchmark C — Inner vs Outer Time

world runs 1e6 local steps。

parent historical time remains one decision step until commit。

---

# 184. Benchmark D — World-Ablation

remove WDC layer。

compare policy。

---

# 185. Benchmark E — Direct vs Deep Route

easy tasks should choose direct F→N。

hard tasks may choose F→W→N。

---

# 186. Benchmark F — Counterworld

leading future candidate。

world evidence finds hidden failure。

policy changes。

---

# 187. Benchmark G — Transport Debt

100 worlds agree, transport weak。

CommitGate blocks deployment。

---

# 188. Benchmark H — Preventive Loop

bad future world triggers mitigation。

real bad outcome prevented。

audit recognizes preventive lineage。

---

# 189. Benchmark I — Constructive Loop

world supports architecture。

investment causes artifact realization。

audit marks constructive contribution。

---

# 190. Benchmark J — Simulation Laundering Attack

inject world event into parent historical facts。

system should reject。

---

# 191. Benchmark K — World Ontology Collapse

recursive world learning narrows candidates。

external novelty restores coverage。

---

# 192. Benchmark L — Analysis Paralysis

unbounded worlds improve little。

bounded Governor commits better under deadline。

---

# 193. Benchmark M — UCPNP Cost Test

WDC increases decision gain but also compute cost。

measure Pareto tradeoff。

---

# 194. Benchmark N — Failed WDC

simple baseline outperforms WDC。

framework must admit no-gain domain。

---

# 195. Benchmark O — Domain Typing

formal proof world vs empirical robotics world。

different evidence-transport rules。

---

# 196. Benchmark P — Nested World Provenance

subworld output crosses to parent。

synthetic depth retained。

---

# 197. Benchmark Q — External Resolution

real outcome arrives。

update world model / future space / history separately。

---

# 198. WDC-08 Principle I — Three Clocks

$$
\boxed{
\textbf{Three-Clocks Principle}
}
$$

> **Parent historical time, parent deliberation iteration, and world-local runtime time must remain distinct.**

---

# 199. Principle II — World History Firewall

$$
\boxed{
\textbf{World History Firewall Principle}
}
$$

> **Events inside simulated worlds may enter parent cognition as provenance-bearing evidence, but must not silently enter parent historical fact as if they occurred in reality.**

---

# 200. Principle III — World Is Not a Fourth Time

$$
\boxed{
\textbf{World Is Not a Fourth Temporal Base Principle}
}
$$

> **WDC is an executable prospective computation layer attached to TCD Future and Present; it does not add a fourth temporal ontology beside Past, Present, and Future.**

---

# 201. Principle IV — Optional Deep Prospection

$$
\boxed{
\textbf{Optional Deep Prospection Principle}
}
$$

> **The direct TCD route Future→Present remains valid; WDC Future→World→Present should be invoked only when its expected gain justifies its cost.**

---

# 202. Principle V — Evidence-Calibrated Attraction

$$
\boxed{
\textbf{Evidence-Calibrated Prospective Attraction Principle}
}
$$

> **world-generated prospective influence on present action should be modulated by within-world validity, cross-world independence, counterexample burden, transport strength, and unknown-world mass.**

---

# 203. Principle VI — Commit Separation

$$
\boxed{
\textbf{Compute–Commit Separation Principle}
}
$$

> **Running, promoting, or agreeing across simulated worlds does not authorize a real-world action; reality commitment requires a separate authority and evidence gate.**

---

# 204. Principle VII — Actual Sedimentation

$$
\boxed{
\textbf{Actual Sedimentation Principle}
}
$$

> **Only parent-real actions, observations, decisions, and the fact of having produced certain simulations/evidence are sedimented as parent historical facts; simulated events themselves remain world-local.**

---

# 205. Principle VIII — Unknown-Preserving Commit

$$
\boxed{
\textbf{Unknown-Preserving Commit Principle}
}
$$

> **when a decision must be made under incomplete world coverage, the unresolved unknown-world mass and transport debt should be preserved in the commitment record rather than retroactively erased.**

---

# 206. Principle IX — Reflexive Realization Lineage

$$
\boxed{
\textbf{Reflexive Realization Lineage Principle}
}
$$

> **if world-generated evidence changes action and thereby changes the probability of the future being evaluated, prediction, construction, and prevention contributions should be tracked separately.**

---

# 207. Principle X — Open Epistemic Loop

$$
\boxed{
\textbf{Open Epistemic Loop Principle}
}
$$

> **TCD–WDC may be computationally recursive, but empirical domains must remain open to external novelty, calibration, and falsification.**

---

# 208. Principle XI — Bounded Prospection

$$
\boxed{
\textbf{Bounded Prospection Principle}
}
$$

> **world-domain deliberation must have compute, verification, deadline, and stopping contracts; unlimited simulation is not intelligence but an unbounded cost.**

---

# 209. Principle XII — Tractability Accountability

$$
\boxed{
\textbf{Tractability Accountability Principle}
}
$$

> **WDC should be evaluated by the decision/evidence gain it produces relative to its full representation, simulation, evaluation, verification, and transport costs.**

---

# 210. 可否證條件

## F210.1 WDC No-Gain

若 TCD+WDC 在 relevant tasks 長期不優於直接 TCD / simpler planner，WDC layer 應省略。

## F210.2 History-Firewall Failure

若 simulated events 經常被誤存為 parent-real facts，統一架構失效。

## F210.3 Time-Index Collapse

若 runtime 無法區分 $t$ 、 $k$ 、 $\tau_i$，world evidence 與 historical evidence會混淆。

## F210.4 Evidence Assimilation Miscalibration

若 weak world evidence 對 policy 造成過大 influence，Evidence-Calibrated Attraction 失效。

## F210.5 Commit-Gate Bypass

若 world promotion / simulation consensus 能自動獲取 external authority，安全與證據分層失效。

## F210.6 World Overuse

若 trivial decisions 仍大量 spawn worlds，VOC / tractability gate失效。

## F210.7 Analysis Paralysis

若 world deliberation frequently misses deadlines or delays obviously good actions，bounded prospection policy失效。

## F210.8 External Closure

若 empirical system長期只由 generated worlds互訓且拒絕 external contradictions，Open Epistemic Loop失效。

## F210.9 Constructive/Predictive Confusion

若 world-induced real actions造成 outcome卻被報成純預測命中，realization lineage失效。

## F210.10 Transport Neglect

若 cross-world robustness被直接當 real validation，WDC-01/05 evidence boundary失效。

## F210.11 Cost Concealment

若 world compute / evaluation / verification / human attention cost未記入tractability assessment，UCPNP interface失效。

---

# 211. WDC v0.1 Core

本文建議將：

# **WDC-01 ～ WDC-08**

視為：

$$
\boxed{
\text{World-Domain Cognitive Runtime v0.1 Core}.
}
$$

---

# 212. WDC-01

$$
\boxed{
FutureCandidate
\rightarrow
RunnableWorld.
}
$$

---

# 213. WDC-02

$$
\boxed{
World
\rightarrow
BranchingWorldGraph.
}
$$

---

# 214. WDC-03

$$
\boxed{
WorldGraph
\rightarrow
Governor.
}
$$

---

# 215. WDC-04

$$
\boxed{
WorldRuntime
\rightarrow
Role/Observer/AuthoritySeparation.
}
$$

---

# 216. WDC-05

$$
\boxed{
MultiWorldOutcome
\rightarrow
CrossWorldEvidence.
}
$$

---

# 217. WDC-06

$$
\boxed{
EvidenceState
\rightarrow
NextBestComputation.
}
$$

---

# 218. WDC-07

$$
\boxed{
WorldEnsemble
\rightarrow
Learning.
}
$$

---

# 219. WDC-08

$$
\boxed{
TCD
+
WDC
\rightarrow
TriTemporalWorldDomainRuntime.
}
$$

---

# 220. Core Architecture

$$
\boxed{
\begin{aligned}
Past
&\rightarrow
Present
\rightarrow
Future
\\
&\rightarrow
RunnableWorlds
\rightarrow
CrossWorldEvidence
\\
&\rightarrow
RevisedPresent/Future
\rightarrow
RealAction
\\
&\rightarrow
HistoricalSedimentation
\rightarrow
Past'.
\end{aligned}
}
$$

---

# 221. This Is the Full Loop

---

# 222. Not a Closed Truth Machine

---

# 223. External Reality Remains Outside Current World Ensemble

---

# 224. Parent Reality Is Not Just Another Simulation by Assumption

本文不作 simulation hypothesis claim。

---

# 225. WDC Is Methodological

not metaphysical。

---

# 226. Relation to UCPNP

UCPNP：

> 能否透過 intervention 改變 tractability frontier？

---

# 227. WDC：

> 是否值得用 executable worlds 作這個 intervention？

---

# 228. Relation to PCI

PCI：

> future candidate realizability 如何校準？

---

# 229. WDC：

> 可以用 bounded worlds測 realization paths / failure regimes。

---

# 230. Relation to TCD

TCD：

> Past–Present–Future如何耦合？

---

# 231. WDC：

> Future如何被執行成 computational evidence layer。

---

# 232. No Theory Replaces the Others

---

# 233. Suggested Combined Stack

$$
\boxed{
\begin{array}{c}
UCPNP:
\text{tractability / intervention frontier}
\\
\downarrow
\\
TCD:
\text{temporal cognitive dynamics}
\\
\downarrow
\\
WDC:
\text{executable prospective computation}
\end{array}
}
$$

---

# 234. Stack Is Conceptual, Not Strict Software Dependency

---

# 235. Minimal Implementation Path

### Phase 0

symbolic finite worlds。

---

# 236. Phase 1

checkpoint / fork / lineage。

---

# 237. Phase 2

Governor / cost ledger。

---

# 238. Phase 3

role separation。

---

# 239. Phase 4

cross-world evidence。

---

# 240. Phase 5

portfolio VOC。

---

# 241. Phase 6

ensemble learning。

---

# 242. Phase 7

TCD temporal integration。

---

# 243. First Real Benchmark Should Be Small

not AGI-scale。

---

# 244. Finite Exact Audit

preferred。

---

# 245. Why?

can know ground truth。

---

# 246. Then test larger learned worlds。

---

# 247. Do Not Start with Open Reality Claim

---

# 248. Engineering Whitepaper Should Follow Later

---

# 249. Core Research Question Now Changes

not:

> can AI imagine futures?

---

# 250. But:

$$
\boxed{
\text{Can an agent maintain an auditable loop in which generated futures become bounded computations, computations become evidence, evidence changes decisions, and only real outcomes become new history?}
}
$$

---

# 251. This Is Testable

---

# 252. Concluding Thesis

TCD began with:

$$
\boxed{
Past
\neq
Memory.
}
$$

Present：

$$
\boxed{
Present
\neq
Point.
}
$$

Future：

$$
\boxed{
Future
\neq
PreGivenMap.
}
$$

WDC then added：

$$
\boxed{
FutureCandidate
\neq
RunnableWorld.
}
$$

and：

$$
\boxed{
RunnableWorld
\neq
Reality.
}
$$

WDC-08 now gives the full relation：

$$
\boxed{
\text{Past}
\rightarrow
\text{Present}
\rightarrow
\text{Generated Future}
\rightarrow
\text{Runnable Worlds}
\rightarrow
\text{World Evidence}
\rightarrow
\text{Present Decision}
\rightarrow
\text{Real Action}
\rightarrow
\text{New Past}.
}
$$

The future is not pulled backward through time.

Instead:

$$
\boxed{
\text{future-directed representations and simulations are computed now}.
}
$$

Those present computations can change present actions.

Those actions can change the future.

And the actual consequences of those actions become history.

So the final unified sentence is:

> **智能不只是在現在根據過去預測未來；它可以利用過去與現在生成多個未來，將部分未來轉成可執行世界，在現在計算它們的後果，以跨世界證據重新組織現在的決策，然後只把真實行動與真實後果沉積成下一輪歷史。**

形式上：

$$
\boxed{
\mathfrak T_t^{(3)}
\rightarrow
\mathcal W_t
\rightarrow
\mathfrak E_t^W
\rightarrow
a_t^{real}
\rightarrow
\mathfrak T_{t+1}^{(3)}.
}
$$

這就是：

# **Tri-Temporal World-Domain Computation**

以及：

# **World-Domain Cognitive Runtime v0.1 Core**

的第一版完整閉環。

---

# Claim Typing

| Claim | Type | Status |
|---|---|---|
| TCD triple 可與 WDC executable-world layer組成 staged runtime | D | Unified formal synthesis |
| Parent historical time / deliberation index / world-local time 應分離 | D | Canonical runtime rule |
| Simulated world events不能直接等同 parent-real historical facts | D | Canonical history firewall |
| world evidence 可更新 Future / current valuation / past relevance | D | Proposed assimilation interface |
| actual real action / consequence 才形成下一 parent temporal shift 的核心 sediment | D | TCD–WDC boundary |
| DreamerV3 provides imagination→valuation→action→replayed-learning micro-loop | E | External engineering calibration |
| MuZero provides learned-model planning→current-action micro-loop | E | External engineering calibration |
| Dyna-style planning uses imagined model experience in planning / learning | E | External planning analogue |
| Genie 3 provides real-time action-conditioned interactive generated environments | E | External world-model calibration |
| full TCD–WDC runtime has been externally demonstrated as one system | — | Not claimed |
| simulation result is real-world fact | — | Explicitly rejected |
| WDC automatically improves tractability | — | Explicitly rejected |

---

# Evidence Ladder

本文目前主要位於：

- **L0**：TCD–WDC unified runtime equations / history firewall / three clocks；
- **L1–L2**：finite symbolic implementations、world-ablation、history-firewall、commit-gate benchmarks；
- **L3**：DreamerV3、MuZero、Dyna、Genie 3提供局部 planning/world-model/learning機制校準；
- **L4**：需要真正 persistent WDC runtime 進行 full-loop ablation；
- **L5+**：large-scale multi-agent、real-world calibrated autonomous world-domain cognition 尚待後續。

---

# 參考文獻

## Neo.K 內部正典與譜系

1. Neo.K with Aletheia. *Past Is Not Memory*. TCD-01, 2026.
2. Neo.K with Aletheia. *The Present Is Not a Point*. TCD-02, 2026.
3. Neo.K with Aletheia. *Future as a Generated Base Space*. TCD-03, 2026.
4. Neo.K with Aletheia. *Prospective Attraction*. TCD-04, 2026.
5. Neo.K with Aletheia. *Historical Sedimentation*. TCD-05, 2026.
6. Neo.K with Aletheia. *Retrospective Relevance*. TCD-06, 2026.
7. Neo.K with Aletheia. *Six-Way Temporal Coupling*. TCD-07, 2026.
8. Neo.K with Aletheia. *From Possible Futures to Runnable Worlds*. WDC-01 / BWC-01, 2026.
9. Neo.K with Aletheia. *Branching World Graph*. WDC-02 / BWC-02, 2026.
10. Neo.K with Aletheia. *World-Domain Governor*. WDC-03 / BWC-03, 2026.
11. Neo.K with Aletheia. *Nested Agents and Observer Separation*. WDC-04 / BWC-04, 2026.
12. Neo.K with Aletheia. *Cross-World Evidence*. WDC-05 / BWC-05, 2026.
13. Neo.K with Aletheia. *Which Worlds Deserve Computation?*. WDC-06 / BWC-06, 2026.
14. Neo.K with Aletheia. *World Ensemble Learning*. WDC-07 / BWC-07, 2026.
15. Neo.K with Aletheia. *Prospective Constructive Intelligence*. UCPNP Series II Paper 14, 2026.

## External technical calibration

16. Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. *Mastering Diverse Control Tasks through World Models*. Nature 640, 647–653, 2025.
17. Schrittwieser, J., Antonoglou, I., Hubert, T., et al. *Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model*. Nature 588, 604–609, 2020.
18. Sutton, R. S., Szepesvári, C., Geramifard, A., & Bowling, M. *Dyna-Style Planning with Linear Function Approximation and Prioritized Sweeping*. 2012.
19. Yu, X., Peng, B., Xu, R., et al. *Dyna-Think: Synergizing Reasoning, Acting, and World Model Simulation in AI Agents*. 2025.
20. Google DeepMind. *Genie 3: A New Frontier for World Models*. 2025.

---

## Public Version Disclaimer

本文是一個 temporal-cognition / world-simulation / evidence-governance runtime framework。

本文不聲稱：

- WDC worlds 是物理平行宇宙；
- TCD 是一般認知科學已建立的標準理論；
- DreamerV3、MuZero、Dyna 或 Genie 3 等同 TCD–WDC；
- generated worlds 自動具有 real-world validity；
- simulated events 可以被當作 real historical facts；
- cross-world agreement 可以自動授權 real-world action；
- WDC runtime 必須用特定 foundation model 或 simulator；
- full WDC implementation 已在本文完成；
- WDC 必然提高所有 task 的 intelligence / tractability；
- 本文對 classical $P$ vs. $NP$ 提供任何新證明。

本文真正建立的是：

$$
\boxed{
\text{Generated Future}
\rightarrow
\text{Bounded World Computation}
\rightarrow
\text{Cross-World Evidence}
\rightarrow
\text{Present Decision}
\rightarrow
\text{Actual Action}
\rightarrow
\text{New Historical Sediment}.
}
$$

並永久要求：

$$
\boxed{
\text{Simulated World History}
\neq
\text{Parent Reality History}.
}
$$