Past–Present–Future 六向認知耦合與時間位移算子
Six-Way Temporal Coupling: A Unified Shift Operator for Past–Present–Future Cognitive Dynamics
Tri-Temporal Cognitive Dynamics — 三生認知耦合動力學系列 TCD-07 / 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 first unified dynamics paper / six-way coupling synthesis
Canonical Non-Identity Statement
TCD-01 至 TCD-06 已分別建立:
B t − , B t 0 , B t + \boxed{
\mathcal B_t^-,
\qquad
\mathcal B_t^0,
\qquad
\mathcal B_t^+
} B t − , B t 0 , B t +
以及六個方向的機制雛型。
本文第一次統一:
T t ( 3 ) = ( B t − , B t 0 , B t + ) \boxed{
\mathfrak T_t^{(3)}
=
(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+
)
} T t ( 3 ) = ( B t − , B t 0 , B t + )
並建立時間位移:
S t : T t ( 3 ) → T t + 1 ( 3 ) . \boxed{
\mathscr S_t:
\mathfrak T_t^{(3)}
\rightarrow
\mathfrak T_{t+1}^{(3)}.
} S t : T t ( 3 ) → T t + 1 ( 3 ) .
但永久保留:
Six Coupling Directions ≠ Six Identical Operators . \boxed{
\text{Six Coupling Directions}
\neq
\text{Six Identical Operators}.
} Six Coupling Directions = Six Identical Operators .
以及:
Temporal Coupling ≠ Physical Retrocausality . \boxed{
\text{Temporal Coupling}
\neq
\text{Physical Retrocausality}.
} Temporal Coupling = Physical Retrocausality .
本文不主張:
Past、Present、Future 是三個本體上獨立的宇宙層;
六條箭頭具有同樣的數學型別;
Future Event 會物理地改變 Past Fact;
任何 agent 都需要完整六向 coupling;
六向 coupling 越強 intelligence 越高;
TCD 是一般認知科學的已證實普遍定律;
同一時間內的 iterative deliberation 是時間旅行;
所有 coupling 都會收斂;
TCD 可直接取代 POMDP、MPC、RL、world models、memory systems 或 causal models;
本文已給出完整跨域實證證明。
摘要
Tri-Temporal Cognitive Dynamics(TCD)前六篇依序完成三個時間底座與六個 directional couplings。
Past:
B t − = compressed historical choice lineage . \boxed{
\mathcal B_t^-
=
\text{compressed historical choice lineage}.
} B t − = compressed historical choice lineage .
Present:
B t 0 = historically conditioned actionable / reachable domain . \boxed{
\mathcal B_t^0
=
\text{historically conditioned actionable / reachable domain}.
} B t 0 = historically conditioned actionable / reachable domain .
Future:
B t + = agent-generated prospective possibility domain . \boxed{
\mathcal B_t^+
=
\text{agent-generated prospective possibility domain}.
} B t + = agent-generated prospective possibility domain .
六個方向分別為:
B − → B 0 : Historical Conditioning , B 0 → B + : Present-Conditioned Future Generation , B + → B 0 : Prospective Attraction , B 0 → B − : Historical Sedimentation , B − → B + : Historical Projection , B + → B − : Retrospective Relevance . \boxed{
\begin{array}{lll}
\mathcal B^- \rightarrow \mathcal B^0
&:&
\text{Historical Conditioning},\\
\mathcal B^0 \rightarrow \mathcal B^+
&:&
\text{Present-Conditioned Future Generation},\\
\mathcal B^+ \rightarrow \mathcal B^0
&:&
\text{Prospective Attraction},\\
\mathcal B^0 \rightarrow \mathcal B^-
&:&
\text{Historical Sedimentation},\\
\mathcal B^- \rightarrow \mathcal B^+
&:&
\text{Historical Projection},\\
\mathcal B^+ \rightarrow \mathcal B^-
&:&
\text{Retrospective Relevance}.
\end{array}
} B − → B 0 B 0 → B + B + → B 0 B 0 → B − B − → B + B + → B − : : : : : : Historical Conditioning , Present-Conditioned Future Generation , Prospective Attraction , Historical Sedimentation , Historical Projection , Retrospective Relevance .
本文最重要的第一個統一原則是:
the six arrows are typed . \boxed{
\textbf{the six arrows are typed}.
} the six arrows are typed .
它們分別可能是:
state construction;
candidate generation;
policy valuation;
provenance / dependency sedimentation;
historical projection;
relevance reweighting。
因此不能把:
B − , B 0 , B + \mathcal B^-
,
\mathcal B^0
,
\mathcal B^+ B − , B 0 , B +
當成三個 ordinary vectors,再把六條 coupling 當同一種線性 matrix edge。
本文定義 typed temporal coupling graph:
G T C D , t = ( V t , E t , τ E ) , \boxed{
G_{\mathrm{TCD},t}
=
(
V_t,
E_t,
\tau_E
),
} G TCD , t = ( V t , E t , τ E ) ,
其中:
V t = { B t − , B t 0 , B t + } , V_t
=
\{
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+
\}, V t = { B t − , B t 0 , B t + } ,
而:
τ E : E t → { C o n d i t i o n , G e n e r a t e , A t t r a c t , S e d i m e n t , P r o j e c t , R e w e i g h t } \tau_E:
E_t
\rightarrow
\{
Condition,
Generate,
Attract,
Sediment,
Project,
Reweight
\} τ E : E t → { C o n d i t i o n , G e n er a t e , A tt r a c t , S e d im e n t , P r o j ec t , R e w e i g h t }
記錄每條 edge 的 operator type。
本文進一步解決一個形式上的核心問題:如果在同一時刻同時寫:
B t + → B t 0 → B t + \mathcal B_t^+
\rightarrow
\mathcal B_t^0
\rightarrow
\mathcal B_t^+ B t + → B t 0 → B t +
以及:
B t + → B t − → B t + , \mathcal B_t^+
\rightarrow
\mathcal B_t^-
\rightarrow
\mathcal B_t^+, B t + → B t − → B t + ,
表面上會形成 circular definition。
因此本文區分:
t = environment / historical time index \boxed{
t
=
\text{environment / historical time index}
} t = environment / historical time index
與:
k = within-step deliberation index . \boxed{
k
=
\text{within-step deliberation index}.
} k = within-step deliberation index .
在同一 physical / historical time t t t 中,agent 可以執行 bounded deliberation:
B t + , ( k ) → w t − , ( k ) → π t ( k ) → B t + , ( k + 1 ) , \boxed{
\mathcal B_t^{+,(k)}
\rightarrow
\mathbf w_t^{-,(k)}
\rightarrow
\pi_t^{(k)}
\rightarrow
\mathcal B_t^{+,(k+1)},
} B t + , ( k ) → w t − , ( k ) → π t ( k ) → B t + , ( k + 1 ) ,
但只要:
a t a_t a t
尚未對環境執行,
就仍然位於:
t . \boxed{
t.
} t .
真正 time shift 發生在 action execution 與 environment transition 之後:
Z t → a t Z t + 1 . \boxed{
Z_t
\xrightarrow{a_t}
Z_{t+1}.
} Z t a t Z t + 1 .
本文因此提出一個完整 staged update:
B ~ t 0 = Φ − 0 ( B t − , Ξ t ) , B t + , ( 0 ) = Γ F ( B t − , B ~ t 0 , Ξ t ) , w t − , ( k ) = R + − ( H t p r o v , B t + , ( k ) , q t ) , π t ( k ) = Π ( B ~ t 0 , B t + , ( k ) , w t − , ( k ) ) , B t + , ( k + 1 ) = Γ F r e f i n e ( B t − , B ~ t 0 , π t ( k ) , Ξ t ) , a t ∼ π t ( K ) , Z t + 1 = T Z ( Z t , a t , ε t ) , B t + 1 − = H − ( B t − , B t 0 , B t + , a t , Z t + 1 , E t ) , B t + 1 0 = Φ − 0 ( B t + 1 − , Ξ t + 1 ) , B t + 1 + = Γ F ( B t + 1 − , B t + 1 0 , Ξ t + 1 ) . \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
),
\\
\mathbf w_t^{-,(k)}
&=
\mathcal R_{+-}
(
\mathcal H_t^{prov},
\mathcal B_t^{+,(k)},
q_t
),
\\
\pi_t^{(k)}
&=
\Pi
(
\widetilde{\mathcal B}_t^0,
\mathcal B_t^{+,(k)},
\mathbf w_t^{-,(k)}
),
\\
\mathcal B_t^{+,(k+1)}
&=
\Gamma_F^{refine}
(
\mathcal B_t^-,
\widetilde{\mathcal B}_t^0,
\pi_t^{(k)},
\Xi_t
),
\\
a_t
&\sim
\pi_t^{(K)},
\\
Z_{t+1}
&=
T_Z(
Z_t,
a_t,
\varepsilon_t
),
\\
\mathcal B_{t+1}^-
&=
\mathcal H_-
(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+,
a_t,
Z_{t+1},
E_t
),
\\
\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}
} B t 0 B t + , ( 0 ) w t − , ( k ) π t ( k ) B t + , ( k + 1 ) a t Z t + 1 B t + 1 − B t + 1 0 B t + 1 + = Φ − 0 ( B t − , Ξ t ) , = Γ F ( B t − , B t 0 , Ξ t ) , = R +− ( H t p r o v , B t + , ( k ) , q t ) , = Π ( B t 0 , B t + , ( k ) , w t − , ( k ) ) , = Γ F r e f in e ( B t − , B t 0 , π t ( k ) , Ξ t ) , ∼ π t ( K ) , = T Z ( Z t , a t , ε t ) , = H − ( B t − , B t 0 , B t + , a t , Z t + 1 , E t ) , = Φ − 0 ( B t + 1 − , Ξ t + 1 ) , = Γ F ( B t + 1 − , B t + 1 0 , Ξ t + 1 ) .
其中:
Ξ t \Xi_t Ξ t :當前 exogenous information / environment / other-agent input;
H t p r o v \mathcal H_t^{prov} H t p r o v :TCD-06 的 stable historical provenance layer;
w t − \mathbf w_t^- w t − :retrospective relevance layer;
K K K :bounded deliberation budget;
T Z T_Z T Z :actual environment / system transition;
H − \mathcal H_- H − :Historical Sedimentation update;
Γ F \Gamma_F Γ F :Future Base Space generation operator。
因此:
Deliberative Recursion ≠ Physical Time Loop . \boxed{
\text{Deliberative Recursion}
\neq
\text{Physical Time Loop}.
} Deliberative Recursion = Physical Time Loop .
同一個 t t t 內可以反覆:
想未來 → 重看過去 → 改估值 → 再想未來。
只有真正 action execution 才讓:
t → t + 1. t
\rightarrow
t+1. t → t + 1.
本文也正式提出 Temporal Shift Operator :
S t ( B t − , B t 0 , B t + ; Ξ t , ε t ) = ( B t + 1 − , B t + 1 0 , B t + 1 + ) . \boxed{
\mathscr S_t
\left(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+;
\Xi_t,
\varepsilon_t
\right)
=
\left(
\mathcal B_{t+1}^-,
\mathcal B_{t+1}^0,
\mathcal B_{t+1}^+
\right).
} S t ( B t − , B t 0 , B t + ; Ξ t , ε t ) = ( B t + 1 − , B t + 1 0 , B t + 1 + ) .
這是 TCD 第一個完整 dynamical-system scaffold。
外部 AI、控制與認知研究提供了六向架構的局部工程對照,但沒有任何單一外部系統等同完整 TCD。MuZero 以 learned future-relevant dynamics 與 tree search 支援 current action selection;DreamerV3 的 world model 預測 potential actions outcomes、critic 評估 imagined outcomes、actor 學習 present policy;Mattar–Daw 的 prioritized-memory theory 則依某 memory 對未來 decision improvement 的 utility 選擇現在應 replay 的 past;Momennejad 等人的 human revaluation experiment 顯示 offline replay of distal past states 與之後 replanning 相關。2025 年的 navigation neuroscience 結果進一步顯示,新資訊可以快速改變 prospective representations of future choices,並伴隨 flexible route switching。這些局部結果共同支持一個保守結論:
past representations, prospective representations, and present choices can participate in recurrent decision loops . \boxed{
\text{past representations, prospective representations, and present choices can participate in recurrent decision loops}.
} past representations, prospective representations, and present choices can participate in recurrent decision loops .
但本文仍不宣稱六向 TCD 已被任何單一 biological or artificial system 實驗完整驗證。
本文最後提出:
Intelligence is not only a function from present state to action; for history-bearing prospective agents, it may be more usefully modeled as a bounded recurrent transformation among remembered/structured past, actionable present, and generated future. \boxed{
\textbf{Intelligence is not only a function from present state to action; for history-bearing prospective agents, it may be more usefully modeled as a bounded recurrent transformation among remembered/structured past, actionable present, and generated future.}
} Intelligence is not only a function from present state to action; for history-bearing prospective agents, it may be more usefully modeled as a bounded recurrent transformation among remembered/structured past, actionable present, and generated future.
中文:
智能不只是在現在輸入狀態後輸出 action;對具有歷史與前瞻能力的 agent,它可以被更完整地描述成:過去、現在與生成未來之間的有限反覆轉換,最後再把 action 寫回世界與歷史。
本文將 TCD-01~07 封裝成第一版:
TCD v0.1 Core
下一個理論接口不再只是增加第七條時間箭頭,而是把:
B t + \boxed{
\mathcal B_t^+
} B t +
中的 prospective branches 從 cognitive objects 升級為:
runnable world instances . \boxed{
\text{runnable world instances}.
} runnable world instances .
也就是後續:
Branching World Computation
分支世界計算/世界域認知 Runtime
的正式入口。
關鍵詞: Tri-Temporal Cognitive Dynamics、Six-Way Coupling、Temporal Shift Operator、Deliberative Recursion、World Models、Replay、Prospective Attraction、Historical Sedimentation、Retrospective Relevance、UCPNP
1. TCD-07 的任務:從六篇局部理論變成一個系統
TCD-01~03 建立三個 object。
TCD-04~06 建立三組重要 feedback mechanisms。
現在需要問:
如果把它們放在同一個 agent 裡,究竟怎麼更新?
2. 三個基本時間域
2.1 Past
B t − \boxed{
\mathcal B_t^-
} B t −
是:
historical choice lineage 的 task-relative compressed structure。
3. Present
B t 0 \boxed{
\mathcal B_t^0
} B t 0
是:
historically conditioned actionable / reachable domain。
4. Future
B t + \boxed{
\mathcal B_t^+
} B t +
是:
agent-generated prospective possibility / choice domain。
5. Triple State
因此:
T t ( 3 ) = ( B t − , B t 0 , B t + ) . \boxed{
\mathfrak T_t^{(3)}
=
(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+
).
} T t ( 3 ) = ( B t − , B t 0 , B t + ) .
6. 為什麼不是三個普通 state vectors?
因為三者內部型別不同。
7. Past 可能包含
B t − = ( H t p r o v , w t − , D t , L t , O t ) . \boxed{
\mathcal B_t^-
=
(
\mathcal H_t^{prov},
\mathbf w_t^-,
D_t,
L_t,
O_t
).
} B t − = ( H t p r o v , w t − , D t , L t , O t ) .
8. Present 可能包含
B t 0 = ( Z t e s t , A t e f f , R ^ t : H e f f , V t , U t 0 ) . \boxed{
\mathcal B_t^0
=
(
Z_t^{est},
\mathcal A_t^{eff},
\widehat{\mathcal R}_{t:H}^{eff},
\mathcal V_t,
U_t^0
).
} B t 0 = ( Z t es t , A t e f f , R t : H e f f , V t , U t 0 ) .
9. Future 可能包含
B t + = ( Ω t F , P t , Π t , G t F , U t + , E t + ) . \boxed{
\mathcal B_t^+
=
(
\Omega_t^F,
P_t,
\Pi_t,
\mathcal G_t^F,
U_t^+,
E_t^+
).
} B t + = ( Ω t F , P t , Π t , G t F , U t + , E t + ) .
10. 所以 coupling 必須 Typed
一條:
B t − → B t 0 \mathcal B_t^-
\rightarrow
\mathcal B_t^0 B t − → B t 0
和:
B t + → B t − \mathcal B_t^+
\rightarrow
\mathcal B_t^- B t + → B t −
根本不是同一種 function。
11. Typed Coupling Graph
定義:
G T C D , t = ( V t , E t , τ E ) . \boxed{
G_{\mathrm{TCD},t}
=
(
V_t,
E_t,
\tau_E
).
} G TCD , t = ( V t , E t , τ E ) .
12. Vertex Set
V t = { B − , B 0 , B + } . \boxed{
V_t
=
\{
B^-,
B^0,
B^+
\}.
} V t = { B − , B 0 , B + } .
13. Edge-Type Map
τ E ( e ) ∈ { C o n d i t i o n , G e n e r a t e , A t t r a c t , S e d i m e n t , P r o j e c t , R e w e i g h t } . \boxed{
\tau_E(e)
\in
\{
Condition,
Generate,
Attract,
Sediment,
Project,
Reweight
\}.
} τ E ( e ) ∈ { C o n d i t i o n , G e n er a t e , A tt r a c t , S e d im e n t , P r o j ec t , R e w e i g h t } .
14. Edge I — Past → Present
Φ − 0 : B t − × Ξ t → B t 0 . \boxed{
\Phi_{-0}:
\mathcal B_t^-
\times
\Xi_t
\rightarrow
\mathcal B_t^0.
} Φ − 0 : B t − × Ξ t → B t 0 .
名稱:
Historical Conditioning
15. 它做什麼?
Past 的:
capability residue;
dependency;
resource history;
permissions;
path dependence;
lost / recovered options;
參與構造現在有效 domain。
16. 它不是 Memory Retrieval
Φ − 0 ≠ R e t r i e v e M e m o r y . \boxed{
\Phi_{-0}
\neq
RetrieveMemory.
} Φ − 0 = R e t r i e v e M e m or y .
因為很多 history 已經固著在 current structure。
17. Edge II — Present → Future
Γ 0 + : B t 0 → B t + . \boxed{
\Gamma_{0+}:
\mathcal B_t^0
\rightarrow
\mathcal B_t^+.
} Γ 0 + : B t 0 → B t + .
名稱:
Present-Conditioned Future Generation
18. 它做什麼?
Present 決定:
current capabilities;
current resources;
current permissions;
current reachable paths;
因此限制/啟發 agent 生成哪些 futures。
19. Edge III — Future → Present
Φ + 0 : ( B t + , B t 0 ) → π t . \boxed{
\Phi_{+0}:
(
\mathcal B_t^+,
\mathcal B_t^0
)
\rightarrow
\pi_t.
} Φ + 0 : ( B t + , B t 0 ) → π t .
名稱:
Prospective Attraction
20. 它做什麼?
Future representations 經:
probability;
value;
risk;
option preservation;
preparation;
改變現在 policy。
21. 永久邊界
Φ + 0 \boxed{
\Phi_{+0}
} Φ + 0
使用的是:
Rep t ( F u t u r e ) , \boxed{
\operatorname{Rep}_t(Future),
} Rep t ( F u t u r e ) ,
不是:
F u t u r e E v e n t t + Δ . \boxed{
FutureEvent_{t+\Delta}.
} F u t u r e E v e n t t + Δ .
22. Edge IV — Present → Past
S e d 0 − : ( B t 0 , a t , Z t + 1 , E t ) → Δ B t + 1 − . \boxed{
\mathsf{Sed}_{0-}:
(
\mathcal B_t^0,
a_t,
Z_{t+1},
E_t
)
\rightarrow
\Delta\mathcal B_{t+1}^-.
} Sed 0 − : ( B t 0 , a t , Z t + 1 , E t ) → Δ B t + 1 − .
名稱:
Historical Sedimentation
23. 它做什麼?
把:
actual action;
alternatives;
dependencies;
plasticity;
resource changes;
permission changes;
decision provenance;
寫入下一時刻 Past。
24. Edge V — Past → Future
Γ − + : B t − → B t + . \boxed{
\Gamma_{-+}:
\mathcal B_t^-
\rightarrow
\mathcal B_t^+.
} Γ −+ : B t − → B t + .
名稱:
Historical Projection
25. 它做什麼?
使用:
analogies;
prior failures;
historical trajectories;
recurring mechanisms;
latent capabilities;
dormant options;
生成/約束 future candidates。
26. Historical Projection 不是簡單 Extrapolation
Γ − + ≠ trend line only . \boxed{
\Gamma_{-+}
\neq
\text{trend line only}.
} Γ −+ = trend line only .
它可以:
recombine past;
invert past failure;
reopen lost branches;
abstract mechanism。
27. Edge VI — Future → Past
R + − : ( H t p r o v , B t + , q t ) → w t − . \boxed{
\mathcal R_{+-}:
(
\mathcal H_t^{prov},
\mathcal B_t^+,
q_t
)
\rightarrow
\mathbf w_t^-.
} R +− : ( H t p r o v , B t + , q t ) → w t − .
名稱:
Retrospective Relevance
28. 它做什麼?
Future target 改變:
retrieval;
attention;
support search;
counterevidence search;
historical reactivation。
29. 它不能改 Past Fact
R + − : weights change \boxed{
\mathcal R_{+-}
:
\text{weights change}
} R +− : weights change
而不是:
H t p r o v : facts change . \boxed{
\mathcal H_t^{prov}
:
\text{facts change}.
} H t p r o v : facts change .
30. 六條 Edge 的 Mechanism Table
Edge
Name
Primary Mechanism
− → 0 -\to0 − → 0
Historical Conditioning
state construction
0 → + 0\to+ 0 → +
Present-Conditioned Future Generation
prospective generation
+ → 0 +\to0 + → 0
Prospective Attraction
valuation / policy
0 → − 0\to- 0 → −
Historical Sedimentation
provenance / structural update
− → + -\to+ − → +
Historical Projection
analogy / mechanism generation
+ → − +\to- + → −
Retrospective Relevance
retrieval / relevance reweighting
31. 六向不等於六個對稱 Arrow
例如:
Φ − 0 ≠ R + − − 1 . \boxed{
\Phi_{-0}
\neq
\mathcal R_{+-}^{-1}.
} Φ − 0 = R +− − 1 .
32. 同樣:
Γ 0 + ≠ S e d 0 − − 1 . \boxed{
\Gamma_{0+}
\neq
\mathsf{Sed}_{0-}^{-1}.
} Γ 0 + = Sed 0 − − 1 .
33. 過去、現在、未來也不是 Information-Conserving Transform
一般:
S \boxed{
\mathscr S
} S
可以:
lose information;
add exogenous information;
generate candidates;
compress history。
34. TCD 因此不是 Reversible Time Physics
它是:
agent-level cognitive dynamics . \boxed{
\text{agent-level cognitive dynamics}.
} agent-level cognitive dynamics .
35. Circularity Problem
如果直接寫:
B 0 → B + → B 0 , B^0
\rightarrow
B^+
\rightarrow
B^0, B 0 → B + → B 0 ,
會問:
哪個先算?
36. 第二個 Circularity
B − → B + → B − . B^-
\rightarrow
B^+
\rightarrow
B^-. B − → B + → B − .
也會問:
是 history 先生成 future,還是 future 先重看 history?
37. 解法:分離兩種時間 Index
本文定義:
t = historical / environment time \boxed{
t
=
\text{historical / environment time}
} t = historical / environment time
與:
k = deliberation iteration . \boxed{
k
=
\text{deliberation iteration}.
} k = deliberation iteration .
38. t t t 的意義
當 agent 真正:
act;
environment transition;
resource change;
external event;
才進入下一個:
t + 1. t+1. t + 1.
39. k k k 的意義
同一 decision moment 中:
想一下。
可以:
k → k + 1. k\rightarrow k+1. k → k + 1.
40. Deliberation Does Not Advance Historical Time
k → k + 1 ⇏ t → t + 1. \boxed{
k\rightarrow k+1
\not\Rightarrow
t\rightarrow t+1.
} k → k + 1 ⇒ t → t + 1.
41. Initial Present Construction
先從 past + exogenous current information:
B ~ t 0 = Φ − 0 ( B t − , Ξ t ) . \boxed{
\widetilde{\mathcal B}_t^0
=
\Phi_{-0}
(
\mathcal B_t^-,
\Xi_t
).
} B t 0 = Φ − 0 ( B t − , Ξ t ) .
42. Initial Future Generation
B t + , ( 0 ) = Γ F ( B t − , B ~ t 0 , Ξ t ) . \boxed{
\mathcal B_t^{+,(0)}
=
\Gamma_F
(
\mathcal B_t^-,
\widetilde{\mathcal B}_t^0,
\Xi_t
).
} B t + , ( 0 ) = Γ F ( B t − , B t 0 , Ξ t ) .
43. Γ F \Gamma_F Γ F 同時接受 Past 與 Present
因此概念上:
Γ F = Couple ( Γ − + , Γ 0 + ) . \boxed{
\Gamma_F
=
\operatorname{Couple}
(
\Gamma_{-+},
\Gamma_{0+}
).
} Γ F = Couple ( Γ −+ , Γ 0 + ) .
這不是線性加法。
44. Retrospective Reweighting
第 k k k 輪:
w t − , ( k ) = R + − ( H t p r o v , B t + , ( k ) , q t ) . \boxed{
\mathbf w_t^{-,(k)}
=
\mathcal R_{+-}
(
\mathcal H_t^{prov},
\mathcal B_t^{+,(k)},
q_t
).
} w t − , ( k ) = R +− ( H t p r o v , B t + , ( k ) , q t ) .
45. Active Past
B t − , a c t , ( k ) = V i e w ( H t p r o v , w t − , ( k ) ) . \boxed{
\mathcal B_t^{-,act,(k)}
=
View(
\mathcal H_t^{prov},
\mathbf w_t^{-,(k)}
).
} B t − , a c t , ( k ) = V i e w ( H t p r o v , w t − , ( k ) ) .
46. Policy Update
π t ( k ) = Π ( B ~ t 0 , B t + , ( k ) , B t − , a c t , ( k ) ) . \boxed{
\pi_t^{(k)}
=
\Pi
(
\widetilde{\mathcal B}_t^0,
\mathcal B_t^{+,(k)},
\mathcal B_t^{-,act,(k)}
).
} π t ( k ) = Π ( B t 0 , B t + , ( k ) , B t − , a c t , ( k ) ) .
47. Future Refinement
policy candidate 本身可以改變:
如果我這樣行動,future 會怎樣?
因此:
B t + , ( k + 1 ) = Γ F r e f i n e ( B t − , B ~ t 0 , π t ( k ) , Ξ t ) . \boxed{
\mathcal B_t^{+,(k+1)}
=
\Gamma_F^{refine}
(
\mathcal B_t^-,
\widetilde{\mathcal B}_t^0,
\pi_t^{(k)},
\Xi_t
).
} B t + , ( k + 1 ) = Γ F r e f in e ( B t − , B t 0 , π t ( k ) , Ξ t ) .
48. 這形成 Deliberation Loop
F u t u r e ( k ) → P a s t V i e w ( k ) → P o l i c y ( k ) → F u t u r e ( k + 1 ) . \boxed{
Future^{(k)}
\rightarrow
PastView^{(k)}
\rightarrow
Policy^{(k)}
\rightarrow
Future^{(k+1)}.
} F u t u r e ( k ) → P a s t V i e w ( k ) → P o l i c y ( k ) → F u t u r e ( k + 1 ) .
49. Deliberation Loop 可以很短
例如:
K = 1. K=1. K = 1.
就是一次:
預測 → 選 action。
50. 也可以多輪
K > 1. K>1. K > 1.
例如:
提方案 → 找歷史反例 → 改方案 → 再模擬。
51. K K K 必須 Bounded
真實 agent:
K < ∞ \boxed{
K<\infty
} K < ∞
受:
latency;
compute;
energy;
deadline;
限制。
52. 不要求 Fixed-Point Convergence
π ( k + 1 ) → π ⋆ \boxed{
\pi^{(k+1)}
\rightarrow
\pi^\star
} π ( k + 1 ) → π ⋆
不保證。
53. Deliberation 可 Oscillate
例如:
f 1 → h 1 → f 2 → h 2 → f 1 . f_1
\rightarrow
h_1
\rightarrow
f_2
\rightarrow
h_2
\rightarrow
f_1. f 1 → h 1 → f 2 → h 2 → f 1 .
54. Oscillation 不是自動 Bug
若 evidence 真 ambiguous,
policy indecision 可能合理。
55. 但需要 Stopping Contract
例如:
S t o p = B u d g e t E x h a u s t e d ∨ P o l i c y S t a b l e ∨ R i s k T h r e s h o l d ∨ D e a d l i n e . \boxed{
Stop
=
BudgetExhausted
\lor
PolicyStable
\lor
RiskThreshold
\lor
Deadline.
} S t o p = B u d g e tE x ha u s t e d ∨ P o l i cy S t ab l e ∨ R i s k T h r es h o l d ∨ D e a d l in e .
56. Policy Stability
d π ( π ( k + 1 ) , π ( k ) ) < ϵ π . \boxed{
d_\pi(
\pi^{(k+1)},
\pi^{(k)}
)
<
\epsilon_\pi.
} d π ( π ( k + 1 ) , π ( k ) ) < ϵ π .
57. Future-Space Stability
d F ( B + , ( k + 1 ) , B + , ( k ) ) < ϵ F . \boxed{
d_F(
\mathcal B^{+,(k+1)},
\mathcal B^{+,(k)}
)
<
\epsilon_F.
} d F ( B + , ( k + 1 ) , B + , ( k ) ) < ϵ F .
58. Historical-View Stability
d H ( w − , ( k + 1 ) , w − , ( k ) ) < ϵ H . \boxed{
d_H(
\mathbf w^{-,(k+1)},
\mathbf w^{-,(k)}
)
<
\epsilon_H.
} d H ( w − , ( k + 1 ) , w − , ( k ) ) < ϵ H .
59. Bounded Cognitive Closure
可定義:
C t ( K ) = Delib K ( T t ( 3 ) ) . \boxed{
\mathfrak C_t^{(K)}
=
\operatorname{Delib}^{K}
(
\mathfrak T_t^{(3)}
).
} C t ( K ) = Delib K ( T t ( 3 ) ) .
60. Cognitive Closure 不是 Logical Completeness
它只表示:
在 bounded deliberation budget 下完成的一輪 internal refinement。
61. Action Commitment
deliberation 結束:
a t ∼ π t ( K ) . \boxed{
a_t
\sim
\pi_t^{(K)}.
} a t ∼ π t ( K ) .
62. Environment Transition
Z t + 1 = T Z ( Z t , a t , ε t ) . \boxed{
Z_{t+1}
=
T_Z(
Z_t,
a_t,
\varepsilon_t
).
} Z t + 1 = T Z ( Z t , a t , ε t ) .
63. ε t \varepsilon_t ε t 包含
noise;
exogenous event;
other-agent action;
model mismatch;
stochasticity。
64. Actual Outcome May Differ from Predicted Future
Z t + 1 r e a l ≠ Z ^ t + 1 . \boxed{
Z_{t+1}^{real}
\neq
\widehat Z_{t+1}.
} Z t + 1 r e a l = Z t + 1 .
65. Prediction Error Becomes New Information
定義:
δ t = d ( Z t + 1 r e a l , Z ^ t + 1 ) . \boxed{
\delta_t
=
d(
Z_{t+1}^{real},
\widehat Z_{t+1}
).
} δ t = d ( Z t + 1 r e a l , Z t + 1 ) .
66. Sedimentation
TCD-05:
B t + 1 − = H − ( B t − , B t 0 , B t + , a t , Z t + 1 , E t ) . \boxed{
\mathcal B_{t+1}^-
=
\mathcal H_-
(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+,
a_t,
Z_{t+1},
E_t
).
} B t + 1 − = H − ( B t − , B t 0 , B t + , a t , Z t + 1 , E t ) .
67. New Past Includes
what happened;
what was chosen;
what was unchosen;
which future influenced decision;
which dependencies changed;
which prediction failed。
68. New Present
新 observation:
Ξ t + 1 \Xi_{t+1} Ξ t + 1
到來後:
B t + 1 0 = Φ − 0 ( B t + 1 − , Ξ t + 1 ) . \boxed{
\mathcal B_{t+1}^0
=
\Phi_{-0}
(
\mathcal B_{t+1}^-,
\Xi_{t+1}
).
} B t + 1 0 = Φ − 0 ( B t + 1 − , Ξ t + 1 ) .
69. New Future
B t + 1 + = Γ F ( B t + 1 − , B t + 1 0 , Ξ t + 1 ) . \boxed{
\mathcal B_{t+1}^+
=
\Gamma_F
(
\mathcal B_{t+1}^-,
\mathcal B_{t+1}^0,
\Xi_{t+1}
).
} B t + 1 + = Γ F ( B t + 1 − , B t + 1 0 , Ξ t + 1 ) .
70. Full Shift
因此:
S t : ( B t − , B t 0 , B t + ) → ( B t + 1 − , B t + 1 0 , B t + 1 + ) . \boxed{
\mathscr S_t
:
(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+
)
\rightarrow
(
\mathcal B_{t+1}^-,
\mathcal B_{t+1}^0,
\mathcal B_{t+1}^+
).
} S t : ( B t − , B t 0 , B t + ) → ( B t + 1 − , B t + 1 0 , B t + 1 + ) .
71. Extended Shift
更完整:
S t = S ( T t ( 3 ) , Ξ t , ε t , B t , κ t ) . \boxed{
\mathscr S_t
=
\mathscr S
(
\mathfrak T_t^{(3)},
\Xi_t,
\varepsilon_t,
\mathbf B_t,
\kappa_t
).
} S t = S ( T t ( 3 ) , Ξ t , ε t , B t , κ t ) .
72. TCD Master Update
本文將第一版 master update 壓縮為:
B ~ t 0 = Φ − 0 ( B t − , Ξ t ) , B t + , ( 0 ) = Γ F ( B t − , B ~ t 0 , Ξ t ) , ( w t − , ( k ) , π t ( k ) , B t + , ( k + 1 ) ) = D ( B t − , B ~ t 0 , B t + , ( k ) ) , a t ∼ π t ( K ) , Z t + 1 = T Z ( Z t , a t , ε t ) , B t + 1 − = H − ( B t − , B t 0 , B t + , a t , Z t + 1 ) , B t + 1 0 = Φ − 0 ( B t + 1 − , Ξ t + 1 ) , B t + 1 + = Γ F ( B t + 1 − , B t + 1 0 , Ξ t + 1 ) . \boxed{
\begin{aligned}
\widetilde B_t^0
&=
\Phi_{-0}(B_t^-,\Xi_t),\\
B_t^{+,(0)}
&=
\Gamma_F(B_t^-,\widetilde B_t^0,\Xi_t),\\
(w_t^{-,(k)},\pi_t^{(k)},B_t^{+,(k+1)})
&=
\mathfrak D(
B_t^-,
\widetilde B_t^0,
B_t^{+,(k)}
),\\
a_t
&\sim
\pi_t^{(K)},\\
Z_{t+1}
&=
T_Z(Z_t,a_t,\varepsilon_t),\\
B_{t+1}^-
&=
\mathcal H_-(B_t^-,B_t^0,B_t^+,a_t,Z_{t+1}),\\
B_{t+1}^0
&=
\Phi_{-0}(B_{t+1}^-,\Xi_{t+1}),\\
B_{t+1}^+
&=
\Gamma_F(B_{t+1}^-,B_{t+1}^0,\Xi_{t+1}).
\end{aligned}
} B t 0 B t + , ( 0 ) ( w t − , ( k ) , π t ( k ) , B t + , ( k + 1 ) ) a t Z t + 1 B t + 1 − B t + 1 0 B t + 1 + = Φ − 0 ( B t − , Ξ t ) , = Γ F ( B t − , B t 0 , Ξ t ) , = D ( B t − , B t 0 , B t + , ( k ) ) , ∼ π t ( K ) , = T Z ( Z t , a t , ε t ) , = H − ( B t − , B t 0 , B t + , a t , Z t + 1 ) , = Φ − 0 ( B t + 1 − , Ξ t + 1 ) , = Γ F ( B t + 1 − , B t + 1 0 , Ξ t + 1 ) .
73. D \mathfrak D D 是 Deliberative Coupling Block
它包含:
R + − + Φ + 0 + Γ F r e f i n e . \boxed{
\mathcal R_{+-}
+
\Phi_{+0}
+
\Gamma_F^{refine}.
} R +− + Φ + 0 + Γ F r e f in e .
仍然不是普通 algebraic sum。
74. Why Deliberative Block Matters
因為 agent 在 action 前可以:
生成 future;
用 future 找 past;
用 past 修正 future;
用 future 重新排序 action。
75. 這就是 Reflexive Cognition
不是:
one-pass inference . \boxed{
\text{one-pass inference}.
} one-pass inference .
而是:
bounded recurrent inference . \boxed{
\text{bounded recurrent inference}.
} bounded recurrent inference .
76. TCD 與普通 Markov Policy 的差別
普通形式:
a t ∼ π ( a ∣ s t ) . \boxed{
a_t
\sim
\pi(a\mid s_t).
} a t ∼ π ( a ∣ s t ) .
77. TCD 形式
更像:
a t ∼ π ( a ∣ B t − , B t 0 , B t + ) . \boxed{
a_t
\sim
\pi
\left(
a
\mid
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+
\right).
} a t ∼ π ( a ∣ B t − , B t 0 , B t + ) .
78. 但如果 s t s_t s t 已充分包含三者呢?
那:
s t = C o m p r e s s ( B − , B 0 , B + ) \boxed{
s_t
=
Compress(
B^-,
B^0,
B^+
)
} s t = C o m p r ess ( B − , B 0 , B + )
完全可以。
79. TCD 不反對 Markovization
如果能找到 sufficient augmented state:
S t ⋆ , S_t^\star, S t ⋆ ,
TCD 可被壓縮進:
S t ⋆ . S_t^\star. S t ⋆ .
80. TCD 真正關心的是:壓縮前你有沒有漏掉功能
所以:
TCD ≠ anti-Markov . \boxed{
\text{TCD}
\neq
\text{anti-Markov}.
} TCD = anti-Markov .
81. Temporal Sufficiency
定義:
S t ⋆ = Ψ T C D ( B t − , B t 0 , B t + ) . \boxed{
S_t^\star
=
\Psi_{TCD}
(
B_t^-,
B_t^0,
B_t^+
).
} S t ⋆ = Ψ T C D ( B t − , B t 0 , B t + ) .
82. 若 S t ⋆ S_t^\star S t ⋆ 足夠
使:
P ( Y t : H ∣ B − , B 0 , B + ) ≈ P ( Y t : H ∣ S t ⋆ ) , P(
Y_{t:H}
\mid
B^-,
B^0,
B^+
)
\approx
P(
Y_{t:H}
\mid
S_t^\star
), P ( Y t : H ∣ B − , B 0 , B + ) ≈ P ( Y t : H ∣ S t ⋆ ) ,
則可使用 compressed state。
83. 這延續 HSV
不是所有 history 都要永久展開。
84. Coupling Strength
如何量一條 edge 的重要性?
用 ablation。
85. Edge-Ablation
對 edge:
e i j , e_{ij}, e ij ,
建立:
G T C D − e i j . \boxed{
G_{\mathrm{TCD}}^{-e_{ij}}.
} G TCD − e ij .
86. Coupling Sensitivity
對 task loss:
L T , \mathcal L_T, L T ,
定義:
χ i j ( T ) = L T ( G − e i j ) − L T ( G ) . \boxed{
\chi_{ij}^{(T)}
=
\mathcal L_T(
G^{-e_{ij}}
)
-
\mathcal L_T(
G
).
} χ ij ( T ) = L T ( G − e ij ) − L T ( G ) .
87. χ > 0 \chi>0 χ > 0
表示移除 coupling 使 performance 變差。
88. χ ≈ 0 \chi\approx0 χ ≈ 0
表示在該 task:
這條 edge 可能不重要。
89. χ < 0 \chi<0 χ < 0
表示移除 edge 反而更好。
這非常重要。
90. More Coupling Is Not Always Better
more temporal coupling ⇏ better cognition . \boxed{
\text{more temporal coupling}
\not\Rightarrow
\text{better cognition}.
} more temporal coupling ⇒ better cognition .
91. Example:Bad Future Representation
如果:
B t + B_t^+ B t +
充滿 phantom futures,
強:
+ → 0 +\to0 + → 0
可能傷害 policy。
92. Example:Bad Retrospective Relevance
若:
+ → − +\to- + → −
只找支持 future 的 history,
形成 narrative capture。
93. Example:Excess Sedimentation
若:
0 → − 0\to- 0 → −
保存太多 old constraints,
plasticity 崩潰。
94. TCD 因此需要 Coupling Governance
每條 edge 都要有:
budget;
uncertainty;
evidence;
stopping;
audit。
95. Coupling Profile
對 agent:
Π T C D A = ( Q − 0 , Q 0 + , Q + 0 , Q 0 − , Q − + , Q + − ) . \boxed{
\Pi_{\mathrm{TCD}}^A
=
(
Q_{-0},
Q_{0+},
Q_{+0},
Q_{0-},
Q_{-+},
Q_{+-}
).
} Π TCD A = ( Q − 0 , Q 0 + , Q + 0 , Q 0 − , Q −+ , Q +− ) .
96. 不是 Scalar
不要:
T C D I Q = 95. \boxed{
TCDIQ=95.
} T C D I Q = 95.
97. Agent A 可能 History 強
Q − 0 ≫ 0 , Q_{-0}\gg0, Q − 0 ≫ 0 ,
但:
Q 0 + Q_{0+} Q 0 +
弱。
98. Agent B 可能 Future Generation 強
但 Sedimentation 差,
反覆重犯。
99. Agent C 可能 Prospective Attraction 太強
容易被 phantom future 操縱。
100. Agent D 可能 Retrospective Relevance 太弱
無法從新目標重新啟用 old knowledge。
101. Six-Edge Bottleneck
定義:
e ⋆ = arg max e χ e ( T ) . \boxed{
e^\star
=
\arg\max_e
\chi_e^{(T)}.
} e ⋆ = arg e max χ e ( T ) .
表示在該 task 移除後造成最大性能損失的 temporal coupling。
102. Coupling Bottleneck Can Migrate
隨 agent 改進:
e t ⋆ ≠ e t + 1 ⋆ . \boxed{
e_t^\star
\neq
e_{t+1}^\star.
} e t ⋆ = e t + 1 ⋆ .
103. Positive Feedback Loop I — Narrative Capture
f → s u p p o r t i v e p a s t → P ( f ) ↑ → m o r e s u p p o r t s e a r c h . \boxed{
f
\rightarrow
supportive\ past
\rightarrow
P(f)\uparrow
\rightarrow
more\ support\ search.
} f → s u pp or t i v e p a s t → P ( f ) ↑→ m or e s u pp or t se a r c h .
104. 這可能形成 Self-Reinforcing Belief
不是因為 evidence 真變強,
而是 retrieval bias。
105. Negative Feedback Loop I — Adversarial Correction
f → c o u n t e r h i s t o r y → P ( f ) ↓ → f u t u r e r e v i s i o n . \boxed{
f
\rightarrow
counterhistory
\rightarrow
P(f)\downarrow
\rightarrow
future\ revision.
} f → co u n t er hi s t or y → P ( f ) ↓→ f u t u r e r e v i s i o n .
106. 這是 Desired Stabilization
TCD 不只研究正 feedback。
107. Positive Feedback Loop II — Constructive Realization
f → a c t i o n → i n f r a s t r u c t u r e → P ( f ) ↑ . \boxed{
f
\rightarrow
action
\rightarrow
infrastructure
\rightarrow
P(f)\uparrow.
} f → a c t i o n → in f r a s t r u c t u r e → P ( f ) ↑ .
108. Negative Feedback Loop II — Prevention
f b a d → m i t i g a t i o n → P ( f b a d ) ↓ . \boxed{
f_{bad}
\rightarrow
mitigation
\rightarrow
P(f_{bad})\downarrow.
} f ba d → mi t i g a t i o n → P ( f ba d ) ↓ .
109. Feedback Sign Depends on Mechanism
所以:
C y c l e ≠ S e l f R e i n f o r c e m e n t . \boxed{
Cycle
\neq
SelfReinforcement.
} C y c l e = S e l f R e in f or ce m e n t .
110. Loop Gain
概念上可定義:
G l o o p = ∏ e ∈ c y c l e g e . \boxed{
G_{loop}
=
\prod_{e\in cycle}
g_e.
} G l oo p = e ∈ cy c l e ∏ g e .
但只在 edge gains 有可比尺度時使用。
111. 一般 TCD 不假設 Linear Stability Theory 直接適用
如果 operator nonlinear / typed,
應使用:
simulation;
ablation;
local linearization;
empirical dynamics。
112. Temporal Coherence
一個 mature agent 的三個 base spaces 不應互相完全矛盾。
113. Example
Present says:
resource = 0 , \text{resource}=0, resource = 0 ,
Future says:
tomorrow deploy giant system。
若沒有 path:
future-present incoherence . \boxed{
\text{future-present incoherence}.
} future-present incoherence .
114. Past-Future Coherence
Future claim 若依賴已被 Past falsified 的 dependency,
也有:
historical-future incoherence . \boxed{
\text{historical-future incoherence}.
} historical-future incoherence .
115. Coherence Checks
可定義:
C − 0 , C 0 + , C − + . \boxed{
C_{-0},
C_{0+},
C_{-+}.
} C − 0 , C 0 + , C −+ .
116. Coherence 不等於 Conformism
new future 可以挑戰 past trend。
只要:
有 explicit mechanism。
117. Surprise Is Allowed
TCD Future 仍保留:
U t + . U_t^+. U t + .
118. Unknown Mass Prevents Temporal Overclosure
如果:
p ⊥ = 0 p_\bot=0 p ⊥ = 0
被濫用,
agent 容易:
把目前三生模型當宇宙全部可能性。
119. TCD Unknown Triple
可寫:
U t = ( U t − , U t 0 , U t + ) . \boxed{
\mathbf U_t
=
(
U_t^-,
U_t^0,
U_t^+
).
} U t = ( U t − , U t 0 , U t + ) .
120. Past Unknown
無法完整重建 historical alternatives。
121. Present Unknown
不知道 hidden capability / path / dependency。
122. Future Unknown
不知道未命名 future regions。
123. Temporal Epistemic Humility
所以:
T ^ t ( 3 ) ≠ T t ( 3 ) , t r u e \boxed{
\widehat{\mathfrak T}_t^{(3)}
\neq
\mathfrak T_t^{(3),true}
} T t ( 3 ) = T t ( 3 ) , t r u e
一般應保留。
124. Observation of TCD Is Observer-Relative
不同 observer:
A , B A,B A , B
可能建出:
T ^ t ( 3 ) , A ≠ T ^ t ( 3 ) , B . \widehat{\mathfrak T}_t^{(3),A}
\neq
\widehat{\mathfrak T}_t^{(3),B}. T t ( 3 ) , A = T t ( 3 ) , B .
125. Observer Difference 不是 Arbitrary Subjectivism
仍受:
evidence;
provenance;
dynamics;
experiment;
resolution;
約束。
126. Multi-Agent TCD
對 agents:
A 1 , … , A n , A_1,\ldots,A_n, A 1 , … , A n ,
每個有:
T t ( 3 ) , i . \mathfrak T_t^{(3),i}. T t ( 3 ) , i .
127. Shared Present
它們可能共享:
E n v t , Env_t, E n v t ,
但:
B t + , i B_t^{+,i} B t + , i
不同。
128. Shared Future Candidate Can Couple Agents
若:
f f f
被多人採用,
可:
f → a t 1 , … , a t n . \boxed{
f
\rightarrow
a_t^1,\ldots,a_t^n.
} f → a t 1 , … , a t n .
129. Collective Prospective Attraction
未來 narrative 可以形成 collective coordination。
130. 也可能形成 Collective Phantom Future
錯誤 future narrative 也能同時改變大量 agents。
131. Collective Sedimentation
multi-agent actions 產生:
standards;
institutions;
markets;
infrastructure。
132. 這為 Branching World / Institution TCD 留接口
本文不展開。
133. MuZero 的外部對照
MuZero 建立 planning-relevant learned dynamics,
用 tree search 評估 action sequences,
再選 current action。
外部意義:
future model → present action . \boxed{
\text{future model}
\rightarrow
\text{present action}.
} future model → present action .
134. DreamerV3 的外部對照
DreamerV3 的 world model:
預測 potential action outcomes;
critic:
評估 imagined outcomes;
actor:
學習導向高 value outcome 的 action。
這是非常清楚的:
F u t u r e R e p r e s e n t a t i o n → V a l u a t i o n → P o l i c y . \boxed{
FutureRepresentation
\rightarrow
Valuation
\rightarrow
Policy.
} F u t u r e R e p r ese n t a t i o n → V a l u a t i o n → P o l i cy .
135. Mattar–Daw 的外部對照
其 prioritized memory access theory:
依 memory 對 future decision improvement 的 utility 決定 replay priority。
對照:
F u t u r e N e e d → P a s t A c c e s s . \boxed{
FutureNeed
\rightarrow
PastAccess.
} F u t u r e N ee d → P a s t A ccess .
136. Momennejad 等人的外部對照
reward revaluation 後:
distal past states 的 offline replay 與後續 replanning 相關。
對照:
N e w I n f o r m a t i o n → P a s t R e p l a y → N e w P o l i c y . \boxed{
NewInformation
\rightarrow
PastReplay
\rightarrow
NewPolicy.
} N e w I n f or ma t i o n → P a s tR e pl a y → N e w P o l i cy .
137. 2025 Prospective Navigation Codes 的外部對照
一項大鼠 flexible navigation 研究顯示:
新 reward-location information 到來後,hippocampal/prefrontal prospective representations of future choices 可以快速調整,並伴隨 route switching。
138. TCD 的最保守讀法
這只支持:
past integration + prospective representation + current choice update \boxed{
\text{past integration}
+
\text{prospective representation}
+
\text{current choice update}
} past integration + prospective representation + current choice update
在至少某些 biological decision tasks 中具有經驗相鄰性。
139. 不宣稱 Rat Brain = TCD Runtime
biological analogue ≠ theory identity . \boxed{
\text{biological analogue}
\neq
\text{theory identity}.
} biological analogue = theory identity .
140. Six-Edge Ablation Benchmark
建立 finite environment。
141. Full Agent
有六條 coupling。
142. Ablation 1
移除:
− → 0. -\to0. − → 0.
143. Ablation 2
移除:
0 → + . 0\to+. 0 → + .
144. Ablation 3
移除:
+ → 0. +\to0. + → 0.
145. Ablation 4
移除:
0 → − . 0\to-. 0 → − .
146. Ablation 5
移除:
− → + . -\to+. − → + .
147. Ablation 6
移除:
+ → − . +\to-. + → − .
148. Measure
比較:
cumulative task return;
safe reachability;
future coverage;
adaptation;
repeated-error rate;
option preservation;
provenance fidelity;
compute cost。
149. Exact Finite Micro-Model
可用 dynamic graph:
G t = ( V t , E t ) . G_t=(V_t,E_t). G t = ( V t , E t ) .
150. Past Choice Alters Graph
action:
a t a_t a t
可以:
add edge;
remove edge;
increase edge cost。
151. Present Domain
agent 只能看到:
V t v i s V_t^{vis} V t v i s
與:
E t v i s . E_t^{vis}. E t v i s .
152. Future Generator
生成 candidate goals:
f 1 , … , f n . f_1,\ldots,f_n. f 1 , … , f n .
153. Prospective Attraction
future value 改 path selection。
154. Sedimentation
action 改 graph + log。
155. Retrospective Relevance
future goal 改 past transition retrieval。
156. 這可以完整執行六條 Edge
而且:
finite \boxed{
\text{finite}
} finite
可 exact audit。
157. TCD Runtime Trace
每輪至少輸出:
t
past_base_version
present_base_version
future_base_version
deliberation_rounds
active_past_items
future_candidates
future_unknown_mass
policy_before_future
policy_after_future
chosen_action
predicted_outcomes
real_outcome
prediction_error
sedimentation_record
dependencies_changed
options_lost
options_opened
next_state_versions
158. Replayability
理想:
T r a c e 0 : t → T ^ t ( 3 ) . \boxed{
Trace_{0:t}
\rightarrow
\widehat{\mathfrak T}_t^{(3)}.
} T r a c e 0 : t → T t ( 3 ) .
159. Replay 不必 Deterministic
若 system stochastic,
保存:
seeds;
distributions;
model versions。
160. Temporal Provenance
每一個:
B − , B 0 , B + B^-,
B^0,B^+ B − , B 0 , B +
都應有:
timestamp;
version;
parent;
evidence;
operator;
cost。
161. Six-Way Coupling Evidence Passport
對每條 edge claim:
edge
agent
task
domain
source_state
target_state
operator
budget
ablation
effect_size
uncertainty
evidence_level
failure_conditions
162. Edge Claim 不能越界
例如:
DreamerV3 有 + → 0 +\to0 + → 0 engineering analogue。
不能推出:
DreamerV3 已具完整 TCD。
163. Temporal Shift Operator
本文正式命名:
TCD Shift Operator
S t . \boxed{
\mathscr S_t.
} S t .
164. Minimal Definition
S t : T t ( 3 ) × Ξ t × E t → T t + 1 ( 3 ) . \boxed{
\mathscr S_t
:
\mathfrak T_t^{(3)}
\times
\Xi_t
\times
\mathcal E_t
\rightarrow
\mathfrak T_{t+1}^{(3)}.
} S t : T t ( 3 ) × Ξ t × E t → T t + 1 ( 3 ) .
165. E t \mathcal E_t E t 表示
actual action;
environment event;
observed consequence;
new evidence。
166. Shift Operator 不是 Closed Autonomous Law
真實 agent 會接收:
Ξ t . \boxed{
\Xi_t.
} Ξ t .
167. 所以 TCD 是 Open Dynamical System
TCD agent + environment + other agents . \boxed{
\text{TCD agent}
+
\text{environment}
+
\text{other agents}.
} TCD agent + environment + other agents .
168. No Closed-World Assumption
future unknown:
U t + U_t^+ U t +
永遠可能非零。
169. Temporal Fixed Point
若:
T t + 1 ( 3 ) ≈ T t ( 3 ) , \mathfrak T_{t+1}^{(3)}
\approx
\mathfrak T_t^{(3)}, T t + 1 ( 3 ) ≈ T t ( 3 ) ,
可稱:
local temporal cognitive fixed regime . \boxed{
\text{local temporal cognitive fixed regime}.
} local temporal cognitive fixed regime .
170. 不一定是好事
可能是:
stable expertise;
rigid institution;
stuck loop;
mature policy。
171. Temporal Cycle
若:
T t + k ( 3 ) ≈ T t ( 3 ) , \mathfrak T_{t+k}^{(3)}
\approx
\mathfrak T_t^{(3)}, T t + k ( 3 ) ≈ T t ( 3 ) ,
可研究 periodic cognitive regime。
172. Temporal Drift
若 state 持續移動但無突然轉變:
drift . \boxed{
\text{drift}.
} drift .
173. Temporal Phase Shift
若:
future ontology;
capability;
reachable domain;
historical relevance;
突然重構,
可標:
temporal cognitive phase shift . \boxed{
\text{temporal cognitive phase shift}.
} temporal cognitive phase shift .
174. 本文不正式建立 Phase Theory
只留接口。
175. Learning Is Temporal-State Change
如果:
C a p t + 1 ≠ C a p t , Cap_{t+1}\neq Cap_t, C a p t + 1 = C a p t ,
這是:
T t + 1 ( 3 ) ≠ T t ( 3 ) . \boxed{
\mathfrak T_{t+1}^{(3)}
\neq
\mathfrak T_t^{(3)}.
} T t + 1 ( 3 ) = T t ( 3 ) .
176. Forgetting Is Temporal-State Change
如果:
B t + 1 − B_{t+1}^- B t + 1 −
壓縮/失活某些 residue,
也會改:
B t + 1 0 , B t + 1 + . B_{t+1}^0,
B_{t+1}^+. B t + 1 0 , B t + 1 + .
177. Prediction Error Can Trigger Temporal Reorganization
若:
δ t ≫ 0 , \delta_t\gg0, δ t ≫ 0 ,
agent 可能:
reweight past;
revise world model;
expand unknown mass;
generate new future ontology。
178. Error-Driven TCD Update
概念:
δ t → ( Δ B − , Δ B 0 , Δ B + ) . \boxed{
\delta_t
\rightarrow
(
\Delta B^-,
\Delta B^0,
\Delta B^+
).
} δ t → ( Δ B − , Δ B 0 , Δ B + ) .
179. Surprise Does Not Automatically Mean Learning
若 agent 不 sediment / update:
δ t \delta_t δ t
可能被忽略。
180. Long-Term Intelligence Requires Temporal Credit Assignment
某 action:
a t a_t a t
結果:
t + 100 t+100 t + 100
才顯現。
181. Credit Assignment Across TCD
需要追:
R t + → a t → S e d i m e n t → F u t u r e O u t c o m e . \boxed{
R_t^+
\rightarrow
a_t
\rightarrow
Sediment
\rightarrow
FutureOutcome.
} R t + → a t → S e d im e n t → F u t u r e O u t co m e .
182. 這是 PCI 長期 Resolution 的 runtime 版本
PCI 保存:
prediction;
realization path;
future resolution。
TCD 現在保存它們之間的 temporal lineage。
183. TCD 與 UCPNP 的關係
UCPNP 問:
哪些 intervention 能改變 agent-relative tractability frontier?
184. TCD 回答其中一個動力來源
同一 intervention:
新 knowledge;
new tool;
new future;
reactivated past;
會經 TCD couplings 改變:
B 0 \boxed{
\mathcal B^0
} B 0
與:
B + . \boxed{
\mathcal B^+.
} B + .
185. UCPNP 是 Frontier Program
TCD 是:
one temporal dynamics layer beneath frontier motion . \boxed{
\text{one temporal dynamics layer beneath frontier motion}.
} one temporal dynamics layer beneath frontier motion .
186. TCD 不能取代 UCPNP
它不直接定義:
verification;
certification;
completion cost;
complexity regime。
187. 兩者接口
U t ⊃ T t ( 3 ) \boxed{
\mathfrak U_t
\supset
\mathfrak T_t^{(3)}
} U t ⊃ T t ( 3 )
可作未來整合方向。
188. TCD v0.1 Core
本文建議將:
TCD-01~TCD-07
視為:
TCD v0.1 Core . \boxed{
\text{TCD v0.1 Core}.
} TCD v0.1 Core .
189. v0.1 包含
Past Base Space;
Present Base Space;
Future Base Space;
Prospective Attraction;
Historical Sedimentation;
Retrospective Relevance;
Six-Way Coupling + Shift Operator。
190. v0.1 不包含
runnable world branching;
world-domain governance;
cross-world evidence;
multi-world allocation;
persistent subworld agents;
full normative deployment layer。
191. 為什麼下一步應該分新系列?
因為:
B t + \boxed{
\mathcal B_t^+
} B t +
目前仍是:
cognitive / modelled future domain。
192. 下一階
若把 candidate:
f i f_i f i
instantiate:
f i → W i , \boxed{
f_i
\rightarrow
W_i,
} f i → W i ,
世界:
W i W_i W i
開始:
run;
accumulate history;
contain agents;
receive interventions;
就不是單純 TCD representation。
193. 這是 Runnable Future
Represented Future → Executable Future World . \boxed{
\text{Represented Future}
\rightarrow
\text{Executable Future World}.
} Represented Future → Executable Future World .
194. 新問題
不再只是:
哪些 futures 值得想?
而是:
哪些 futures 值得實際投入計算? \boxed{
\text{哪些 futures 值得實際投入計算?}
} 哪些 futures 值得實際投入計算?
195. 這就是下一系列
Branching World Computation
分支世界計算/世界域認知 Runtime
196. 可否證條件
F196.1 Edge-Indistinguishability
若六條 coupling 在 controlled ablation 中無法操作性區分,
taxonomy 應簡化。
F196.2 Deliberation No-Gain
若 K > 1 K>1 K > 1 的 reflective loop 長期不比 K = 1 K=1 K = 1 好,
bounded recursion 不應被神化。
F196.3 Circularity Failure
若 implementation 無法區分 deliberation index k k k 與 historical time t t t ,
runtime formalization 需要重構。
F196.4 Past-Fact Contamination
若 + → − +\to- + → − 會 silently rewrite provenance,
統一系統失效。
F196.5 Phantom-Future Instability
若錯誤 future representation 經 + → 0 +\to0 + → 0 造成大規模 harmful lock-in,
需要降低 coupling gain / 增加 adversarial gate。
F196.6 Over-Sedimentation
若 0 → − 0\to- 0 → − 造成 memory rigidity / plasticity collapse,
需增加 forgetting / archive。
F196.7 Historical Confirmation Loop
若 + → − → + +\to-\to+ + → − → + 只強化同一 narrative,
需強制 counter-retrieval。
F196.8 External-Shock Failure
若 TCD 忽略 Ξ t , ε t \Xi_t,\varepsilon_t Ξ t , ε t 後只能解釋封閉世界,
必須保留 open-system formulation。
F196.9 No Predictive Gain
若完整 TCD state 對 planning / adaptation / audit 完全不優於更簡單 sufficient state,
應使用簡單模型。
197. 結論
TCD-01 說:
P a s t ≠ M e m o r y . \boxed{
Past\neq Memory.
} P a s t = M e m or y .
TCD-02 說:
P r e s e n t ≠ P o i n t . \boxed{
Present\neq Point.
} P r ese n t = P o in t .
TCD-03 說:
F u t u r e ≠ P r e G i v e n M a p . \boxed{
Future\neq PreGivenMap.
} F u t u r e = P r e G i v e n M a p .
TCD-04 說:
F u t u r e R e p r e s e n t a t i o n → P r e s e n t P o l i c y . \boxed{
FutureRepresentation
\rightarrow
PresentPolicy.
} F u t u r e R e p r ese n t a t i o n → P r ese n tP o l i cy .
TCD-05 說:
P r e s e n t A c t i o n → P a s t S e d i m e n t . \boxed{
PresentAction
\rightarrow
PastSediment.
} P r ese n t A c t i o n → P a s tS e d im e n t .
TCD-06 說:
F u t u r e T a r g e t → P a s t R e l e v a n c e . \boxed{
FutureTarget
\rightarrow
PastRelevance.
} F u t u r e T a r g e t → P a s tR e l e v an ce .
TCD-07 現在把它們全部放進同一個系統。
真正的核心不再是:
P a s t , P r e s e n t , F u t u r e . \boxed{
Past,
Present,
Future.
} P a s t , P r ese n t , F u t u r e .
而是:
P a s t ⇆ P r e s e n t ⇆ F u t u r e ⇆ P a s t \boxed{
Past
\leftrightarrows
Present
\leftrightarrows
Future
\leftrightarrows
Past
} P a s t ⇆ P r ese n t ⇆ F u t u r e ⇆ P a s t
但每一條箭頭都有不同 operator semantics。
在 agent 的同一 decision moment:
t t t
內,
可以有:
F u t u r e ( k ) → P a s t V i e w ( k ) → P o l i c y ( k ) → F u t u r e ( k + 1 ) . \boxed{
Future^{(k)}
\rightarrow
PastView^{(k)}
\rightarrow
Policy^{(k)}
\rightarrow
Future^{(k+1)}.
} F u t u r e ( k ) → P a s t V i e w ( k ) → P o l i c y ( k ) → F u t u r e ( k + 1 ) .
這不是時間旅行。
它只是:
bounded deliberation . \boxed{
\text{bounded deliberation}.
} bounded deliberation .
當 agent 真正執行:
a t , a_t, a t ,
世界才:
t → t + 1. \boxed{
t\rightarrow t+1.
} t → t + 1.
於是:
P r e s e n t A c t i o n t → P a s t t + 1 . \boxed{
PresentAction_t
\rightarrow
Past_{t+1}.
} P r ese n t A c t i o n t → P a s t t + 1 .
新的 Past 又重新構造:
P r e s e n t t + 1 , Present_{t+1}, P r ese n t t + 1 ,
新的 Present 再與新的 Past 一起生成:
F u t u r e t + 1 . Future_{t+1}. F u t u r e t + 1 .
所以第一版完整 TCD 可以壓縮成一句話:
智能不只是在現在解題;它利用沉積的過去生成可行的現在與想像的未來,再讓被生成的未來重新組織現在的行動與過去的相關性,而行動又被沉積成下一輪歷史。
形式上:
S t : ( B t − , B t 0 , B t + ) ⟶ ( B t + 1 − , B t + 1 0 , B t + 1 + ) . \boxed{
\mathscr S_t:
(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+
)
\longrightarrow
(
\mathcal B_{t+1}^-,
\mathcal B_{t+1}^0,
\mathcal B_{t+1}^+
).
} S t : ( B t − , B t 0 , B t + ) ⟶ ( B t + 1 − , B t + 1 0 , B t + 1 + ) .
這就是:
Tri-Temporal Cognitive Dynamics v0.1 Core
第一次完整閉合。
下一步不再需要多畫一條時間箭頭。
真正的新問題是:
如果 Future Base Space 中的候選不只被想像,而可以被 instantiate 成可持續執行、具有自身 agent、history、rules 與 interventions 的子世界,智能體要如何管理大量並行世界?
那已經是另一個研究層:
Branching World Computation
Claim Typing
Claim
Type
Status
TCD 可表示為三個 typed temporal base spaces
D
Canonical synthesis
六個 coupling 方向具有不同 operator semantics
D
Canonical boundary
deliberation index k k k 應與 historical time t t t 分離
D / methodology
Core anti-circularity formalization
TCD Shift Operator 可作完整一步更新 scaffold
D
Proposed unified dynamics
coupling importance 可用 edge ablation 衡量
D / experiment design
Proposed operationalization
MuZero / Dreamer 提供 future-model-to-present-action engineering analogues
E
External calibration
prioritized / offline replay 提供 future-need-to-past-access analogues
E
External calibration
prospective navigation codes adapt with new information
E
External neuroscience evidence
完整六向 TCD 已被單一實驗系統證明
—
Not claimed
Future → \to → Past 是物理逆因果
—
Explicitly rejected
更多 coupling 一定更智能
—
Explicitly rejected
Evidence Ladder
本文目前主要位於:
L0 :typed six-way coupling + shift operator;
L1–L2 :finite six-edge ablation / deliberation benchmarks 可實作;
L3 :world-model planning、replay、revaluation、prospective-navigation studies 提供局部機制外部對照;
L4 :需要 persistent AI / robot / institution runtime 做 longitudinal six-edge ablation;
L5+ :跨 domain replication、multi-agent TCD、runnable-world extension 尚待後續。
參考文獻
Neo.K 內部正典與譜系
Neo.K. 歷史作為狀態變量:路徑依賴、記憶增廣與複雜系統的動力身份 . 2026.
Neo.K with Aletheia. Past Is Not Memory . TCD-01, 2026.
Neo.K with Aletheia. The Present Is Not a Point . TCD-02, 2026.
Neo.K with Aletheia. Future as a Generated Base Space . TCD-03, 2026.
Neo.K with Aletheia. Prospective Attraction . TCD-04, 2026.
Neo.K with Aletheia. Historical Sedimentation . TCD-05, 2026.
Neo.K with Aletheia. Retrospective Relevance . TCD-06, 2026.
Neo.K with Aletheia. Neo.K Ultimate Cognitive P/NP Unified Theory . UCPNP Paper 09, 2026.
Neo.K with Aletheia. Generative Forecasting . UCPNP Series II Paper 13, 2026.
Neo.K with Aletheia. Prospective Constructive Intelligence . UCPNP Series II Paper 14, 2026.
External technical calibration
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.
Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. Mastering Diverse Control Tasks through World Models . Nature 640, 647–653, 2025.
Mattar, M. G., & Daw, N. D. Prioritized Memory Access Explains Planning and Hippocampal Replay . Nature Neuroscience 21, 1609–1617, 2018.
Momennejad, I., Otto, A. R., Daw, N. D., & Norman, K. A. Offline Replay Supports Planning in Human Reinforcement Learning . eLife 7:e32548, 2018.
Prince, S. M., Cushing, S. D., Yassine, T. A., et al. New Information Triggers Prospective Codes to Adapt for Flexible Navigation . Nature Communications 16, 4822, 2025.
Schaul, T., Quan, J., Antonoglou, I., & Silver, D. Prioritized Experience Replay . arXiv:1511.05952, 2015.
Andrychowicz, M., Wolski, F., Ray, A., et al. Hindsight Experience Replay . Advances in Neural Information Processing Systems 30, 2017.
Public Version Disclaimer
本文是 agent-level cognitive / decision / temporal dynamics framework。
本文不聲稱:
六向 coupling 是標準認知科學定律;
Past、Present、Future 是三個物理本體層;
Future physically causes Past;
recurrent deliberation 是時間旅行;
六條 edge 必須存在於所有 agent;
coupling strength 可以跨 domain 直接比較;
world models、replay 或 hippocampal prospective coding 等同完整 TCD;
TCD 必然提高 performance;
TCD v0.1 已經完成 runnable-world / multi-world architecture;
本文對 classical P P P vs. N P NP N P 提供任何新證明。
本文真正建立的是:
T t ( 3 ) = ( B t − , B t 0 , B t + ) \boxed{
\mathfrak T_t^{(3)}
=
(
\mathcal B_t^-,
\mathcal B_t^0,
\mathcal B_t^+
)
} T t ( 3 ) = ( B t − , B t 0 , B t + )
與:
S t : T t ( 3 ) → T t + 1 ( 3 ) . \boxed{
\mathscr S_t:
\mathfrak T_t^{(3)}
\rightarrow
\mathfrak T_{t+1}^{(3)}.
} S t : T t ( 3 ) → T t + 1 ( 3 ) .
並明確要求:
six temporal couplings are typed, bounded, auditable, and empirically ablatable . \boxed{
\text{six temporal couplings are typed, bounded, auditable, and empirically ablatable}.
} six temporal couplings are typed, bounded, auditable, and empirically ablatable .