TADC-07:外部認知支架與人—AI 認知拓樸——有效距離、回返成本與混合認知系統
英文題名: External Cognitive Scaffolds and Human–AI Cognitive Topology: Effective Distance, Re-entry Cost, and Hybrid Cognitive Systems系列: Topological Attention and Dynamic Cognitive Domains — Conjecture Series(TADC)中文系列名: 拓樸注意力與動態認知域命題系列編號: TADC-07版本: v0.1日期: 2026-08-17作者: Neo.K(許筌崴)協作: GPT-5.6 Sol文件性質: 理論命題/人—AI 混合認知模型/可證偽研究綱領文獻檢索截點: 2026-08-17
摘要
人類長期使用紙筆、書籍、索引、提醒器、搜尋系統與其他人來降低內部記憶與推理負擔。認知科學通常將其中一部分描述為 cognitive offloading:透過外部行動或環境資源改變資訊處理需求。生成式 AI、長期記憶系統與 agent architecture 則進一步提供傳統外部記憶工具較少具備的功能:主動檢索、重新表示、跨資料源建立橋接、生成候選路徑、保持多個研究分支,以及在使用者未直接處理某支線時持續執行局部工作。
本文提出一個比「AI 提升工作效率」更強、但也更需要反證的命題:
外部認知支架可能改變一個人—工具混合系統的有效認知可達結構,而不只是提高某個既有認知節點上的處理速度。
本文定義人類內部認知結構:
C t H = ( X t H , R t H , κ t H , N t H , A t H , G t ) , \mathcal C_t^H
=
(
X_t^H,
\mathcal R_t^H,
\kappa_t^H,
\mathcal N_t^H,
A_t^H,
G_t
), C t H = ( X t H , R t H , κ t H , N t H , A t H , G t ) ,
外部支架:
S t E = ( M t , I t , Q t , P t ) , \mathcal S_t^E
=
(
M_t,
I_t,
Q_t,
P_t
), S t E = ( M t , I t , Q t , P t ) ,
以及 AI / Agent 層:
A t A I = ( X t A I , R t A I , Π t , V t ) . \mathcal A_t^{AI}
=
(
X_t^{AI},
R_t^{AI},
\Pi_t,
V_t
). A t A I = ( X t A I , R t A I , Π t , V t ) .
三者形成候選混合認知系統:
H t = C t H ⊕ S t E ⊕ A t A I . \boxed{
\mathcal H_t
=
\mathcal C_t^H
\oplus
\mathcal S_t^E
\oplus
\mathcal A_t^{AI}.
} H t = C t H ⊕ S t E ⊕ A t A I .
本文不以「extended mind」作為已成立前提,也不主張 AI 系統與人腦具有同等心理或主體地位;符號 ⊕ \oplus ⊕ 僅表示在特定任務中,人類行為的有效資訊處理路徑可以跨越內部與外部節點。
本文提出三個核心猜想:
Effective Cognitive Distance Transformation Conjecture(ECDTC) :外部支架可以改變任務中的有效認知距離;
Re-entry Topology Conjecture(RTC) :可尋址、可恢復的外部狀態可以降低中斷後回返舊分支的重建成本;
Hybrid Reachability Expansion Conjecture(HREC) :AI / agent support 可擴張混合系統在有限時間與資源下實際可到達的問題狀態集合。
定義內部距離:
d H ( x , y ∣ G ) , d_H(x,y\mid G), d H ( x , y ∣ G ) ,
以及混合系統有效距離:
d e f f ( x , y ∣ G , S E , A A I ) . d_{\mathrm{eff}}
(
x,y
\mid
G,\mathcal S^E,\mathcal A^{AI}
). d eff ( x , y ∣ G , S E , A A I ) .
若:
d e f f < d H , d_{\mathrm{eff}}
<
d_H, d eff < d H ,
則定義 cognitive-distance gain:
Γ d = d H − d e f f . \boxed{
\Gamma_d
=
d_H-d_{\mathrm{eff}}.
} Γ d = d H − d eff .
本文進一步區分四種 AI / 外部支架作用:Externalization、Retrieval、Bridge Discovery、Parallel Execution。前兩者接近傳統 cognitive offloading;後兩者可能更直接改變關係圖、可達域與系統級並行分支數。
然而,外部支架也帶來 verification、coordination、miscalibration、dependency、internal-learning loss 與 motivational costs。因此本文提出:
K e f f = K H − Δ K m e m o r y − Δ K r e t r i e v a l − Δ K r e e n t r y + K v e r i f i c a t i o n + K c o o r d i n a t i o n + K d e p e n d e n c y . \boxed{
K_{\mathrm{eff}}
=
K_H
-
\Delta K_{\mathrm{memory}}
-
\Delta K_{\mathrm{retrieval}}
-
\Delta K_{\mathrm{reentry}}
+
K_{\mathrm{verification}}
+
K_{\mathrm{coordination}}
+
K_{\mathrm{dependency}}.
} K eff = K H − Δ K memory − Δ K retrieval − Δ K reentry + K verification + K coordination + K dependency .
現有 cognitive-offloading meta-analysis 支持外部支架可改善特定記憶任務表現並降低個體差異,但研究亦顯示 offloading 可能降低內部編碼或後續無支架表現。2025 年 PNAS 的大規模高中數學 field experiment 更顯示,無 guardrails 的 GPT 支援可在練習時大幅提升表現,卻可能在移除 AI 後降低獨立考試表現;有教學 guardrails 的 AI tutor 能大幅減輕此負面效果。2025 年四項 online experiments(總 N = 3562 N=3562 N = 3562 )亦顯示 human–GenAI collaboration 的即時產出優勢不必然轉移到後續 human-only tasks,並伴隨動機與無聊感變化。2026 年複雜臨床推理實驗則顯示 human–AI collaboration 可提高平均準確率並降低主觀 cognitive burden,但合作失敗時常與人類接受錯誤 AI insight 有關。
因此本文的核心不是:
AI ⇒ better cognition . \text{AI}\Rightarrow\text{better cognition}. AI ⇒ better cognition .
而是:
AI changes the cost structure and reachable paths of cognition under specific interface, memory, verification, and control conditions. \boxed{
\text{AI changes the cost structure and reachable paths
of cognition under specific interface,
memory, verification, and control conditions.}
} AI changes the cost structure and reachable paths of cognition under specific interface, memory, verification, and control conditions.
如果混合系統模型不能比普通「更快檢索/更多資訊/較低工作記憶負荷」模型提供額外預測,則本文所稱「認知拓樸變換」應被降級。
關鍵詞: cognitive offloading;human–AI collaboration;external memory;agents;cognitive scaffolding;task switching;re-entry;distributed cognition;effective cognitive distance;human–AI topology;TADC
0. 邊界聲明
本文不是臨床研究、醫療建議、AI 安全保證或人類能力增強的既成事實報告。
本文提出的是:
general human–tool cognitive architecture conjecture . \boxed{
\text{general human–tool cognitive architecture conjecture}.
} general human–tool cognitive architecture conjecture .
本文不主張:
AI is part of the human brain . \boxed{
\text{AI is part of the human brain}.
} AI is part of the human brain .
不主張:
human attention becomes literally parallel because AI agents work in parallel . \boxed{
\text{human attention becomes literally parallel
because AI agents work in parallel}.
} human attention becomes literally parallel because AI agents work in parallel .
也不主張:
using AI necessarily improves learning, memory, creativity, or reasoning . \boxed{
\text{using AI necessarily improves learning,
memory, creativity, or reasoning}.
} using AI necessarily improves learning, memory, creativity, or reasoning .
本文只問:
若外部系統能保存、檢索、轉換、連接並執行認知相關狀態,那麼「人類完成一個任務時真正可走的有效資訊路徑」是否因此改變?
1. 從 TADC-06 到外部支架
TADC-06 定義:
d r e l ( x , y ∣ G ) d_{\mathrm{rel}}(x,y\mid G) d rel ( x , y ∣ G )
作為 goal-conditioned relational cognitive distance。
但此前主要假設:
x , y x,y x , y
都存在於主體當前可使用的內部認知結構中。
現在加入外部系統:
notes;
file system;
search engine;
knowledge graph;
LLM;
memory-augmented assistant;
autonomous / semi-autonomous agent;
shared workspace。
此時從:
x x x
到:
y y y
的最佳路徑未必完全位於:
C H . \mathcal C^H. C H .
2. 人類內部認知結構
沿用 TADC:
C t H = ( X t H , R t H , κ t H , N t H , A t H , G t ) . \mathcal C_t^H
=
(
X_t^H,
\mathcal R_t^H,
\kappa_t^H,
\mathcal N_t^H,
A_t^H,
G_t
). C t H = ( X t H , R t H , κ t H , N t H , A t H , G t ) .
其中:
X t H X_t^H X t H :人類當前可內部操作的 cognitive objects;
R t H \mathcal R_t^H R t H :內部有效 relations;
κ t H \kappa_t^H κ t H :內部可達強度;
N t H \mathcal N_t^H N t H :內部 neighborhoods;
A t H A_t^H A t H :當前 active set;
G t G_t G t :高階 goal。
3. 外部支架
定義:
S t E = ( M t , I t , Q t , P t ) . \mathcal S_t^E
=
(
M_t,
I_t,
Q_t,
P_t
). S t E = ( M t , I t , Q t , P t ) .
其中:
M t M_t M t :external memory state;
I t I_t I t :index / address structure;
Q t Q_t Q t :retrieval mechanism;
P t P_t P t :provenance / persistence structure。
最簡紙本筆記:
S E \mathcal S^E S E
已具有:
M t M_t M t
與部分:
I t . I_t. I t .
搜尋引擎強化:
Q t . Q_t. Q t .
版本控制與資料庫強化:
P t . P_t. P t .
4. AI / Agent 層
生成式 AI 不只儲存既有資料。
定義:
A t A I = ( X t A I , R t A I , Π t , V t ) . \mathcal A_t^{AI}
=
(
X_t^{AI},
R_t^{AI},
\Pi_t,
V_t
). A t A I = ( X t A I , R t A I , Π t , V t ) .
其中:
X t A I X_t^{AI} X t A I :AI 可生成或暫存的候選 states;
R t A I R_t^{AI} R t A I :AI 建議的 relations / bridges;
Π t \Pi_t Π t :AI / agent operation policy;
V t V_t V t :verification state。
因此 AI 可以候選地:
retrieve;
transform;
propose;
branch;
execute;
summarize;
critique;
reconnect。
5. 混合認知系統
定義:
H t = C t H ⊕ S t E ⊕ A t A I . \boxed{
\mathcal H_t
=
\mathcal C_t^H
\oplus
\mathcal S_t^E
\oplus
\mathcal A_t^{AI}.
} H t = C t H ⊕ S t E ⊕ A t A I .
這裡:
⊕ \oplus ⊕
不是代數直和的嚴格宣告。
它表示:
在任務執行中,資訊處理路徑可以跨 human-internal、external-store 與 AI-computation 三種節點。
6. 有效狀態集合
令:
X t + = X t H ∪ X t E ∪ X t A I . X_t^+
=
X_t^H
\cup
X_t^E
\cup
X_t^{AI}. X t + = X t H ∪ X t E ∪ X t A I .
這是一個 task-effective state set。
它不表示:
X t E X_t^E X t E
和:
X t H X_t^H X t H
具有相同心理性質。
只是:
both can causally support task completion . \boxed{
\text{both can causally support task completion}.
} both can causally support task completion .
7. 混合關係圖
建立:
G t + = ( X t + , E H , E E , E A I , E H E , E H A , E E A ) . \mathcal G_t^+
=
(
X_t^+,
E_H,
E_E,
E_{AI},
E_{HE},
E_{HA},
E_{EA}
). G t + = ( X t + , E H , E E , E A I , E H E , E H A , E E A ) .
其中:
E H E_H E H :human-internal relation;
E E E_E E E :external-memory links;
E A I E_{AI} E A I :AI-generated relation;
E H E E_{HE} E H E :human ↔ external memory;
E H A E_{HA} E H A :human ↔ AI;
E E A E_{EA} E E A :external state ↔ AI。
8. 有效認知距離
內部-only:
d H ( x , y ∣ G ) . d_H(x,y\mid G). d H ( x , y ∣ G ) .
混合系統:
d e f f ( x , y ∣ G , H t ) = min γ ⊆ G t + K ( γ ) . \boxed{
d_{\mathrm{eff}}
(
x,y
\mid
G,\mathcal H_t
)
=
\min_{\gamma\subseteq\mathcal G_t^+}
K(\gamma).
} d eff ( x , y ∣ G , H t ) = γ ⊆ G t + min K ( γ ) .
因此最佳 path 可以是:
x H → m E → q A I → y H . x_H
\rightarrow
m_E
\rightarrow
q_{AI}
\rightarrow
y_H. x H → m E → q A I → y H .
9. Effective Cognitive Distance Transformation Conjecture(ECDTC)
ECDTC 宣稱:
可用外部支架會使至少部分 task-state pairs 的有效認知距離系統性改變。
形式:
d e f f ( x , y ∣ G , H ) ≠ d H ( x , y ∣ G ) . \boxed{
d_{\mathrm{eff}}
(
x,y\mid G,\mathcal H
)
\neq
d_H(x,y\mid G).
} d eff ( x , y ∣ G , H ) = d H ( x , y ∣ G ) .
若:
d e f f < d H , d_{\mathrm{eff}}<d_H, d eff < d H ,
定義:
Γ d = d H − d e f f > 0. \boxed{
\Gamma_d
=
d_H-d_{\mathrm{eff}}
>0.
} Γ d = d H − d eff > 0.
10. 這不是「AI 讓人變聰明」的同義詞
若:
Γ d > 0 , \Gamma_d>0, Γ d > 0 ,
只能表示:
在給定工具可用時,完成某個 transition 的有效成本下降。
這不保證:
unaided reasoning 變強;
internal memory 變強;
skill acquisition 變強;
long-term learning 變強。
所以:
augmented performance ≠ internalized capability . \boxed{
\text{augmented performance}
\neq
\text{internalized capability}.
} augmented performance = internalized capability .
11. Cognitive Offloading 的既有基礎
Risko 與 Gilbert 將 cognitive offloading 定義為利用 physical action 改變 task 的 information-processing requirements,以降低 cognitive demand。
後續 intention-offloading 與 memory-offloading 研究已反覆顯示:
external reminders can improve performance on supported memory tasks . \boxed{
\text{external reminders can improve
performance on supported memory tasks}.
} external reminders can improve performance on supported memory tasks .
因此 TADC-07 不需要重新發明:
external memory helps memory . \text{external memory helps memory}. external memory helps memory .
真正的新問題是:
可尋址、可重組、可生成的外部系統是否還會改變 branch reachability 與 cognitive switching geometry?
12. 2026 Cognitive-Offloading Meta-analysis
Burnett 與 Richmond 的 meta-analysis 聚合 memory-based cognitive offloading studies。
結果指出:
offloading 整體可改善 memory-task performance;
forced offloading 的 benefit 較 choice offloading 大;
offloading 也降低 performance 的 interindividual variability;
效果依 task design、prospective / retrospective memory 等因素而變。
這支持:
external support changes effective task performance . \boxed{
\text{external support changes effective task performance}.
} external support changes effective task performance .
但 meta-analysis 不證明:
cognitive topology changes . \boxed{
\text{cognitive topology changes}.
} cognitive topology changes .
13. Offloading 是決策,不只是工具有沒有存在
Gilbert(2024)提出 value-based cognitive-offloading model:
人類在:
internal memory cost \text{internal memory cost} internal memory cost
與:
external reminder cost \text{external reminder cost} external reminder cost
之間做 trade-off。
可寫:
U E = V r e m e m b e r − K e x t e r n a l , U_E
=
V_{\mathrm{remember}}
-
K_{\mathrm{external}}, U E = V remember − K external ,
U H = V r e m e m b e r − K i n t e r n a l . U_H
=
V_{\mathrm{remember}}
-
K_{\mathrm{internal}}. U H = V remember − K internal .
選擇:
max ( U E , U H ) . \max(U_E,U_H). max ( U E , U H ) .
因此外部支架是否改變 cognition,
還取決於:
whether the user chooses and knows how to use it . \boxed{
\text{whether the user chooses and knows how to use it}.
} whether the user chooses and knows how to use it .
14. Metacognitive Calibration
2025 MOOT 研究顯示:
offloading 策略品質和:
externalization cost;
memory accuracy;
有系統關係。
2026 的 metacognitive-training experiment 進一步顯示:
短期 prediction + feedback training 可改善 metacognitive calibration 與更 optimal reminder-setting。
所以:
tool availability ≠ optimal tool use . \boxed{
\text{tool availability}
\neq
\text{optimal tool use}.
} tool availability = optimal tool use .
需要:
metacognitive control . \boxed{
\text{metacognitive control}.
} metacognitive control .
15. 四種外部支架作用
TADC-07 將外部系統拆成四類功能:
F E = { E x , R x , B x , P x } . \boxed{
\mathcal F_E
=
\{
E_x,R_x,B_x,P_x
\}.
} F E = { E x , R x , B x , P x } .
分別:
Externalization;
Retrieval;
Bridge Discovery;
Parallel Execution。
16. Externalization
將:
x H x_H x H
寫入:
m E . m_E. m E .
形式:
W : x H → m E . W:
x_H
\rightarrow
m_E. W : x H → m E .
外部化可以降低:
K m a i n t e n a n c e . K_{\mathrm{maintenance}}. K maintenance .
17. Retrieval
需要時:
Q : m E → x ~ H . Q:
m_E
\rightarrow
\widetilde x_H. Q : m E → x H .
如果:
I t I_t I t
良好,
則:
K r e t r i e v a l ↓ . K_{\mathrm{retrieval}}\downarrow. K retrieval ↓ .
但:
x ~ H \widetilde x_H x H
不必等於原:
x H . x_H. x H .
仍有:
reconstruction error . \boxed{
\text{reconstruction error}.
} reconstruction error .
18. Bridge Discovery
AI 可以給:
x x x
提出候選:
y 1 , … , y n y_1,\ldots,y_n y 1 , … , y n
以及:
R x y . R_{xy}. R x y .
因此:
B x : G t → G ~ t + 1 . B_x:
\mathcal G_t
\rightarrow
\widetilde{\mathcal G}_{t+1}. B x : G t → G t + 1 .
這是 TADC-06:
Relation-First Cognition \text{Relation-First Cognition} Relation-First Cognition
最直接的人—AI擴展。
19. Bridge Discovery 不是有效 Bridge
AI 提出:
R x y A I R_{xy}^{AI} R x y A I
不代表:
R x y A I R_{xy}^{AI} R x y A I
是真的。
所以:
proposed edge ≠ validated edge . \boxed{
\text{proposed edge}
\neq
\text{validated edge}.
} proposed edge = validated edge .
必須有:
V ( R x y A I ) . V(R_{xy}^{AI}). V ( R x y A I ) .
20. Parallel Execution
agent system 可以同時執行:
B 1 , B 2 , … , B n . B_1,B_2,\ldots,B_n. B 1 , B 2 , … , B n .
例如:
branch 1 搜尋;
branch 2 寫 code;
branch 3 驗證 citation;
branch 4 整理文檔。
因此 system-level activity:
ν S \nu_S ν S
可以很高。
21. AI 平行不等於人類注意力平行
定義:
ν H = human attentional transition rate , \nu_H
=
\text{human attentional transition rate}, ν H = human attentional transition rate ,
ν S = hybrid-system branch event rate . \nu_S
=
\text{hybrid-system branch event rate}. ν S = hybrid-system branch event rate .
可能:
ν S ≫ ν H . \boxed{
\nu_S\gg\nu_H.
} ν S ≫ ν H .
這不代表:
human working attention became n-way parallel . \boxed{
\text{human working attention became n-way parallel}.
} human working attention became n-way parallel .
而是:
execution parallelism moved outside the human bottleneck . \boxed{
\text{execution parallelism moved outside the human bottleneck}.
} execution parallelism moved outside the human bottleneck .
22. Human Attention / System Throughput 分離
人類層:
H t . H_t. H t .
Agent / system 層:
S t . S_t. S t .
Artifact 層:
A t . A_t. A t .
因此:
A t ≠ S t ≠ H t . \boxed{
A_t
\neq
S_t
\neq
H_t.
} A t = S t = H t .
大量 commits、files 或 agent events 不能直接換算成人類 cognitive switches。
23. Hybrid Reachability Expansion Conjecture(HREC)
在有限時間:
T T T
與資源:
B B B
下,
人類內部可達集合:
Reach H T , B ( x ) . \operatorname{Reach}_H^{T,B}(x). Reach H T , B ( x ) .
加入支架:
Reach + T , B ( x ) . \operatorname{Reach}_+^{T,B}(x). Reach + T , B ( x ) .
HREC 宣稱:
Reach + T , B ( x ) ⊃ Reach H T , B ( x ) \boxed{
\operatorname{Reach}_+^{T,B}(x)
\supset
\operatorname{Reach}_H^{T,B}(x)
} Reach + T , B ( x ) ⊃ Reach H T , B ( x )
在部分 task conditions 成立。
24. Reachability Gain
定義:
Γ R = ∣ Reach + T , B ∣ − ∣ Reach H T , B ∣ ∣ Reach H T , B ∣ . \Gamma_R
=
\frac{
|
\operatorname{Reach}_+^{T,B}
|
-
|
\operatorname{Reach}_H^{T,B}
|
}{
|
\operatorname{Reach}_H^{T,B}
|
}. Γ R = ∣ Reach H T , B ∣ ∣ Reach + T , B ∣ − ∣ Reach H T , B ∣ .
若:
Γ R > 0 , \Gamma_R>0, Γ R > 0 ,
混合系統在同資源窗內能探索更多狀態。
但:
more reachable ≠ more correct . \boxed{
\text{more reachable}
\neq
\text{more correct}.
} more reachable = more correct .
25. Reachability Precision
定義:
P R = ∣ Reach v a l i d ∣ ∣ Reach a l l ∣ . P_R
=
\frac{
|\operatorname{Reach}_{\mathrm{valid}}|
}{
|\operatorname{Reach}_{\mathrm{all}}|
}. P R = ∣ Reach all ∣ ∣ Reach valid ∣ .
AI 可能:
Γ R ↑ \Gamma_R\uparrow Γ R ↑
但:
P R ↓ . P_R\downarrow. P R ↓ .
即產生大量錯誤 branch。
所以:
reachability + precision \boxed{
\text{reachability}
+
\text{precision}
} reachability + precision
必須共同測。
26. 外部支架與任務中斷
task-interruption literature 已反覆觀察:
resumption cost . \boxed{
\text{resumption cost}.
} resumption cost .
中斷後回到原 task:
reaction time 增加;
errors 可能增加;
working-memory / attentional reorientation 成本增加。
2024 的 interruption study 亦支持 suspended task goal 在切換後具有 persisting activation / inhibition-related dynamics。
因此:
returning to a task is not free . \boxed{
\text{returning to a task is not free}.
} returning to a task is not free .
27. Re-entry Cost
對 branch:
b , b, b ,
定義:
K r e e n t r y ( b ) = K l o c a t e + K r e t r i e v e + K r e c o n s t r u c t + K v e r i f y + K r e s u m e . K_{\mathrm{reentry}}(b)
=
K_{\mathrm{locate}}
+
K_{\mathrm{retrieve}}
+
K_{\mathrm{reconstruct}}
+
K_{\mathrm{verify}}
+
K_{\mathrm{resume}}. K reentry ( b ) = K locate + K retrieve + K reconstruct + K verify + K resume .
傳統多專案工作中:
K r e c o n s t r u c t K_{\mathrm{reconstruct}} K reconstruct
往往很高。
28. Re-entry Topology Conjecture(RTC)
若 external system 保存:
branch identity;
last state;
dependencies;
unresolved questions;
next action;
provenance;
則:
K l o c a t e ↓ , K_{\mathrm{locate}}\downarrow, K locate ↓ ,
K r e t r i e v e ↓ , K_{\mathrm{retrieve}}\downarrow, K retrieve ↓ ,
K r e c o n s t r u c t ↓ . K_{\mathrm{reconstruct}}\downarrow. K reconstruct ↓ .
RTC 宣稱:
K r e e n t r y E < K r e e n t r y H \boxed{
K_{\mathrm{reentry}}^{E}
<
K_{\mathrm{reentry}}^{H}
} K reentry E < K reentry H
在適當 external-state design 下成立。
29. External Memory 不只是 Storage Capacity
如果只有容量:
∣ M ∣ ↑ |M|\uparrow ∣ M ∣ ↑
但沒有:
I t , Q t , P t , I_t,
Q_t,
P_t, I t , Q t , P t ,
則:
K r e t r i e v a l K_{\mathrm{retrieval}} K retrieval
仍可很高。
因此:
memory capacity ≠ addressable cognitive support . \boxed{
\text{memory capacity}
\neq
\text{addressable cognitive support}.
} memory capacity = addressable cognitive support .
TADC-07 特別重視:
addressability . \boxed{
\text{addressability}.
} addressability .
30. Addressability
定義:
A M ( b ) = P ( correct state retrieved ∣ branch query b ) . A_M(b)
=
P(
\text{correct state retrieved}
\mid
\text{branch query }b
). A M ( b ) = P ( correct state retrieved ∣ branch query b ) .
高:
A M A_M A M
才真正降低:
K r e e n t r y . K_{\mathrm{reentry}}. K reentry .
大型 archive 若沒有索引:
∣ M ∣ ↑ |M|\uparrow ∣ M ∣ ↑
但:
A M ↓ , A_M\downarrow, A M ↓ ,
可能反而造成負擔。
31. Branch Persistence
branch:
b i b_i b i
在時間:
t t t
被暫停。
若:
P ( b i recoverable at t + Δ ) P(
b_i
\text{ recoverable at }
t+\Delta
) P ( b i recoverable at t + Δ )
因 external memory 上升,
則稱:
branch persistence gain . \boxed{
\text{branch persistence gain}.
} branch persistence gain .
定義:
Γ B = P E ( return ) − P H ( return ) . \Gamma_B
=
P_E(\text{return})
-
P_H(\text{return}). Γ B = P E ( return ) − P H ( return ) .
32. External State as Cognitive Bookmark
一個高品質 bookmark:
m b m_b m b
至少包含:
m b = ( G b , S b , D b , Q b , N b ) , m_b
=
(
G_b,
S_b,
D_b,
Q_b,
N_b
), m b = ( G b , S b , D b , Q b , N b ) ,
其中:
goal;
current state;
dependencies;
unresolved questions;
next move。
所以:
bookmark ≠ file name only . \boxed{
\text{bookmark}
\neq
\text{file name only}.
} bookmark = file name only .
33. AI Summary 作為 Lossy Compression
AI 可將:
C b C_b C b
壓成:
C ^ b . \widehat C_b. C b .
壓縮率:
r = ∣ C ^ b ∣ ∣ C b ∣ . r
=
\frac{
|\widehat C_b|
}{
|C_b|
}. r = ∣ C b ∣ ∣ C b ∣ .
但重點是:
L c r i t i c a l L_{\mathrm{critical}} L critical
是否保留。
因此 summary utility:
U S = Δ K r e e n t r y − λ L c r i t i c a l . U_S
=
\Delta K_{\mathrm{reentry}}
-
\lambda
L_{\mathrm{critical}}. U S = Δ K reentry − λ L critical .
34. Summary 可能降低也可能提高回返成本
如果 hallucinated summary:
C ^ b \widehat C_b C b
改寫了:
premise;
version;
unresolved status;
negative result;
則:
K v e r i f i c a t i o n ↑ K_{\mathrm{verification}}\uparrow K verification ↑
甚至:
wrong-branch probability ↑ . \text{wrong-branch probability}\uparrow. wrong-branch probability ↑ .
所以:
compression quality \boxed{
\text{compression quality}
} compression quality
是拓樸保持的核心條件。
35. Provenance
外部支架若保留:
P t = provenance graph , P_t
=
\text{provenance graph}, P t = provenance graph ,
使用者能追:
claim → source → version → derivation . \text{claim}
\rightarrow
\text{source}
\rightarrow
\text{version}
\rightarrow
\text{derivation}. claim → source → version → derivation .
則:
K v e r i f i c a t i o n ↓ . K_{\mathrm{verification}}\downarrow. K verification ↓ .
沒有 provenance:
K v e r i f i c a t i o n ↑ . K_{\mathrm{verification}}\uparrow. K verification ↑ .
因此:
memory without provenance ≠ reliable cognitive scaffold . \boxed{
\text{memory without provenance}
\neq
\text{reliable cognitive scaffold}.
} memory without provenance = reliable cognitive scaffold .
36. AI 改變切換幾何的最小形式
傳統 transition:
x → y x
\rightarrow
y x → y
成本:
K H ( x , y ) . K_H(x,y). K H ( x , y ) .
加入 AI:
x → q → m → y . x
\rightarrow
q
\rightarrow
m
\rightarrow
y. x → q → m → y .
若:
K ( x , q ) + K ( q , m ) + K ( m , y ) < K H ( x , y ) , K(x,q)
+
K(q,m)
+
K(m,y)
<
K_H(x,y), K ( x , q ) + K ( q , m ) + K ( m , y ) < K H ( x , y ) ,
則:
d e f f ( x , y ) < d H ( x , y ) . \boxed{
d_{\mathrm{eff}}(x,y)
<
d_H(x,y).
} d eff ( x , y ) < d H ( x , y ) .
這就是:
AI changes the effective geometry of switching . \boxed{
\text{AI changes the effective geometry of switching}.
} AI changes the effective geometry of switching .
注意:
effective 是必要限定詞。
37. AI 並沒有把內部大腦距離直接變短
若拿走 AI:
A A I = 0 , \mathcal A^{AI}=0, A A I = 0 ,
可能:
d H d_H d H
完全沒變。
所以:
d e f f ↓ ⇏ d H ↓ . \boxed{
d_{\mathrm{eff}}\downarrow
\not\Rightarrow
d_H\downarrow.
} d eff ↓ ⇒ d H ↓ .
這是 augmented cognition 與 learned/internalized cognition 的關鍵區分。
38. Scaffold-Dependent Topology
若:
d e f f + A I ≪ d H , d_{\mathrm{eff}}^{+AI}
\ll
d_H, d eff + A I ≪ d H ,
但移除 AI:
d e f f − A I ≈ d H , d_{\mathrm{eff}}^{-AI}
\approx
d_H, d eff − A I ≈ d H ,
稱:
scaffold-dependent topology . \boxed{
\text{scaffold-dependent topology}.
} scaffold-dependent topology .
這可能高效,
也可能脆弱。
39. Internalization
若長期使用支架後:
d H p o s t < d H p r e , d_H^{post}
<
d_H^{pre}, d H p os t < d H p r e ,
則存在:
internalization gain . \boxed{
\text{internalization gain}.
} internalization gain .
定義:
Γ I = d H p r e − d H p o s t . \Gamma_I
=
d_H^{pre}
-
d_H^{post}. Γ I = d H p r e − d H p os t .
只有:
Γ I > 0 \Gamma_I>0 Γ I > 0
才表示工具協作真正壓縮了 unaided cognitive distance。
40. Offloading 的代價:內部記憶可能變弱
2025 的 cognitive-offloading studies 顯示:
預期 external memory 可用時,
participants 可能降低 internal encoding。
2026 prospective-memory experiments 也顯示:
external reminders 改善被 offloaded intention 的當下成功,
但在後續移除 reminder 時,
原先被 offload 的 prospective-memory learning 可能受損。
因此:
Γ d > 0 \boxed{
\Gamma_d>0
} Γ d > 0
可以同時:
Γ I < 0. \boxed{
\Gamma_I<0.
} Γ I < 0.
也就是即時有效距離下降,
但 internalized capability 反而下降。
41. Generative AI 與 Learning Cost
Bastani 等人 2025 PNAS field experiment 在近千名高中學生中比較:
control;
GPT Base;
有 safeguard 的 GPT Tutor。
AI 在練習階段可大幅提高 performance。
但 GPT Base 組在移除 AI 的 exam 上低於 control;
有 guardrails 的 GPT Tutor 大幅減輕此負面 learning effect。
這是 TADC-07 非常重要的反例:
better scaffolded performance ⇏ better unaided learning . \boxed{
\text{better scaffolded performance}
\not\Rightarrow
\text{better unaided learning}.
} better scaffolded performance ⇒ better unaided learning .
42. 所以必須加入 Scaffold Removal Test
任何 AI cognitive-augmentation study 都應同時測:
Supported phase
P + A I . P_{+AI}. P + A I .
Removal phase
P − A I p o s t . P_{-AI}^{post}. P − A I p os t .
Baseline
P − A I p r e . P_{-AI}^{pre}. P − A I p r e .
只有三者都有,
才能分:
augmentation;
internalization;
dependency。
43. Immediate Augmentation
定義:
Γ A = P + A I − P b a s e l i n e . \Gamma_A
=
P_{+AI}
-
P_{baseline}. Γ A = P + A I − P ba se l in e .
44. Transfer / Internalization
定義:
Γ I = P − A I p o s t − P − A I p r e . \Gamma_I
=
P_{-AI}^{post}
-
P_{-AI}^{pre}. Γ I = P − A I p os t − P − A I p r e .
45. Dependency Cost
若:
P − A I p o s t < P − A I p r e , P_{-AI}^{post}
<
P_{-AI}^{pre}, P − A I p os t < P − A I p r e ,
定義:
K D = P − A I p r e − P − A I p o s t . \boxed{
K_D
=
P_{-AI}^{pre}
-
P_{-AI}^{post}.
} K D = P − A I p r e − P − A I p os t .
這是 performance-level dependency index。
46. Human–GenAI Collaboration 不是必然有持續 spillover
Wu 等人(2025)進行四項 online experiments,總:
N = 3562. N=3562. N = 3562.
整體上 human–GenAI collaboration 提升 immediate task performance,
但提升不穩定地延續到之後 human-only task。
同時,從 AI collaboration 切回 solo work 與:
intrinsic motivation 降低;
boredom 增加;
sense of control 改變;
相關。
所以:
performance topology \boxed{
\text{performance topology}
} performance topology
還不等於:
motivation topology . \boxed{
\text{motivation topology}.
} motivation topology .
47. Motivation 也會改變有效距離
如果 AI collaboration 後:
M t = motivation M_t
=
\text{motivation} M t = motivation
下降,
某些 task transitions:
K e n g a g e m e n t K_{\mathrm{engagement}} K engagement
反而可能上升。
因此:
d e f f d_{\mathrm{eff}} d eff
不能只用資訊 retrieval cost 計算。
應包含:
motivational access cost . \boxed{
\text{motivational access cost}.
} motivational access cost .
48. Clinical Human–AI Collaboration 的警告
2026 複雜眼科 reasoning experiment:
human-only 平均 accuracy 約:
0.45. 0.45. 0.45.
human–AI collaboration 約:
0.60. 0.60. 0.60.
LLM-only 約:
0.70. 0.70. 0.70.
human–AI collaboration 同時提高 confidence 並降低 subjective cognitive burden。
但有約:
20 % 20\% 20%
participants performance 下降。
失敗經常涉及:
human accepts incorrect AI insight . \boxed{
\text{human accepts incorrect AI insight}.
} human accepts incorrect AI insight .
因此低 cognitive burden:
⇏ \not\Rightarrow ⇒
高 epistemic quality。
49. Verification Cost
AI 建議:
z A I . z_{AI}. z A I .
真正採用需要:
V ( z A I ) . V(z_{AI}). V ( z A I ) .
定義:
K V = K ( source checking , logic checking , replication , cross-model verification ) . K_V
=
K(
\text{source checking},
\text{logic checking},
\text{replication},
\text{cross-model verification}
). K V = K ( source checking , logic checking , replication , cross-model verification ) .
若:
K V K_V K V
被忽略,
表面:
K e f f K_{\mathrm{eff}} K eff
會被嚴重低估。
50. Trust Calibration
令:
p C = P ( A I correct ) , p_C
=
P(
AI\text{ correct}
), p C = P ( A I correct ) ,
使用者主觀估計:
p ^ C . \widehat p_C. p C .
calibration error:
E c a l = ∣ p C − p ^ C ∣ . E_{\mathrm{cal}}
=
|
p_C-\widehat p_C
|. E cal = ∣ p C − p C ∣.
若:
E c a l ↑ , E_{\mathrm{cal}}\uparrow, E cal ↑ ,
可能:
過度信任;
過度拒絕;
verification policy 失衡。
所以:
AI availability + poor calibration \boxed{
\text{AI availability}
+
\text{poor calibration}
} AI availability + poor calibration
可能比沒有 AI 更差。
51. Effective Cost Equation
本文提出候選總成本:
K e f f = K H − Δ K M − Δ K R − Δ K R E + K V + K C + K D . \boxed{
K_{\mathrm{eff}}
=
K_H
-
\Delta K_M
-
\Delta K_R
-
\Delta K_{RE}
+
K_V
+
K_C
+
K_D.
} K eff = K H − Δ K M − Δ K R − Δ K R E + K V + K C + K D .
其中:
Δ K M \Delta K_M Δ K M :memory maintenance reduction;
Δ K R \Delta K_R Δ K R :retrieval reduction;
Δ K R E \Delta K_{RE} Δ K R E :re-entry reduction;
K V K_V K V :verification;
K C K_C K C :coordination;
K D K_D K D :dependency / skill-loss cost。
只有:
K e f f < K H K_{\mathrm{eff}}
<
K_H K eff < K H
才是淨收益。
52. Coordination Cost
多 agent 並行:
n ↑ n\uparrow n ↑
可能:
raw throughput ↑ . \text{raw throughput}\uparrow. raw throughput ↑ .
但:
K C ( n ) K_C(n) K C ( n )
也上升。
包括:
duplicate work;
conflicting answers;
merge cost;
state divergence;
version mismatch;
provenance reconciliation。
因此:
more agents ≠ monotonic cognitive gain . \boxed{
\text{more agents}
\neq
\text{monotonic cognitive gain}.
} more agents = monotonic cognitive gain .
53. Agent Fan-out
人類在 state:
x x x
建立:
n n n
個 agent branches:
b 1 , … , b n . b_1,\ldots,b_n. b 1 , … , b n .
記:
F A = n . F_A=n. F A = n .
system reachability:
Reach S \operatorname{Reach}_S Reach S
可能快速增加。
但 human review capacity:
B H B_H B H
有限。
若:
F A ≫ B H , F_A
\gg
B_H, F A ≫ B H ,
會產生:
verification backlog . \boxed{
\text{verification backlog}.
} verification backlog .
54. Verification Backlog
定義:
Q V ( t ) = N u n v e r i f i e d ( t ) . Q_V(t)
=
N_{\mathrm{unverified}}(t). Q V ( t ) = N unverified ( t ) .
若:
d Q V d t > 0 \frac{dQ_V}{dt}>0 d t d Q V > 0
長期成立,
混合系統產出速度超過人類吸收/驗證速度。
此時:
Γ R ↑ \Gamma_R\uparrow Γ R ↑
但:
P R ↓ P_R\downarrow P R ↓
或:
K V → ∞ . K_V\rightarrow\infty. K V → ∞.
55. 可達域爆炸
若 AI 每個 node 提出:
b b b
個 branches,
深度:
d , d, d ,
候選量:
O ( b d ) . O(b^d). O ( b d ) .
因此 AI 的問題可能從:
not enough options \text{not enough options} not enough options
轉成:
too many reachable options . \boxed{
\text{too many reachable options}.
} too many reachable options .
所以 Expansion 必須搭配 TADC-03 的:
C = Contraction . C
=
\text{Contraction}. C = Contraction .
56. 人—AI 系統需要 Selection Operator
AI 生成:
Y = { y 1 , … , y n } . Y
=
\{y_1,\ldots,y_n\}. Y = { y 1 , … , y n } .
人類/另一個 agent 必須:
S : Y → Y ∗ . S:
Y
\rightarrow
Y^*. S : Y → Y ∗ .
如果:
∣ Y ∣ ↑ |Y|\uparrow ∣ Y ∣ ↑
但:
Q ( S ) ↓ , Q(S)\downarrow, Q ( S ) ↓ ,
混合 cognition 變差。
57. AI 可能降低跨域 Bridge Cost
TADC-06:
d r e l ( x , y ∣ G ) . d_{\mathrm{rel}}(x,y\mid G). d rel ( x , y ∣ G ) .
AI 搜尋與類比生成可以提出 bridge:
B x y . B_{xy}. B x y .
若驗證後成立:
d r e l + A I ( x , y ) < d r e l − A I ( x , y ) . d_{\mathrm{rel}}^{+AI}(x,y)
<
d_{\mathrm{rel}}^{-AI}(x,y). d rel + A I ( x , y ) < d rel − A I ( x , y ) .
這是 AI 最接近真正「改變關係拓樸」的作用之一。
58. 但 AI 也可能製造 False Bridge
若:
B x y A I B_{xy}^{AI} B x y A I
只保 surface similarity,
不保 constraints,
則:
cross-domain hallucination . \boxed{
\text{cross-domain hallucination}.
} cross-domain hallucination .
所以:
G = Gluing G
=
\text{Gluing} G = Gluing
必須搭配:
D = Detachment . D
=
\text{Detachment}. D = Detachment .
59. AI 作為 Re-indexing Engine
AI 可以:
summarize;
outline;
decompose;
abstract;
expand。
因此:
R − : U → z U R^-:
U
\rightarrow
z_U R − : U → z U
與:
R + : z U → U ~ R^+:
z_U
\rightarrow
\widetilde U R + : z U → U
都可以外部化。
這可能:
K R ↓ . K_R\downarrow. K R ↓ .
所以 AI 不只降低 horizontal switching cost,
也可能降低:
vertical scale-switching cost . \boxed{
\text{vertical scale-switching cost}.
} vertical scale-switching cost .
60. AI 改變 TADC-04 多尺度 routing
沒有 AI:
x f → y f x_f
\rightarrow
y_f x f → y f
可能需要長 fine-scale path。
有 AI:
x f ⟶ R − U c ⟶ T V c ⟶ R + y f . x_f
\overset{R^-}{\longrightarrow}
U_c
\overset{T}{\longrightarrow}
V_c
\overset{R^+}{\longrightarrow}
y_f. x f ⟶ R − U c ⟶ T V c ⟶ R + y f .
若:
K R − + K T + K R + < K f i n e , K_{R^-}+K_T+K_{R^+}
<
K_{\mathrm{fine}}, K R − + K T + K R + < K fine ,
AI 降低了 multiscale routing cost。
61. AI 與 Topological Hyperfocus
TADC-05 定義 THF:
K 0 → K 1 → ⋯ \mathcal K_0
\rightarrow
\mathcal K_1
\rightarrow\cdots K 0 → K 1 → ⋯
保持 structural continuity。
外部 memory 可以保存:
K t \mathcal K_t K t
使 interrupted episode 後:
P r e t u r n ↑ . P_{\mathrm{return}}\uparrow. P return ↑ .
因此 AI / external state 可能:
support persistent domain attachment without continuous internal maintenance . \boxed{
\text{support persistent domain attachment
without continuous internal maintenance}.
} support persistent domain attachment without continuous internal maintenance .
62. Continuous Attention 不再是 Continuous State Maintenance
如果 state 可外存,
人類可以:
U A → U B → U C → U A U_A
\rightarrow
U_B
\rightarrow
U_C
\rightarrow
U_A U A → U B → U C → U A
而不必在 working memory 中一直保留:
U A . U_A. U A .
所以:
long-term project continuity ≠ continuous internal activation . \boxed{
\text{long-term project continuity}
\neq
\text{continuous internal activation}.
} long-term project continuity = continuous internal activation .
這對長時間尺度研究非常重要。
63. External Persistence as a New Form of Continuity
傳統:
continuity ≈ internal state persistence . \text{continuity}
\approx
\text{internal state persistence}. continuity ≈ internal state persistence .
混合系統:
continuity \text{continuity} continuity
可以由:
external state persistence + reliable re-entry \boxed{
\text{external state persistence}
+
\text{reliable re-entry}
} external state persistence + reliable re-entry
維持。
因此 persistent cognition 可變成:
discontinuous human activation over continuous addressable state . \boxed{
\text{discontinuous human activation
over continuous addressable state}.
} discontinuous human activation over continuous addressable state .
64. Hybrid Cognitive Continuity Conjecture(HCCC)
本文增加一個衍生命題:
若 external state:
M t M_t M t
能長期保存 task-relevant invariants,
則:
A t H A_t^H A t H
即使離開,
整個混合系統的 project-state continuity:
C S C_S C S
仍可保持。
形式:
A t H = 0 ⇏ C S = 0. \boxed{
A_t^H=0
\not\Rightarrow
C_S=0.
} A t H = 0 ⇒ C S = 0.
65. 這不是把 AI 說成人格主體
HCCC 只描述:
task-state persistence . \boxed{
\text{task-state persistence}.
} task-state persistence .
它不推論:
AI consciousness;
AI personhood;
AI intention;
shared phenomenology。
這些是不同問題。
66. 混合系統的四個層級
Level 0 — No External Scaffold
H = C H . \mathcal H=\mathcal C^H. H = C H .
Level 1 — Passive Memory
notes / files:
M . M. M .
Level 2 — Addressable Scaffold
search / index / graph:
M + I + Q . M+I+Q. M + I + Q .
Level 3 — Interactive AI
M + I + Q + B x + R . M+I+Q+B_x+R. M + I + Q + B x + R .
AI 可主動重表徵與提出 bridge。
Level 4 — Persistent Agent System
M + I + Q + B x + R + P x . M+I+Q+B_x+R+P_x. M + I + Q + B x + R + P x .
支援:
branch persistence;
parallel execution;
asynchronous state updates。
67. Topological Gain 不應只看工具複雜度
Level 4 不一定優於 Level 2。
如果:
K V + K C + K D K_V+K_C+K_D K V + K C + K D
過大,
則:
K e f f L 4 > K e f f L 2 . K_{\mathrm{eff}}^{L4}
>
K_{\mathrm{eff}}^{L2}. K eff L 4 > K eff L 2 .
所以:
more agentic ≠ better cognitive scaffold . \boxed{
\text{more agentic}
\neq
\text{better cognitive scaffold}.
} more agentic = better cognitive scaffold .
68. Null Model 1:Speed-Up Only
假設 AI 只降低:
K l o o k u p . K_{\mathrm{lookup}}. K lookup .
所有其他 graph structure 不變。
如果這就能解釋:
switching;
reachability;
return;
performance;
則 ECDTC 的拓樸語言不需要。
69. Null Model 2:Memory Capacity Only
假設:
∣ M ∣ ↑ |M|\uparrow ∣ M ∣ ↑
已經解釋所有 benefit。
如果 index / graph / agent architecture 沒有額外作用,
RTC / HREC 過度複雜。
70. Null Model 3:More Information Only
AI 只是給更多:
I . I. I .
如果資訊量:
∣ I ∣ |I| ∣ I ∣
控制後,
AI-specific bridge / re-indexing / re-entry 效應消失,
TADC-07 應縮減。
71. Null Model 4:Motivation / Novelty Only
AI interface 可能比較有趣。
若所有:
T e n g a g e m e n t T_{\mathrm{engagement}} T engagement
增加都由 novelty / motivation 解釋,
不能說拓樸變了。
72. Null Model 5:Ordinary Collaboration
人類一直會把 cognition offload 給其他人。
Armitage 與 Redshaw 的研究直接顯示:
other humans can serve as offloading targets . \boxed{
\text{other humans can serve as offloading targets}.
} other humans can serve as offloading targets .
因此 AI 若只是另一個 partner,
則無需新 topology。
TADC-07 必須找:
persistent addressability;
scalable branching;
rapid re-indexing;
machine retrieval;
等增量特徵。
73. Null Model 6:Artifact Throughput Illusion
如果:
A t ↑ A_t\uparrow A t ↑
只是 AI 自動生成大量 artifacts,
但:
P v a l i d a t e d ↓ P_{\mathrm{validated}}\downarrow P validated ↓
或:
P i n t e g r a t e d ↓ , P_{\mathrm{integrated}}\downarrow, P integrated ↓ ,
不能說 cognition 變強。
所以:
artifact count ≠ cognitive reachability gain . \boxed{
\text{artifact count}
\neq
\text{cognitive reachability gain}.
} artifact count = cognitive reachability gain .
74. 實驗一:Support Ladder
同一 participant 完成 multi-branch reasoning task。
條件:
no aid;
static notes;
searchable notes;
LLM without persistent memory;
LLM + persistent memory;
LLM + memory + agents。
測:
K s w i t c h , K r e e n t r y , Reach , P R , A c c u r a c y , I n t e r n a l R e c a l l . K_{\mathrm{switch}},
K_{\mathrm{reentry}},
\operatorname{Reach},
P_R,
Accuracy,
InternalRecall. K switch , K reentry , Reach , P R , A cc u r a cy , I n t er na l R ec a l l .
75. 關鍵預測
若 TADC-07 成立,
不同工具層級不只提高:
speed , \text{speed}, speed ,
還會改變:
which branches are revisited, which bridges are used, and which states become reachable . \boxed{
\text{which branches are revisited,
which bridges are used,
and which states become reachable}.
} which branches are revisited, which bridges are used, and which states become reachable .
76. 實驗二:Branch Suspension / Re-entry
建立:
b 1 , b 2 , b 3 , b 4 . b_1,b_2,b_3,b_4. b 1 , b 2 , b 3 , b 4 .
participants 反覆被迫:
b i → b j . b_i\rightarrow b_j. b i → b j .
比較:
internal-only;
note;
structured checkpoint;
AI-generated checkpoint;
provenance-preserving checkpoint。
測:
τ r e e n t r y , \tau_{\mathrm{reentry}}, τ reentry ,
state reconstruction error , \text{state reconstruction error}, state reconstruction error ,
branch completion . \text{branch completion}. branch completion .
77. 實驗三:AI Bridge Discovery
兩個外部 taxonomy 遠的 domains:
U A , U B . U_A,U_B. U A , U B .
control:
人工搜尋。
experimental:
AI 提供候選 bridges。
最終所有 bridge 都需 blind validation。
測:
d r e l p o s t , d_{\mathrm{rel}}^{post}, d rel p os t ,
K A B , K_{AB}, K A B ,
novel inference . \text{novel inference}. novel inference .
78. 實驗四:False Bridge Load
AI 故意混入:
p p p
比例的 invalid bridges。
測:
K V ( p ) , K_V(p), K V ( p ) ,
P R ( p ) , P_R(p), P R ( p ) ,
Γ d ( p ) . \Gamma_d(p). Γ d ( p ) .
預測存在臨界:
p ∗ p^* p ∗
使:
Γ d \Gamma_d Γ d
由正轉負。
79. 實驗五:Agent Fan-out
設定:
n = 1 , 2 , 4 , 8 , 16. n
=
1,2,4,8,16. n = 1 , 2 , 4 , 8 , 16.
agents 平行處理 branches。
測:
raw artifact throughput , \text{raw artifact throughput}, raw artifact throughput ,
validated throughput , \text{validated throughput}, validated throughput ,
Q V , Q_V, Q V ,
K C . K_C. K C .
預測:
validated gain 對:
n n n
不是單調增加。
80. 實驗六:Scaffold Removal
participants 經多輪 AI-assisted task 後,
移除 AI。
測:
P − A I p o s t . P_{-AI}^{post}. P − A I p os t .
和:
P − A I p r e P_{-AI}^{pre} P − A I p r e
比較。
這直接區分:
augmentation \boxed{
\text{augmentation}
} augmentation
與:
internalization . \boxed{
\text{internalization}.
} internalization .
81. 實驗七:Guardrail Design
比較:
Answer-first AI
直接給解。
Hint-first AI
提示、要求 human step。
Verify-first AI
要求 human 先提出 candidate,
AI 再 critique。
Memory-first AI
只保存/恢復狀態,
不代替 reasoning。
測:
supported performance;
later unaided performance;
cognitive load;
motivation;
retention;
verification accuracy。
82. 實驗八:System vs Human Switch Rate
使用精確 logs 分離:
ν H \nu_H ν H
與:
ν S . \nu_S. ν S .
Human switch 必須由:
explicit interaction;
eye / input focus;
task declaration;
experience sample;
估計。
Agent background events:
∉ ν H . \notin\nu_H. ∈ / ν H .
測:
ν S / ν H . \nu_S/\nu_H. ν S / ν H .
這是多 agent cognition 研究必要的 measurement correction。
83. 實驗九:External-State Continuity
讓人類離開 project:
Δ t = 1 h , 1 d , 1 w . \Delta t
=
1\text{h},1\text{d},1\text{w}. Δ t = 1 h , 1 d , 1 w .
比較不同 scaffold:
P r e t u r n , P_{\mathrm{return}}, P return ,
τ r e e n t r y , \tau_{\mathrm{reentry}}, τ reentry ,
state fidelity . \text{state fidelity}. state fidelity .
RTC 預測:
structured external state reduces decay of project continuity . \boxed{
\text{structured external state
reduces decay of project continuity}.
} structured external state reduces decay of project continuity .
84. 九個核心可證偽命題
TADC7-H1 — Effective Distance Reduction
至少部分 transitions:
d e f f + E < d H . d_{\mathrm{eff}}^{+E}
<
d_H. d eff + E < d H .
TADC7-H2 — Addressability Matters Beyond Capacity
控制:
∣ M ∣ |M| ∣ M ∣
後,
A M A_M A M
仍預測:
K r e e n t r y . K_{\mathrm{reentry}}. K reentry .
TADC7-H3 — Re-entry Gain
structured state support:
K r e e n t r y s t r u c t u r e d < K r e e n t r y u n s t r u c t u r e d . K_{\mathrm{reentry}}^{structured}
<
K_{\mathrm{reentry}}^{unstructured}. K reentry s t r u c t u r e d < K reentry u n s t r u c t u r e d .
TADC7-H4 — Hybrid Reachability Expansion
同資源窗:
∣ Reach + ∣ > ∣ Reach H ∣ . |\operatorname{Reach}_+|
>
|\operatorname{Reach}_H|. ∣ Reach + ∣ > ∣ Reach H ∣.
TADC7-H5 — Verification Moderates Gain
Γ R \Gamma_R Γ R
只有在:
P R P_R P R
足夠高時轉化成有效 performance gain。
TADC7-H6 — Agent Parallelism Is System-Level
ν S ↑ \nu_S\uparrow ν S ↑
不必伴隨:
ν H ↑ \nu_H\uparrow ν H ↑
同等幅度。
TADC7-H7 — Supported and Unaided Performance Dissociate
存在:
Γ A > 0 \Gamma_A>0 Γ A > 0
但:
Γ I ≤ 0. \Gamma_I\leq0. Γ I ≤ 0.
這已與既有 offloading / AI-learning evidence 相容。
TADC7-H8 — Tool Design Changes Internalization
不同 guardrails:
Γ I ( 1 ) ≠ Γ I ( 2 ) . \Gamma_I^{(1)}
\neq
\Gamma_I^{(2)}. Γ I ( 1 ) = Γ I ( 2 ) .
TADC7-H9 — Hybrid Model Adds Prediction
加入:
A M , K R E , Γ R , K V , K C A_M,
K_{RE},
\Gamma_R,
K_V,
K_C A M , K R E , Γ R , K V , K C
後,
out-of-sample prediction 應優於:
speed + information volume \text{speed + information volume} speed + information volume
模型。
85. 什麼會殺掉「AI 改變認知拓樸」?
F1 — Pure Speed Model Wins
如果所有效應都只需:
K l o o k u p ↓ K_{\mathrm{lookup}}\downarrow K lookup ↓
解釋,
則 topology language 多餘。
F2 — Reachability Does Not Change
若 AI 只讓原本 path 更快,
但:
Reach + = Reach H , \operatorname{Reach}_+
=
\operatorname{Reach}_H, Reach + = Reach H ,
HREC 失敗。
F3 — Re-entry Does Not Improve
若 structured external state 對:
K r e e n t r y K_{\mathrm{reentry}} K reentry
沒有穩定 effect,
RTC 失敗。
F4 — Addressability Adds Nothing
若:
∣ M ∣ |M| ∣ M ∣
足以解釋所有效果,
index / topology claim 過度。
F5 — AI Bridges Add No Valid Transfer
若 AI 只增加候選但不增加:
validated cross-domain inference , \text{validated cross-domain inference}, validated cross-domain inference ,
Bridge Discovery 不構成拓樸增益。
F6 — Agent Parallelism Produces Only Artifact Noise
若:
n ↑ n\uparrow n ↑
只使:
Q V ↑ Q_V\uparrow Q V ↑
而 validated output 不增,
system-level reachability gain 不成立。
F7 — Internal Cost Dominates
若:
K V + K C + K D > Δ K M + Δ K R + Δ K R E , K_V+K_C+K_D
>
\Delta K_M+\Delta K_R+\Delta K_{RE}, K V + K C + K D > Δ K M + Δ K R + Δ K R E ,
則:
K e f f > K H . K_{\mathrm{eff}}>K_H. K eff > K H .
此工具配置應被視為負增益。
86. 「AI 讓領域消失」不是本文主張
AI 可以降低:
d e f f ( D i , D j ) , d_{\mathrm{eff}}(D_i,D_j), d eff ( D i , D j ) ,
但不表示:
D i = D j . D_i=D_j. D i = D j .
因此:
distance reduction ≠ domain identity . \boxed{
\text{distance reduction}
\neq
\text{domain identity}.
} distance reduction = domain identity .
TADC-06 的 constraint-preserving mapping 仍然必要。
87. AI 的真正特殊性可能是「低成本重建」
傳統 external memory:
M M M
需要人類自己讀回並重建。
LLM / agent 可以:
M → C ^ → next action . M
\rightarrow
\widehat C
\rightarrow
\text{next action}. M → C → next action .
所以它不只:
store , \text{store}, store ,
而是:
store + retrieve + reconstruct candidate context . \boxed{
\text{store + retrieve + reconstruct candidate context}.
} store + retrieve + reconstruct candidate context .
這可能是:
K r e e n t r y K_{\mathrm{reentry}} K reentry
大幅下降的關鍵。
88. 但重建候選必須可驗證
如果 AI 自動補完 missing context:
C ^ = C + ϵ , \widehat C
=
C+\epsilon, C = C + ϵ ,
其中:
ϵ \epsilon ϵ
可能包含 false inference。
因此高品質 scaffold 需要:
state recovery + uncertainty disclosure + provenance . \boxed{
\text{state recovery}
+
\text{uncertainty disclosure}
+
\text{provenance}.
} state recovery + uncertainty disclosure + provenance .
89. Hybrid Cognitive Atlas
TADC-02:
A t = { ( U α , ϕ α ) } . \mathfrak A_t
=
\{(U_\alpha,\phi_\alpha)\}. A t = {( U α , ϕ α )} .
加入外部系統:
A t + = { ( U α , ϕ α , M α , Q α , V α ) } . \boxed{
\mathfrak A_t^+
=
\{
(U_\alpha,\phi_\alpha,M_\alpha,Q_\alpha,V_\alpha)
\}.
} A t + = {( U α , ϕ α , M α , Q α , V α )} .
每個 domain 不只具有 internal chart,
還可以具有:
external state;
retrieval path;
verification state。
90. 外部 Atlas 允許 Sparse Human Activation
人類不需要:
A t H A_t^H A t H
同時涵蓋:
U 1 , … , U n . U_1,\ldots,U_n. U 1 , … , U n .
只要:
A t + \mathfrak A_t^+ A t +
保留其狀態,
就能:
U 1 → U 7 → U 3 → U 1 . U_1
\rightarrow
U_7
\rightarrow
U_3
\rightarrow
U_1. U 1 → U 7 → U 3 → U 1 .
因此:
many active project branches ≠ many simultaneously active human attention states . \boxed{
\text{many active project branches}
\neq
\text{many simultaneously active human attention states}.
} many active project branches = many simultaneously active human attention states .
91. 這會改變「多工」的定義
傳統 multitasking 常研究:
rapid internal task switching . \text{rapid internal task switching}. rapid internal task switching .
混合系統則可能是:
serial human control over parallel persistent external branches . \boxed{
\text{serial human control
over parallel persistent external branches}.
} serial human control over parallel persistent external branches .
這不是傳統意義的 simultaneous human multitasking。
可以叫:
Externally Parallelized Serial Cognition . \boxed{
\text{Externally Parallelized Serial Cognition}.
} Externally Parallelized Serial Cognition .
簡稱:
E P S C . \boxed{
EPSC.
} E P S C .
92. EPSC 候選模型
Human controller:
H : b i → b j . H:
b_i\rightarrow b_j. H : b i → b j .
agents:
A i A_i A i
在背景維持:
b i . b_i. b i .
所以:
human control channel is serial-ish, execution fabric is parallel . \boxed{
\text{human control channel is serial-ish,
execution fabric is parallel}.
} human control channel is serial-ish, execution fabric is parallel .
這可能是 agentic AI 時代新的工作拓樸。
93. EPSC 的限制
若 human review:
B H B_H B H
固定,
agent throughput:
B A B_A B A
持續增加,
則:
B A B H → ∞ \frac{B_A}{B_H}
\rightarrow\infty B H B A → ∞
不可持續。
因此需要:
automated verification;
hierarchical summaries;
branch prioritization;
provenance;
exception routing。
否則:
parallelism becomes queue overload . \boxed{
\text{parallelism becomes queue overload}.
} parallelism becomes queue overload .
94. Cognitive Topology Engineering
如果 TADC-07 部分成立,
未來工具設計目標就不只是:
minimize clicks . \text{minimize clicks}. minimize clicks .
而可能是:
engineer effective cognitive topology . \boxed{
\text{engineer effective cognitive topology}.
} engineer effective cognitive topology .
包括:
shorten valid paths;
preserve branch state;
expose bridges;
reduce re-entry cost;
preserve invariants;
prevent false gluing;
maintain human controllability。
95. 六個工程指標
1. Addressability
A M . A_M. A M .
2. Re-entry cost
K R E . K_{RE}. K R E .
3. Reachability gain
Γ R . \Gamma_R. Γ R .
4. Precision
P R . P_R. P R .
5. Verification burden
K V . K_V. K V .
6. Internalization
Γ I . \Gamma_I. Γ I .
真正好的 cognitive scaffold 應同時優化,
而不是只提高:
output count . \text{output count}. output count .
96. 系列統一
TADC-01:
attention may transform effective cognitive space . \boxed{
\text{attention may transform effective cognitive space}.
} attention may transform effective cognitive space .
TADC-02:
domains may be dynamically induced . \boxed{
\text{domains may be dynamically induced}.
} domains may be dynamically induced .
TADC-03:
O = { E , C , T , G , D , R } . \boxed{
\mathcal O=\{E,C,T,G,D,R\}.
} O = { E , C , T , G , D , R } .
TADC-04:
object/domain status may be scale-relative . \boxed{
\text{object/domain status may be scale-relative}.
} object/domain status may be scale-relative .
TADC-05:
focus persistence need not imply state immobility . \boxed{
\text{focus persistence need not imply state immobility}.
} focus persistence need not imply state immobility .
TADC-06:
d e x t ≠ d r e l . \boxed{
d_{\mathrm{ext}}
\neq
d_{\mathrm{rel}}.
} d ext = d rel .
TADC-07:
d e f f = d ( C H ⊕ S E ⊕ A A I ) . \boxed{
d_{\mathrm{eff}}
=
d(
\mathcal C^H
\oplus
\mathcal S^E
\oplus
\mathcal A^{AI}
).
} d eff = d ( C H ⊕ S E ⊕ A A I ) .
97. 最小總模型
Human state:
C t H . \mathcal C_t^H. C t H .
External state:
S t E . \mathcal S_t^E. S t E .
AI system:
A t A I . \mathcal A_t^{AI}. A t A I .
Hybrid state:
H t = C t H ⊕ S t E ⊕ A t A I . \mathcal H_t
=
\mathcal C_t^H
\oplus
\mathcal S_t^E
\oplus
\mathcal A_t^{AI}. H t = C t H ⊕ S t E ⊕ A t A I .
有效 transition:
d e f f ( x , y ) = min γ ⊆ H t K ( γ ) . \boxed{
d_{\mathrm{eff}}(x,y)
=
\min_{\gamma\subseteq\mathcal H_t}
K(\gamma).
} d eff ( x , y ) = γ ⊆ H t min K ( γ ) .
98. 最小效益條件
外部系統有淨效益需:
Δ K M + Δ K R + Δ K R E > K V + K C + K D . \boxed{
\Delta K_M
+
\Delta K_R
+
\Delta K_{RE}
>
K_V
+
K_C
+
K_D.
} Δ K M + Δ K R + Δ K R E > K V + K C + K D .
而長期學習若重要,
還應要求:
Γ I ≥ Γ I min . \boxed{
\Gamma_I
\geq
\Gamma_I^{\min}.
} Γ I ≥ Γ I m i n .
否則即時效能可能以人類內部能力下降為代價。
99. 結論
本文提出:
ECDTC \boxed{
\textbf{ECDTC}
} ECDTC
即外部認知支架可改變任務中的有效認知距離;
RTC \boxed{
\textbf{RTC}
} RTC
即可尋址、可恢復的 external state 可降低中斷後 re-entry cost;
以及:
HREC \boxed{
\textbf{HREC}
} HREC
即 AI / agent system 可在有限資源窗內擴張混合系統的有效 reachability。
本文最重要的區分是:
internal cognition ≠ supported human performance ≠ hybrid-system throughput . \boxed{
\text{internal cognition}
\neq
\text{supported human performance}
\neq
\text{hybrid-system throughput}.
} internal cognition = supported human performance = hybrid-system throughput .
因此:
ν S ≫ ν H \nu_S\gg\nu_H ν S ≫ ν H
不能被解釋成人類 attention 無限制平行化;
A t ↑ A_t\uparrow A t ↑
也不能直接被解釋為 human cognitive throughput 上升。
本文真正提出的是:
AI may alter the effective geometry through which cognition is executed . \boxed{
\text{AI may alter the effective geometry
through which cognition is executed}.
} AI may alter the effective geometry through which cognition is executed .
這個改變可以來自:
Externalization , \text{Externalization}, Externalization ,
Retrieval , \text{Retrieval}, Retrieval ,
Bridge Discovery , \text{Bridge Discovery}, Bridge Discovery ,
Parallel Execution . \text{Parallel Execution}. Parallel Execution .
但每項都有成本:
Verification , \text{Verification}, Verification ,
Coordination , \text{Coordination}, Coordination ,
Dependency , \text{Dependency}, Dependency ,
Internal-learning loss . \text{Internal-learning loss}. Internal-learning loss .
因此最終不是:
AI = cognitive amplification . \boxed{
\text{AI = cognitive amplification}.
} AI = cognitive amplification .
而是:
AI creates a new cost-and-reachability landscape whose value depends on architecture and control . \boxed{
\text{AI creates a new cost-and-reachability landscape
whose value depends on architecture and control}.
} AI creates a new cost-and-reachability landscape whose value depends on architecture and control .
現有 cognitive-offloading evidence 已充分說明外部資源能改變 task performance;現有 GenAI experiments 也已顯示即時增益、長期學習、動機、confidence 與 correctness 可以彼此分離。
因此 TADC-07 的強版本只有在:
外部 state 不只增加資訊量;
addressability 能獨立降低 re-entry;
AI bridge 能增加 validated cross-domain transfer;
hybrid system 擴張有限資源下的有效 reachable states;
這些量比純 speed / memory-capacity 模型提供額外預測;
時才成立。
若不能,
本文應降級成:
cognitive offloading + human–AI task support model . \boxed{
\text{cognitive offloading + human–AI task support model}.
} cognitive offloading + human–AI task support model .
而不再使用「認知拓樸」這個較強名稱。
但是若成立,
就會得到一個非常重要的結果:
認知切換成本並不完全由大腦當下保存了多少上下文決定;它也取決於環境是否替這個認知系統保存了可尋址、可驗證、可回返的狀態。
因此:
cognitive continuity \boxed{
\text{cognitive continuity}
} cognitive continuity
可以從:
continuous internal activation \text{continuous internal activation} continuous internal activation
擴展為:
reliably recoverable state continuity . \boxed{
\text{reliably recoverable state continuity}.
} reliably recoverable state continuity .
這也是下一篇 TADC-08 的最終任務:
把整個 TADC 系列從理論語言收斂成可量測變量、preregistered experiments、模型比較與明確淘汰條件,決定「拓樸注意力」究竟能不能從一個命題猜想升級為真正的研究理論。
參考文獻
Risko EF, Gilbert SJ. Cognitive Offloading. Trends in Cognitive Sciences . 2016;20(9):676–688. doi:10.1016/j.tics.2016.07.002. PMID: 27542527.
Gilbert SJ, Boldt A, Sachdeva C, Scarampi C, Tsai P-C. Outsourcing Memory to External Tools: A Review of 'Intention Offloading'. Psychonomic Bulletin & Review . 2023;30(1):60–76. doi:10.3758/s13423-022-02139-4. PMID: 35789477.
Gilbert SJ. Cognitive offloading is value-based decision making: Modelling cognitive effort and the expected value of memory. Cognition . 2024;247:105783. doi:10.1016/j.cognition.2024.105783. PMID: 38583321.
Burnett LK, Richmond LL. Meta-analytic investigations of the effect of cognitive offloading on memory-based task performance and interindividual variability. Memory & Cognition . 2026;54(1):144–168. doi:10.3758/s13421-025-01743-8. PMID: 40500483.
Kelly MO, Karimjee B, Pereira AE, Lu X, Risko EF. Does expecting external memory support cost recognition memory? Memory & Cognition . 2025. doi:10.3758/s13421-025-01688-y. PMID: 39890707.
Magen H, Tomer-Offen M. Reduced relational and item-specific processing in cognitive offloading. Cognitive Research: Principles and Implications . 2025;10(1):41. doi:10.1186/s41235-025-00647-0. PMID: 40660049.
Fellers C, Storm BC. Offloading reduces prospective memory learning. Journal of Experimental Psychology: Learning, Memory, and Cognition . 2026. doi:10.1037/xlm0001630. PMID: 42241083.
Armitage KL, Redshaw J. Can you help me? Using others to offload cognition. Memory & Cognition . 2025;53(3):946–959. doi:10.3758/s13421-024-01621-9. PMID: 39172203.
Murphy DH, Metcalfe J. The Metacognitive Optimization of Offloading Task (MOOT): Both higher costs to offload and the accuracy of memory predict goodness of offloading performance. Journal of Experimental Psychology: General . 2025;154(4):1149–1166. doi:10.1037/xge0001726. PMID: 39847000.
Metacognitive training facilitates optimal cognitive offloading. Cognitive Research: Principles and Implications . 2026. doi:10.1186/s41235-026-00714-0. PMID: 41817942.
Hirsch P, Moretti L, Askin S, Koch I. Examining the cognitive processes underlying resumption costs in task-interruption contexts: Decay or inhibition of suspended task goals? Memory & Cognition . 2024;52(2):271–284. doi:10.3758/s13421-023-01458-8. PMID: 37674056.
Wu S, Liu Y, Ruan M, Chen S, et al. Human-generative AI collaboration enhances task performance but undermines human’s intrinsic motivation. Scientific Reports . 2025;15:15105. doi:10.1038/s41598-025-98385-2.
Bastani H, Bastani O, Sungu A, Ge H, Kabakcı Ö, Mariman R. Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences USA . 2025;122(26):e2422633122. doi:10.1073/pnas.2422633122. PMID: 40560616.
Ong KT-I, Seo J, Kim H, Kim J, Kim J, Kim S, Yeo J, Choi EY. Success and failure of human–AI collaboration in clinical reasoning: An experimental study on challenging real-world cases. International Journal of Medical Informatics . 2026;211:106342. doi:10.1016/j.ijmedinf.2026.106342. PMID: 41689881.
Tsai P-C, Scarampi C, Kliegel M, Gilbert SJ. Optimal cognitive offloading: Increased reminder usage but reduced proreminder bias in older adults. Psychology and Aging . 2023. doi:10.1037/pag0000751. PMID: 37289516.
Strategic reminder setting for time-based intentions: Influence of metacognition, delay length, and cue visibility. 2025. PMID: 40238032.
Between Alarms and Scheduling: The Effect of Cognitive Offloading on Prospective and Retrospective Memory. 2026. PMID: 42352705.
Dong CV, Lu Q, Norman KA, Michelmann S. Towards large language models with human-like episodic memory. Trends in Cognitive Sciences . 2025. doi:10.1016/j.tics.2025.06.016. PMID: 40713240.
Peer M, Epstein RA. Cognitive maps for hierarchical spaces in the human brain. Cerebral Cortex . 2025;35(9):bhaf261. doi:10.1093/cercor/bhaf261. PMID: 40982478.
Tan L, Qiu Y, Qiu L, et al. The medial and lateral orbitofrontal cortex jointly represent the cognitive map of task space. Communications Biology . 2025;8:163. doi:10.1038/s42003-025-07588-w. PMID: 39900714.
與系列的關係
已完成:
TADC-01:《注意力不是單點選擇——可變認知空間與注意—空間轉換猜想》
TADC-02:《動態認知域——領域作為局部座標圖》
TADC-03:《拓樸注意力六算子——展開、收斂、遍歷、黏合、切離與重索引》
TADC-04:《嵌套注意域與觀察尺度——宏觀/微觀的相對性與多尺度重索引》
TADC-05:《從單點超專注到拓樸超專注——域級持續性、內部高熵遍歷與可控退出》
TADC-06:《關係優先認知與跨域連續性——從學科距離到關係距離的認知拓樸猜想》
TADC-07:《外部認知支架與人—AI 認知拓樸——有效距離、回返成本與混合認知系統》
下一篇:
TADC-08:《拓樸注意力的測量、反證與工程化》
狀態: TADC-07 v0.1原始人體/臨床數據: 無理論狀態: 猜想/研究綱領;未經一般性實驗驗證AI 狀態: 不主張 AI 使用必然提升內部人類認知能力;supported performance、internalization 與 hybrid-system throughput 必須分離拓樸狀態: effective cognitive topology 為 task-level hybrid accessibility 的候選形式;不宣稱外部 AI 與人腦構成單一心理主體