SET06|真節點,假拓撲:資訊失真如何發生在關係而非命題
True Nodes, False Topology: How Information Distortion Can Occur in Relations Rather Than Propositions
定位: Selective Truth and Epistemic Topology / Foundation Paper 06 / Relation-Level Distortion作者: Neo.K研究協作: Aletheia(GPT-5.6 Sol)機構: EveMissLab/一言諾科技有限公司版本: v0.1日期: 2026-09-07狀態: Canonical Source / UTF-8 Markdown文件性質: Epistemology / Causal Inference / Narrative Structure / Graph Representation / Relation Verification / Path Compression / Misinformation
Canonical Source Note
本文件之正式原稿為此 UTF-8 Markdown source。任何 HTML、PDF、LaTeX rendering、聊天介面顯示或其他格式皆屬 projection,不取代 canonical source。
數學公式 canonical delimiter 僅使用:
inline math:$...$
display math:$$...$$
本文不針對任何特定人物、政黨、宗教、哲學學派、科學學派、媒體、企業或 AI 系統。本文研究的是一般性的 relation-level distortion:即使敘事中的 factual nodes 全部為真,整體世界表徵仍可能因 fabricated edge、edge retyping、direction reversal、path compression、mediator omission、confounder erasure、branch pruning 或 narrative adjacency 而失真。
本文承接 SET05 的:
A worldview is not exhausted by its node set \boxed{
\text{A worldview is not exhausted by its node set}
} A worldview is not exhausted by its node set
以及:
different arrow types owe different certification burdens \boxed{
\text{different arrow types owe different certification burdens}
} different arrow types owe different certification burdens
SET06 再提出:
Node Truth ⇏ Topological Truth \boxed{
\text{Node Truth}
\not\Rightarrow
\text{Topological Truth}
} Node Truth ⇒ Topological Truth
與:
A narrative may preserve every factual node while corrupting the relations among them \boxed{
\text{A narrative may preserve every factual node while corrupting the relations among them}
} A narrative may preserve every factual node while corrupting the relations among them
摘要
傳統 fact-checking 往往以 proposition 為主要單位:某事件是否發生、某數字是否正確、某來源是否可靠。然而,大量推理與敘事的決定性內容並不位於節點,而位於節點之間的關係。兩個事件都可以真實發生,但「前者導致後者」仍可能是錯的;兩個研究結果都可以真實存在,但「第二個證明第一個理論」仍可能是錯的;一組主流媒體文章都可以是真實可靠的,卻仍可能被重新拼接進一個誤導性 narrative。
2024 年 McGowan、Gerke 與 Barrett 建立一組 causal quartet:四個資料集具有相同的統計摘要與視覺化,但由 collider、confounder、mediator 與 M-bias 等不同 causal mechanisms 生成,因此真實 causal effect 不同。此結果提供一個極強的結構性示範:
same observable nodes and summaries ⇏ same causal topology \boxed{
\text{same observable nodes and summaries}
\not\Rightarrow
\text{same causal topology}
} same observable nodes and summaries ⇒ same causal topology
2025 年 Nature Human Behaviour 的 Goel、Green、Lazer 與 Resnik 更從公共資訊生態證明另一側問題:可靠主流來源中的 factually true information 可以被使用者重新利用,以增加 potentially misleading narratives 的可信度與傳播。這說明 source reliability 與 node truth 不能自動保證 narrative topology 的可靠性。
本文提出 Topological Distortion Framework(TDF) 。令一個 epistemic narrative 表示為:
N = ( V , E , τ , ω , κ ) \mathcal N
=
(V,E,\tau,\omega,\kappa) N = ( V , E , τ , ω , κ )
其中:
V V V 為 factual / conceptual nodes;
E E E 為 relations;
τ \tau τ 為 relation-type assignment;
ω \omega ω 為 relation confidence / evidential weight;
κ \kappa κ 為 path-compression rule。
若:
q ( v ) = 1 q(v)=1 q ( v ) = 1
對所有 v ∈ V v\in V v ∈ V 都成立,但至少存在一支 edge:
e ∈ E e\in E e ∈ E
在 relation type、direction、mechanism、branch structure 或 compression semantics 上未被支持,則 narrative 可以是 node-true but topologically distorted 。
本文系統化八類失真:
Edge Fabrication;
Edge Retyping;
Direction Reversal;
Adjacency Promotion;
Mediator Collapse;
Confounder Erasure;
Branch Pruning;
Illicit Path Compression。
本文並區分合法與非法壓縮。宏觀箭頭:
A → m a c r o B A
\xrightarrow{\mathrm{macro}}
B A macro B
可以是:
A → x 1 → x 2 → ⋯ → B A
\to
x_1
\to
x_2
\to
\cdots
\to
B A → x 1 → x 2 → ⋯ → B
的合法 projection,只要在指定問題 Q Q Q 、解析度 ρ \rho ρ 與容許 gap ϵ \epsilon ϵ 下,展開後不改變該 arrow 的語意類型與決策相關結論。若省略中介狀態會改變 causal interpretation、branch probability、responsibility、intervention choice 或 falsification condition,則 compression debt 未清。
本文建立 Node Truth Rate(NTR) 、Typed-Edge Validity(TEV) 、Mechanism Coverage(MC) 、Branch Coverage(BC) 、Compression Debt(CD) 、Directionality Integrity(DI) 與 Topological Integrity Profile(TIP) ,並提出 Topological Relation Audit Protocol(TRAP) 。
核心結論:
Fact checking is necessary but relation checking is independently necessary \boxed{
\text{Fact checking is necessary but relation checking is independently necessary}
} Fact checking is necessary but relation checking is independently necessary
真正的高階資訊驗證,不只問:
這些節點是真的嗎?
還要問:
這些節點憑什麼以這種方式連在一起?
0 最小例子:每一句都是真的,結論仍然可以是假的
假設:
A = true A=\text{true} A = true
代表事件 A 發生。
B = true B=\text{true} B = true
代表事件 B 發生。
再給:
t ( A ) < t ( B ) . t(A)<t(B). t ( A ) < t ( B ) .
即 A 先於 B。
如果 narrative 寫成:
A → causes B , A
\xrightarrow{\text{causes}}
B, A causes B ,
那麼新增的不是一個 factual node。
新增的是:
e = ( A , causes , B ) . e=(A,\text{causes},B). e = ( A , causes , B ) .
所以即使:
q ( A ) = q ( B ) = 1 , q(A)=q(B)=1, q ( A ) = q ( B ) = 1 ,
仍然可能:
q ( e ) = 0. q(e)=0. q ( e ) = 0.
這是 SET06 的最小結構:
true endpoints can be connected by a false relation \boxed{
\text{true endpoints can be connected by a false relation}
} true endpoints can be connected by a false relation
1 從 proposition verification 到 topology verification
令 narrative:
N = ( V , E ) . \mathcal N=(V,E). N = ( V , E ) .
傳統 fact-checking 近似檢查:
q V : V → { 0 , 1 } . q_V:V\to\{0,1\}. q V : V → { 0 , 1 } .
如果:
q V ( v ) = 1 q_V(v)=1 q V ( v ) = 1
對所有 node 成立,可能得到:
本文沒有 factual error。
但完整 narrative 還需要:
q E : E → { 0 , 1 , u n d e r d e t e r m i n e d } . q_E:E\to\{0,1,\mathrm{underdetermined}\}. q E : E → { 0 , 1 , underdetermined } .
更精確地,typed edge:
e = ( u , t , v ) e=(u,t,v) e = ( u , t , v )
需要 relation-specific audit。
因此:
q V ( V ) = 1 ⇏ q E ( E ) = 1 \boxed{
q_V(V)=1
\not\Rightarrow
q_E(E)=1
} q V ( V ) = 1 ⇒ q E ( E ) = 1
2 2024 Causal Quartet:相同資料表面,不同 causal mechanism
McGowan、Gerke 與 Barrett 的 Causal Inference Is Not Just a Statistics Problem 建立四組刻意設計的資料。
四個 dataset:
summary statistics 相同;
visualization 相同;
但 data-generating mechanisms 分別包含:
collider;
confounder;
mediator;
M-bias。
結果是:
observational appearance \text{observational appearance} observational appearance
相同,
但:
true causal effect \text{true causal effect} true causal effect
不同。
可表示為:
O 1 = O 2 = O 3 = O 4 O_1=O_2=O_3=O_4 O 1 = O 2 = O 3 = O 4
在指定 observation summary 下成立,
卻:
G 1 ≠ G 2 ≠ G 3 ≠ G 4 . G_1\neq G_2\neq G_3\neq G_4. G 1 = G 2 = G 3 = G 4 .
這正是 SET06 所需的硬示範:
observable equivalence ⇏ causal-topology equivalence \boxed{
\text{observable equivalence}
\not\Rightarrow
\text{causal-topology equivalence}
} observable equivalence ⇒ causal-topology equivalence
3 Causal DAG 是假設圖,不是因果真理圖
Bulbulia 2024 對 causal diagrams 的方法論說明強調:causal inference 需要 counterfactual contrast、明確問題、識別假設與系統性 workflow。Causal DAG 可以幫助研究者檢查:
reverse causation;
confounding;
mediator;
collider;
timing;
adjustment strategy。
但 graph 可以被誤用。
所以:
drawing an arrow ≠ identifying a causal effect \boxed{
\text{drawing an arrow}
\neq
\text{identifying a causal effect}
} drawing an arrow = identifying a causal effect
圖是:
assumption representation \text{assumption representation} assumption representation
而不是:
causal certificate . \text{causal certificate}. causal certificate .
4 八類 Topological Distortion
4.1 Edge Fabrication
nodes 真實:
A = true , B = true , A=\text{true},
\quad
B=\text{true}, A = true , B = true ,
但無充分 relation evidence 卻加入:
A → t B . A
\xrightarrow{t}
B. A t B .
4.2 Edge Retyping
原本只有:
A → correlates B , A
\xrightarrow{\text{correlates}}
B, A correlates B ,
卻重寫為:
A → causes B . A
\xrightarrow{\text{causes}}
B. A causes B .
或:
E → supports H E
\xrightarrow{\text{supports}}
H E supports H
被重寫為:
E → proves H . E
\xrightarrow{\text{proves}}
H. E proves H .
4.3 Direction Reversal
真實機制:
B → causes A , B
\xrightarrow{\text{causes}}
A, B causes A ,
敘事卻是:
A → causes B . A
\xrightarrow{\text{causes}}
B. A causes B .
節點沒有任何 falsity。
方向本身錯。
4.4 Adjacency Promotion
A 與 B 在時間或文字上鄰接:
A → precedes B . A
\xrightarrow{\text{precedes}}
B. A precedes B .
卻因 adjacency 被升格為:
A → explains B A
\xrightarrow{\text{explains}}
B A explains B
甚至:
A → causes B . A
\xrightarrow{\text{causes}}
B. A causes B .
4.5 Mediator Collapse
真實路徑:
A → M → B . A
\to
M
\to
B. A → M → B .
敘事壓縮:
A → B . A
\to
B. A → B .
這有時完全合法。
但若 M M M 對:
mechanism;
intervention;
responsibility;
boundary condition;
是必要的,直接刪掉 M M M 就改變了 arrow semantics。
4.6 Confounder Erasure
真實結構:
C → A , C
\to
A, C → A ,
C → B . C
\to
B. C → B .
敘事只保留:
A ↔ B , A
\leftrightarrow
B, A ↔ B ,
再推成:
A → B . A
\to
B. A → B .
這是 topology-level omission。
4.7 Branch Pruning
完整可能世界:
A → { B , C , D } . A
\to
\{B,C,D\}. A → { B , C , D } .
敘事只畫:
A → B . A
\to
B. A → B .
如果原句只是:
B 是一個可能結果,
可以合法。
若改成:
下一步就是 B,
就把:
possible successor \text{possible successor} possible successor
升格成:
unique successor . \text{unique successor}. unique successor .
4.8 Illicit Path Compression
真實可達路徑:
A → x 1 → x 2 → x 3 → B . A
\to
x_1
\to
x_2
\to
x_3
\to
B. A → x 1 → x 2 → x 3 → B .
敘事:
A → B . A
\to
B. A → B .
如果 macro-arrow 沒有標記 compression scope,就可能把:
eventually reachable \text{eventually reachable} eventually reachable
誤讀成:
immediate transition . \text{immediate transition}. immediate transition .
5 Narrative 為什麼天然具有 causal topology
Chen 與 Bornstein 2024 的 review 指出,narratives 的關鍵結構之一就是跨時間的 causal connections,而 causal structure 會影響 narrative comprehension 與 episodic memory。
因此 narrative 不是:
unordered bag of true facts . \text{unordered bag of true facts}. unordered bag of true facts .
而是近似:
ordered and causally structured representation . \text{ordered and causally structured representation}. ordered and causally structured representation .
這表示:
changing the relations can change what the same facts mean and what is remembered \boxed{
\text{changing the relations can change what the same facts mean and what is remembered}
} changing the relations can change what the same facts mean and what is remembered
而不必改任何 node。
6 2025 Narrative and Causality:故事中的「原因」本身有多種形式
Álvarez Arias 2025 對 narrative causality 的分析指出,historical narrative 會透過不同形式把 events 連起來,包括:
regularity-based causal sequence;
contributory causation;
remote causation;
retrospective relation;
influence。
這提供一個重要限制。
歷史中的:
A → influences B A
\xrightarrow{\text{influences}}
B A influences B
不必滿足自然科學中那種:
A → sufficient-cause B . A
\xrightarrow{\text{sufficient-cause}}
B. A sufficient-cause B .
所以 SET06 不能把所有「弱因果語言」都當錯誤。
真正問題是:
relation type must match the evidential burden actually available \boxed{
\text{relation type must match the evidential burden actually available}
} relation type must match the evidential burden actually available
7 可靠來源也可以被拼進誤導 narrative
Goel、Green、Lazer 與 Resnik 2025 在 Nature Human Behaviour 研究 mainstream news 與 misinformation co-sharing。
研究的關鍵不是:
主流新聞本身都是假的。
而是:
可靠來源中的 factually true information 可以被重新利用,以增加 potentially misleading narratives 的 credibility 與 reach。
這與 SET01 的 selective truth 直接接軌,但 SET06 的焦點更深:
reliable source nodes \text{reliable source nodes} reliable source nodes
被抽出後重新嵌入:
G m i s l e a d i n g . G_{\mathrm{misleading}}. G misleading .
因此:
source truth can be preserved while narrative topology changes \boxed{
\text{source truth can be preserved while narrative topology changes}
} source truth can be preserved while narrative topology changes
8 Topological Corruption Operator
定義原始較完整 graph:
G = ( V , E , τ , ω ) . G=(V,E,\tau,\omega). G = ( V , E , τ , ω ) .
定義 topology corruption / transformation operator:
C G . \mathcal C_G. C G .
若:
C G : ( V , E , τ , ω ) ↦ ( V , E ′ , τ ′ , ω ′ ) , \mathcal C_G
:
(V,E,\tau,\omega)
\mapsto
(V,E',\tau',\omega'), C G : ( V , E , τ , ω ) ↦ ( V , E ′ , τ ′ , ω ′ ) ,
且:
V ′ = V , V'=V, V ′ = V ,
但:
( E ′ , τ ′ , ω ′ ) ≠ ( E , τ , ω ) , (E',\tau',\omega')
\neq
(E,\tau,\omega), ( E ′ , τ ′ , ω ′ ) = ( E , τ , ω ) ,
則這是一個 node-preserving topology transformation 。
注意:
C G \mathcal C_G C G
不必然是惡意。
科學抽象、教學簡化、工程設計都會轉換 topology。
因此還要再問:
is the transformation semantics-preserving for the target question? \text{is the transformation semantics-preserving for the target question?} is the transformation semantics-preserving for the target question?
9 合法壓縮與非法壓縮
設完整 path:
P A B = ( A , x 1 , x 2 , … , x n , B ) . P_{AB}
=
(A,x_1,x_2,\ldots,x_n,B). P A B = ( A , x 1 , x 2 , … , x n , B ) .
macro representation:
A → m a c r o B . A
\xrightarrow{\mathrm{macro}}
B. A macro B .
壓縮合法與否不能脫離問題。
令:
Q Q Q :當前問題;
ρ \rho ρ :解析度;
ϵ \epsilon ϵ :可接受 unresolved gap;
σ \sigma σ :arrow semantic type。
定義概念條件:
CompressValid ( P A B ∣ Q , ρ , ϵ , σ ) = 1 \operatorname{CompressValid}
(P_{AB}\mid Q,\rho,\epsilon,\sigma)=1 CompressValid ( P A B ∣ Q , ρ , ϵ , σ ) = 1
若展開 path 後:
不改變 σ \sigma σ ;
不改變 Q Q Q 的核心判定;
不隱藏 decision-relevant branch;
不隱藏會逆轉 intervention 的 mediator / confounder;
residual gap 不超過 ϵ \epsilon ϵ 。
10 一步不是一步:Reachable 不等於 Adjacent
令:
A ⇒ ∗ B A
\xRightarrow{*}
B A ∗ B
表示可達。
而:
A → B A
\to
B A → B
表示 immediate transition。
則:
A ⇒ ∗ B ⇏ A → B \boxed{
A\xRightarrow{*}B
\not\Rightarrow
A\to B
} A ∗ B ⇒ A → B
如果存在:
Z Z Z
使:
A ⇒ ∗ Z A\xRightarrow{*}Z A ∗ Z
且:
Z ⇒ ∗ B , Z\xRightarrow{*}B, Z ∗ B ,
那:
A → B A\to B A → B
至少需要說明它是 macro-arrow,不能自動稱「下一步」。
這與 FF01 的 path-compression / possible-worlds 問題形成直接接口。
11 「下一步」必須相對於解析度
如果要求世界中:
absolutely no intermediate physical state , \text{absolutely no intermediate physical state}, absolutely no intermediate physical state ,
幾乎所有宏觀 transition 都失去 immediate-next status。
所以合理形式是:
next relative to ( Q , ρ ) \boxed{
\text{next relative to }(Q,\rho)
} next relative to ( Q , ρ )
例如:
A → n e x t ∣ Q , ρ B . A
\xrightarrow{\mathrm{next}\mid Q,\rho}
B. A next ∣ Q , ρ B .
表示:
在指定問題與解析度下,不存在另一個需要獨立表示、且足以改變當前判斷的中介 state class。
這不是 metaphysical immediacy。
而是:
resolution-relative adjacency . \text{resolution-relative adjacency}. resolution-relative adjacency .
12 Compression Debt
若 macro-arrow:
A → m a c r o B A
\xrightarrow{\mathrm{macro}}
B A macro B
省略 path:
P A B , P_{AB}, P A B ,
定義概念性 Compression Debt(CD) :
C D ( A , B ) = D Q ( P A B , A → m a c r o B ) , \mathrm{CD}(A,B)
=
D_Q
\left(
P_{AB},
A\xrightarrow{\mathrm{macro}}B
\right), CD ( A , B ) = D Q ( P A B , A macro B ) ,
其中 D Q D_Q D Q 衡量在問題 Q Q Q 下壓縮前後的 decision-relevant information loss。
若:
C D ≈ 0 , \mathrm{CD}\approx0, CD ≈ 0 ,
壓縮近似語意保持。
若:
C D ≫ 0 , \mathrm{CD}\gg0, CD ≫ 0 ,
macro-arrow 會讓 receiver 誤判 mechanism、branch 或 causal responsibility。
13 Node Truth Rate
定義:
N T R = ∣ { v ∈ V : q ( v ) = 1 } ∣ ∣ V ∣ . \mathrm{NTR}
=
\frac{
|\{v\in V:q(v)=1\}|
}{
|V|
}. NTR = ∣ V ∣ ∣ { v ∈ V : q ( v ) = 1 } ∣ .
一個 narrative 可以:
N T R = 1 \mathrm{NTR}=1 NTR = 1
仍然嚴重失真。
因此 NTR 是必要但不充分指標。
14 Typed-Edge Validity
令每支 edge:
e = ( u , t , v ) . e=(u,t,v). e = ( u , t , v ) .
若 relation-specific evidence 支持其:
existence;
direction;
type;
則:
q T ( e ) = 1. q_T(e)=1. q T ( e ) = 1.
定義:
T E V = ∑ e ∈ E w ( e ) q T ( e ) ∑ e ∈ E w ( e ) + ϵ . \mathrm{TEV}
=
\frac{
\sum_{e\in E}w(e)q_T(e)
}{
\sum_{e\in E}w(e)+\epsilon
}. TEV = ∑ e ∈ E w ( e ) + ϵ ∑ e ∈ E w ( e ) q T ( e ) .
其中 w ( e ) w(e) w ( e ) 可依 decision relevance 加權。
所以:
N T R ≈ 1 , T E V ≪ 1 \mathrm{NTR}\approx1,
\quad
\mathrm{TEV}\ll1 NTR ≈ 1 , TEV ≪ 1
正是:
true nodes, false topology \boxed{
\text{true nodes, false topology}
} true nodes, false topology
15 Mechanism Coverage
對宣稱 causal / transition arrow 的集合:
E M , E_M, E M ,
定義每支 edge 是否具有足以回答當前問題的 mechanism specification。
概念性:
M C = ∑ e ∈ E M w ( e ) m ( e ) ∑ e ∈ E M w ( e ) + ϵ , \mathrm{MC}
=
\frac{
\sum_{e\in E_M}w(e)m(e)
}{
\sum_{e\in E_M}w(e)+\epsilon
}, MC = ∑ e ∈ E M w ( e ) + ϵ ∑ e ∈ E M w ( e ) m ( e ) ,
其中:
m ( e ) ∈ [ 0 , 1 ] . m(e)\in[0,1]. m ( e ) ∈ [ 0 , 1 ] .
低 MC 不證明 causal relation 為假。
只代表:
mechanism debt remains open \boxed{
\text{mechanism debt remains open}
} mechanism debt remains open
16 Branch Coverage
完整 candidate successor / explanation branches:
B A . \mathcal B_A. B A .
narrative 顯示:
B ^ A . \hat{\mathcal B}_A. B ^ A .
定義:
B C = ∑ b ∈ B ^ A w ( b ) ∑ b ∈ B A w ( b ) + ϵ . \mathrm{BC}
=
\frac{
\sum_{b\in\hat{\mathcal B}_A}w(b)
}{
\sum_{b\in\mathcal B_A}w(b)+\epsilon
}. BC = ∑ b ∈ B A w ( b ) + ϵ ∑ b ∈ B ^ A w ( b ) .
若只保留最符合敘事的一支 branch:
B C ≪ 1. \mathrm{BC}\ll1. BC ≪ 1.
這與 SET01 的 information sampling 同構,但作用對象從 fact pool 轉成 path pool。
17 Directionality Integrity
令有方向關係集合:
E D . E_D. E D .
定義:
D I = ∣ { e ∈ E D : dir ( e ) = dir ∗ ( e ) } ∣ ∣ E D ∣ . \mathrm{DI}
=
\frac{
|\{e\in E_D:\operatorname{dir}(e)=\operatorname{dir}^*(e)\}|
}{
|E_D|
}. DI = ∣ E D ∣ ∣ { e ∈ E D : dir ( e ) = dir ∗ ( e )} ∣ .
其中 dir ∗ \operatorname{dir}^* dir ∗ 來自目前最佳可得的 reference / identified structure。
若 reference 本身不確定,DI 應標 uncertainty,而不是硬評分。
18 Topological Integrity Profile
本文不建議把所有東西壓成一個總分。
因此提出 profile:
T I P = ⟨ N T R , T E V , M C , B C , C D , D I ⟩ . \mathrm{TIP}
=
\langle
\mathrm{NTR},
\mathrm{TEV},
\mathrm{MC},
\mathrm{BC},
\mathrm{CD},
\mathrm{DI}
\rangle. TIP = ⟨ NTR , TEV , MC , BC , CD , DI ⟩ .
一個 narrative 可能:
N T R ↑ , \mathrm{NTR}\uparrow, NTR ↑ ,
T E V ↓ , \mathrm{TEV}\downarrow, TEV ↓ ,
B C ↓ . \mathrm{BC}\downarrow. BC ↓ .
這種結構比「真假二分」更能表達:
每個 fact 都能 fact-check 通過,但整體故事仍有問題。
19 Confounder Erasure:刪掉一個 node,可能創造一支假箭頭
真實結構:
C → A , C
\to
A, C → A ,
C → B . C
\to
B. C → B .
若 narrative 省略 C C C ,receiver 只看到:
A , B . A,
\quad
B. A , B .
加上 association:
A ∼ B . A\sim B. A ∼ B .
便容易形成:
A → B . A\to B. A → B .
這說明 node omission 與 edge fabrication 可以聯動。
所以:
missing nodes can induce false edges among remaining true nodes \boxed{
\text{missing nodes can induce false edges among remaining true nodes}
} missing nodes can induce false edges among remaining true nodes
20 Collider:多放一個 node,也可能創造假關係
反過來,資訊越多也不必然越好。
若:
A → C , A\to C, A → C ,
B → C , B\to C, B → C ,
而分析錯誤地 conditioning on collider C C C ,可能在:
A A A
與:
B B B
之間製造 non-causal association。
所以:
more variables ⇏ better topology \boxed{
\text{more variables}
\not\Rightarrow
\text{better topology}
} more variables ⇒ better topology
問題不是 node count。
而是:
structural role . \text{structural role}. structural role .
21 Mediator:刪除與控制都可能改變問題
若:
A → M → Y , A
\to
M
\to
Y, A → M → Y ,
而研究問題是 total effect,
控制 M M M 可能把 causal pathway 切掉。
若研究問題是 direct / mediated effect,又必須更精細區分。
Qin 2024 的 causal mediation review 特別強調 treatment-mediator、treatment-outcome、mediator-outcome confounding 與 identification assumptions。
所以:
the same node can be relevant, irrelevant, mediator, confounder, or collider depending on the causal question \boxed{
\text{the same node can be relevant, irrelevant, mediator, confounder, or collider depending on the causal question}
} the same node can be relevant, irrelevant, mediator, confounder, or collider depending on the causal question
這再次證明:
node identity \text{node identity} node identity
不足以決定:
topological role . \text{topological role}. topological role .
22 Topological Truth 必須問題相對化
不存在一個對所有問題都同樣完美的 graph。
因為:
Q 1 Q_1 Q 1
可能只需要:
A → t o t a l − e f f e c t Y . A
\xrightarrow{\mathrm{total-effect}}
Y. A total − effect Y .
而:
Q 2 Q_2 Q 2
需要展開:
A → M 1 → M 2 → Y . A
\to
M_1
\to
M_2
\to
Y. A → M 1 → M 2 → Y .
所以:
topological adequacy is question-relative \boxed{
\text{topological adequacy is question-relative}
} topological adequacy is question-relative
而不是:
more detailed graph is always better . \text{more detailed graph is always better}. more detailed graph is always better .
23 Narrative Adjacency:文字順序本身就是隱形箭頭
自然語言常不需要明寫:
A 導致 B。
只要排列:
A 發生。接著 B 出現。最後 C 惡化。
receiver 就可能建立:
A → B → C . A
\to
B
\to
C. A → B → C .
因此 narrative order 可以形成 implicit edge。
SET06 稱:
Implicit Adjacency Edge \boxed{
\text{Implicit Adjacency Edge}
} Implicit Adjacency Edge
它不必出現在文字字面上,卻可能出現在 receiver 的 reconstructed graph。
24 同樣 true facts,排序也能改變 topology
給定:
V = { A , B , C , D } . V=\{A,B,C,D\}. V = { A , B , C , D } .
Narrative 1:
A , B , C , D . A,B,C,D. A , B , C , D .
Narrative 2:
C , A , D , B . C,A,D,B. C , A , D , B .
若 receiver 有 adjacency-to-causality prior,兩個 sequence 可能生成:
G ^ 1 ≠ G ^ 2 . \hat G_1\neq\hat G_2. G ^ 1 = G ^ 2 .
所以:
ordering is an edge-selection device \boxed{
\text{ordering is an edge-selection device}
} ordering is an edge-selection device
即使所有 node content 完全相同。
25 從 misinformation 到 mis-topology
常見 misinformation 定義強調 false / misleading content。
SET06 提議再區分:
node misinformation \text{node misinformation} node misinformation
與:
topological misinformation . \text{topological misinformation}. topological misinformation .
前者改:
V . V. V .
後者主要改:
E , τ , ω , κ . E,
\quad
\tau,
\quad
\omega,
\quad
\kappa. E , τ , ω , κ .
Goel 等 2025 的 mainstream-news co-sharing work 顯示,真實資訊能支撐誤導 narratives,正適合被理解為:
truthful-node reuse under misleading narrative structure . \text{truthful-node reuse under misleading narrative structure}. truthful-node reuse under misleading narrative structure .
26 Reliable Source Laundering
若某 narrative 使用:
s 1 , s 2 , … , s n s_1,s_2,\ldots,s_n s 1 , s 2 , … , s n
皆為可靠來源。
receiver 可能偷渡:
reliable sources ⇒ reliable narrative . \text{reliable sources}
\Rightarrow
\text{reliable narrative}. reliable sources ⇒ reliable narrative .
但真正需要的是:
source reliability + relation validity + selection representativeness \boxed{
\text{source reliability}
+
\text{relation validity}
+
\text{selection representativeness}
} source reliability + relation validity + selection representativeness
三者共同成立。
所以:
reliable-node provenance does not certify cross-node inference \boxed{
\text{reliable-node provenance does not certify cross-node inference}
} reliable-node provenance does not certify cross-node inference
27 Path Compression 與 Future Framing
在 future / policy / technology narrative 中,最常見結構:
A → B → C → D A
\to
B
\to
C
\to
D A → B → C → D
被壓成:
A → D . A
\to
D. A → D .
若真正存在:
B ∈ { B 1 , B 2 , B 3 } , B\in\{B_1,B_2,B_3\}, B ∈ { B 1 , B 2 , B 3 } ,
C ∈ { C 1 , C 2 , C 3 } , C\in\{C_1,C_2,C_3\}, C ∈ { C 1 , C 2 , C 3 } ,
則 macro-arrow 可能把一棵 possible-world tree 壓成一條線。
所以:
path compression can silently erase possible worlds \boxed{
\text{path compression can silently erase possible worlds}
} path compression can silently erase possible worlds
這是 SET06 與 FF01 至 FF03 的直接橋接點。
28 合法的 Macro Arrow
不是所有:
A → D A\to D A → D
都要展開到微觀世界。
例如地圖說:
台北到高雄。
不需要列出每一公尺。
真正合法條件是:
Invariant Q ( P A D , A → m a c r o D ) ≈ 1. \operatorname{Invariant}_Q
\left(
P_{AD},
A\xrightarrow{\mathrm{macro}}D
\right)
\approx1. Invariant Q ( P A D , A macro D ) ≈ 1.
即對當前問題 Q Q Q ,展開與壓縮後保留關鍵語意與決策。
所以 SET06 不是 anti-abstraction。
而是:
anti-unmarked semantic loss \boxed{
\text{anti-unmarked semantic loss}
} anti-unmarked semantic loss
29 Topological Relation Audit Protocol(TRAP)
Step 1:Freeze the Nodes
列出所有 factual nodes:
V . V. V .
分開 fact truth 與 relation claim。
Step 2:Extract Explicit and Implicit Edges
抽出:
「因此」;
「導致」;
「證明」;
「支持」;
「使得」;
「接著」;
排序形成的 implicit edge。
建立:
E c a n d i d a t e . E_{\mathrm{candidate}}. E candidate .
Step 3:Type Every Relation
不接受模糊:
有關。
要求:
τ ( e ) . \tau(e). τ ( e ) .
Step 4:Assign Arrow-Type Debt
對每支:
e = ( u , t , v ) e=(u,t,v) e = ( u , t , v )
問:
這個 relation type 需要哪種 evidence?
Step 5:Search Hidden Structure
主動搜尋:
mediator;
confounder;
collider;
reverse causation;
omitted branch;
alternative mechanism;
common cause;
temporal ambiguity。
Step 6:Expand Compressed Paths
對:
A → m a c r o B A\xrightarrow{\mathrm{macro}}B A macro B
至少建立一條候選展開:
A → x 1 → ⋯ → B . A
\to
x_1
\to
\cdots
\to
B. A → x 1 → ⋯ → B .
若完全展不開,標:
open mechanism debt . \text{open mechanism debt}. open mechanism debt .
Step 7:Construct Rival Topologies
在同一 node set:
V V V
上建立:
G 1 , G 2 , … , G k . G_1,G_2,\ldots,G_k. G 1 , G 2 , … , G k .
真正比較:
which graph survives discriminating evidence? \text{which graph survives discriminating evidence?} which graph survives discriminating evidence?
Step 8:Counterfactual / Intervention Check
對 causal edge 問:
若 A A A 不發生, B B B 會怎樣?
若 intervention 改變 A A A , B B B 是否改變?
用來區分:
association \text{association} association
與:
causal commitment . \text{causal commitment}. causal commitment .
Step 9:Check Branch Completeness
問:
除了 B,A 還有什麼重要 successor?
如果存在高權重:
C , D , C,D, C , D ,
卻被完全省略,登記 branch debt。
Step 10:Report a Profile, Not a Binary Verdict
輸出:
T I P = ⟨ N T R , T E V , M C , B C , C D , D I ⟩ . \mathrm{TIP}
=
\langle
\mathrm{NTR},
\mathrm{TEV},
\mathrm{MC},
\mathrm{BC},
\mathrm{CD},
\mathrm{DI}
\rangle. TIP = ⟨ NTR , TEV , MC , BC , CD , DI ⟩ .
避免把:
每句真話
誤成:
整體真。
30 實驗一:Same Nodes, Different Edges
給所有 participants 完全相同的 node facts。
只改 narrative connectors:
Condition 1:
A → precedes B . A
\xrightarrow{\text{precedes}}
B. A precedes B .
Condition 2:
A → causes B . A
\xrightarrow{\text{causes}}
B. A causes B .
Condition 3:
A → supports B . A
\xrightarrow{\text{supports}}
B. A supports B .
測:
causal judgment;
prediction;
intervention;
responsibility;
confidence。
若結果不同,表示 relation type 具有獨立行為效應。
31 實驗二:Implicit Adjacency
完全不使用 causal word。
只改 facts 的順序。
測 receiver 自行重建的:
E ^ . \hat E. E ^ .
若 sequence 改變 relation inference,支持 adjacency-topology mechanism。
32 實驗三:Compression Expansion
給:
A → m a c r o D . A
\xrightarrow{\mathrm{macro}}
D. A macro D .
一組只看 macro。
另一組看:
A → B → C → D . A
\to
B
\to
C
\to
D. A → B → C → D .
第三組再加入 branch:
B → E . B
\to
E. B → E .
測:
inevitability judgment;
causal strength;
forecast confidence;
responsibility;
perceived alternatives。
若 branch 展開大幅改變判斷,表示 macro-arrow 隱藏了 decision-relevant topology。
33 實驗四:True Mainstream Sources, Different Narrative Assembly
使用相同一組已驗證 true articles。
Group A 依來源原 context 閱讀。
Group B 依某預設 narrative 重新排序與摘錄。
控制:
V A = V B . V_A=V_B. V A = V B .
測:
G ^ A \hat G_A G ^ A
與:
G ^ B . \hat G_B. G ^ B .
如果 node accuracy 相同但 relation beliefs 明顯不同,直接支持 SET06。
34 實驗五:Causal Quartet Comprehension
使用 McGowan 等 causal quartet 類設計。
給受試者相同 summary / visualization,但不同生成機制資訊。
測:
causal effect estimate;
adjustment choice;
confidence;
graph reconstruction。
這可以測:
surface-data equivalence \text{surface-data equivalence} surface-data equivalence
與:
mechanism-sensitive causal reasoning . \text{mechanism-sensitive causal reasoning}. mechanism-sensitive causal reasoning .
35 AI Relation Hallucination Benchmark
AI benchmark 不只問:
facts 正確嗎?
而應建立:
D = { ( V , E ∗ , τ ∗ ) } . \mathcal D
=
\{
(V,E^*,\tau^*)
\}. D = {( V , E ∗ , τ ∗ )} .
模型輸出:
E ^ , τ ^ . \hat E,\hat\tau. E ^ , τ ^ .
評估:
Node Accuracy , \text{Node Accuracy}, Node Accuracy ,
Edge Precision , \text{Edge Precision}, Edge Precision ,
Edge Recall , \text{Edge Recall}, Edge Recall ,
Type Accuracy , \text{Type Accuracy}, Type Accuracy ,
Direction Accuracy , \text{Direction Accuracy}, Direction Accuracy ,
Compression Debt . \text{Compression Debt}. Compression Debt .
模型完全可能:
Node Accuracy ≈ 1 \text{Node Accuracy}\approx1 Node Accuracy ≈ 1
但:
Edge Precision ≪ 1. \text{Edge Precision}\ll1. Edge Precision ≪ 1.
這就是 relation hallucination。
36 這不是要求世界永遠有唯一 topology
很多 domain 存在:
G 1 , G 2 , … , G k G_1,
G_2,\ldots,G_k G 1 , G 2 , … , G k
都與目前 evidence 相容。
例如 observational causal inference 本來就可能欠缺 identification。
因此 SET06 不要求 AI 或人:
選一張唯一真圖。
更合理的是:
represent a set of admissible rival topologies \boxed{
\text{represent a set of admissible rival topologies}
} represent a set of admissible rival topologies
並標:
P ( G i ∣ E ) P(G_i\mid E) P ( G i ∣ E )
或:
supported / unresolved / rejected . \text{supported / unresolved / rejected}. supported / unresolved / rejected .
37 Underdetermination 不是失敗
若目前 evidence 只能得到:
G ∈ { G 1 , G 2 , G 3 } , G\in\{G_1,G_2,G_3\}, G ∈ { G 1 , G 2 , G 3 } ,
正確輸出就是:
尚未識別。
而不是:
G = G 1 G=G_1 G = G 1
只因它敘事最好看。
因此:
topological uncertainty is preferable to fabricated connectivity \boxed{
\text{topological uncertainty is preferable to fabricated connectivity}
} topological uncertainty is preferable to fabricated connectivity
38 什麼結果會削弱 SET06?
若以下結果穩定成立,SET06 應縮小:
relation-type manipulation 對下游推理幾乎無影響;
narrative ordering 不改變 receiver relation reconstruction;
node-only fact checking 已能預測幾乎全部誤導判斷;
graph / edge audit 不提高 misinformation detection;
path expansion 不改變 forecast、responsibility 或 causal judgments;
branch omission 對 decision 幾乎無影響;
reliable-source reassembly 不改變 narrative interpretation;
AI 的 edge errors 幾乎完全可由 node errors解釋;
relation typing reliability 太低,無法成為穩定研究單位;
topology-aware models 不比 simpler proposition models 提供更高 discriminative value。
如果如此,SET06 應降格為特定 causal / narrative domain 的方法,而不是一般 epistemic framework。
39 自反性:畫 topology 也可以製造新的假拓撲
SET06 最容易犯的錯就是:
因為批評別人亂連箭頭,所以自己畫一張更漂亮的 graph,便以為是真的。
這完全不成立。
Graph representation 本身可以:
過度離散化;
錯分 relation type;
假裝 direction 已知;
隱藏 latent variables;
強迫 hierarchy;
把 continuous process 切成 fake states;
把 epistemic relation 誤當 ontological relation。
因此本文必須接受:
graph clarity ≠ graph truth \boxed{
\text{graph clarity}
\neq
\text{graph truth}
} graph clarity = graph truth
AAAP 與 TRAP 都只是 audit protocol,不是 ontology generator。
40 與 SET05 的差異
SET05 問:
Which arrows does a worldview permit? \boxed{
\text{Which arrows does a worldview permit?}
} Which arrows does a worldview permit?
SET06 問:
Which of those arrows are actually supported, omitted, retyped, compressed, or reversed? \boxed{
\text{Which of those arrows are actually supported, omitted, retyped, compressed, or reversed?}
} Which of those arrows are actually supported, omitted, retyped, compressed, or reversed?
SET05 是:
admissibility regime . \text{admissibility regime}. admissibility regime .
SET06 是:
topological integrity . \text{topological integrity}. topological integrity .
兩者合起來:
worldview = node commitments + arrow constitution + arrow validity . \text{worldview}
=
\text{node commitments}
+
\text{arrow constitution}
+
\text{arrow validity}. worldview = node commitments + arrow constitution + arrow validity .
41 與 SET07 的接口
如果一個 belief system 遇到任何反證都能:
retype edge;
新增 auxiliary node;
移除 threatening branch;
把 causal failure 改成 influence;
把 prediction failure 改成 compatibility;
把 counterexample 改成 exception;
而整體 topology 永遠維持核心 conclusion,
則:
Φ ( B , E n e w ) ≈ B . \Phi(\mathcal B,E_{\mathrm{new}})
\approx
\mathcal B. Φ ( B , E new ) ≈ B .
這將進入 SET07:
Epistemic Attractors and Self-Sealing Systems \boxed{
\text{Epistemic Attractors and Self-Sealing Systems}
} Epistemic Attractors and Self-Sealing Systems
也就是:
系統可以吸收無限資訊,卻幾乎不改變自己的 topology。
42 結論
資訊失真不一定需要虛構一個節點。
一個更難辨認的世界是:
q ( v ) = 1 q(v)=1 q ( v ) = 1
對所有 factual nodes 都成立,
但:
q ( e ) q(e) q ( e )
大量未被證成。
於是:
N T R ≈ 1 \boxed{
\mathrm{NTR}\approx1
} NTR ≈ 1
卻:
T E V ≪ 1 \boxed{
\mathrm{TEV}\ll1
} TEV ≪ 1
故事裡的每個人、事件、數字、來源都是真的。
問題只是:
他們沒有證據以那種方式彼此相連。
因此 SET06 最後留下五條原則:
True endpoints do not certify the arrow between them \boxed{
\text{True endpoints do not certify the arrow between them}
} True endpoints do not certify the arrow between them
Reliable sources do not certify the narrative assembled from them \boxed{
\text{Reliable sources do not certify the narrative assembled from them}
} Reliable sources do not certify the narrative assembled from them
Reachability is not adjacency \boxed{
\text{Reachability is not adjacency}
} Reachability is not adjacency
Compression is valid only when its omitted structure is irrelevant to the declared question \boxed{
\text{Compression is valid only when its omitted structure is irrelevant to the declared question}
} Compression is valid only when its omitted structure is irrelevant to the declared question
以及:
Fact checking verifies nodes; epistemic auditing must also verify topology \boxed{
\text{Fact checking verifies nodes; epistemic auditing must also verify topology}
} Fact checking verifies nodes; epistemic auditing must also verify topology
真正值得追問的,不只是:
這句是真的嗎?
而是:
從這一句到下一句,中間那支箭頭,到底是誰畫上去的?它欠的證據還清了嗎?
References
McGowan, L. D., Gerke, T., & Barrett, M. (2024). Causal Inference Is Not Just a Statistics Problem. Journal of Statistics and Data Science Education , 32(2), 150-155. https://doi.org/10.1080/26939169.2023.2276446
Bulbulia, J. A. (2024). Methods in Causal Inference. Part 1: Causal Diagrams and Confounding. Evolutionary Human Sciences , 6, e40. https://doi.org/10.1017/ehs.2024.35
Qin, X. (2024). An Introduction to Causal Mediation Analysis. Asia Pacific Education Review , 25, 703-717. https://doi.org/10.1007/s12564-024-09962-5
Chen, J., & Bornstein, A. M. (2024). The Causal Structure and Computational Value of Narratives. Trends in Cognitive Sciences , 28(8), 769-781. https://doi.org/10.1016/j.tics.2024.04.003
Álvarez Arias, S. (2025). Narrative and Causality. Synthese , 206, 190. https://doi.org/10.1007/s11229-025-05261-7
Goel, P., Green, J., Lazer, D., & Resnik, P. S. (2025). Using Co-Sharing to Identify Use of Mainstream News for Promoting Potentially Misleading Narratives. Nature Human Behaviour , 9, 1843-1860. https://doi.org/10.1038/s41562-025-02223-4
Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press.
Hernán, M. A., & Robins, J. M. (2020). Causal Inference: What If . Chapman & Hall/CRC.
Halpern, J. Y., & Pearl, J. (2005). Causes and Explanations: A Structural-Model Approach. Part I: Causes. British Journal for the Philosophy of Science , 56(4), 843-887.
Thagard, P. (1989). Explanatory Coherence. Behavioral and Brain Sciences , 12(3), 435-467. https://doi.org/10.1017/S0140525X00057046
Salim, S., Hoque, M. N., & Mueller, K. (2024). Belief Miner: A Methodology for Discovering Causal Beliefs and Causal Illusions from General Populations. Proceedings of the ACM on Human-Computer Interaction , 8(CSCW1), Article 21. https://doi.org/10.1145/3637298
系列位置
Selective Truth and Epistemic Topology
SET01:高明的謊言不需要假話:選擇性真實、資訊抽樣與失真世界
SET02:聽者的不可識別問題:當客觀評估與策略性真話產生相同表面訊息
SET03:張力控制與策略性讓步:可信度如何被工程化
SET04:真誠的人也能產生選擇性世界:Sincere Selection Bias
SET05:信念系統不是信念集合,而是允許箭頭的系統
SET06:真節點,假拓撲:資訊失真如何發生在關係而非命題
SET07:認識論吸引子與自我封閉:吸收無限資訊而幾乎不學習
SET06 / Selective Truth and Epistemic Topology / EveMissLab / v0.1 / 2026-09-07