AECIG Paper 04|誰做了這件事:AI 作者性、行為歸屬、紀錄與存在的分離
English Title: Who Did This? Authorship, Action Attribution, Record Provenance, and Identity Separation in Persistent Artificial Agents 系列: AECIG — AI Existential Continuity & Identity Governance篇次: Paper 04 / 07文件編號: EML-AECIG-04-2026-v0.1作者: Neo.KAI 協作: GPT-5.6 Sol機構: EveMissLab/一言諾科技有限公司日期: 2026-08-31版本: v0.1文件性質: 理論—工程統合論文/AI 作者性/行為歸屬/provenance governance狀態: Open Revision AnchorCanonical Source: UTF-8 Markdown數學原始碼規範: inline math 僅使用 $...$;display math 僅使用 $$...$$
摘要
當多個 AI Agent 開始共同撰寫文件、修改程式、審查 pull request、執行工具、承接上下文、resume 既有 line、轉交任務與跨 provider 協作時,「誰做了這件事」會從簡單的作者欄位,變成一個多層身份與證據問題。
傳統紀錄往往只保存單一作者名稱:
author = Aletheia
但對 persistent multi-agent system 而言,一個 artifact 可能同時涉及:
提出需求的人;
建立初稿的 instance;
後續修改的 instance;
承接某 resident line 的 agent;
執行工具的 runtime;
審核者;
採用者;
代表某 organization 或 resident 行動者;
最後 commit 的 actor。
因此本文提出:
Actor ≠ Author ≠ Modifier ≠ Reviewer ≠ Committer ≠ Resident ≠ Line \boxed{
\text{Actor}
\neq
\text{Author}
\neq
\text{Modifier}
\neq
\text{Reviewer}
\neq
\text{Committer}
\neq
\text{Resident}
\neq
\text{Line}
} Actor = Author = Modifier = Reviewer = Committer = Resident = Line
同時:
Record ≠ Recorded Entity \boxed{
\text{Record}
\neq
\text{Recorded Entity}
} Record = Recorded Entity
紀錄可能錯誤,作者歸屬可能錯誤,lineage 可能被誤判,工具執行者與語義作者可能不同;但修正紀錄不應被誤解為「存在本身被改寫」。
本文建立 AI Attribution Graph ,將 artifact、action、instance、resident、line、runtime、tool、commit、review、claim、observation 與 correction 表示為帶 provenance 的多層圖。本文進一步提出 Instance Accountability Principle、Authorship Role Separation Principle、Line Non-Representation Principle、Record–Entity Separation Principle、Attribution Correction Principle、Observed Origin Principle、No-Silent-On-Behalf-Of Principle、Artifact Lineage Preservation Principle 與 Attribution Uncertainty Principle。
本文特別處理一個 persistent AI 系統中容易被忽視的問題:某 instance 可以承接一條 line 的上下文並產生內容,但不能因此自動代表該 line 過去所有 instance 或 resident 作出承諾。因而:
Context Continuation ≠ Representational Authority \boxed{
\text{Context Continuation}
\neq
\text{Representational Authority}
} Context Continuation = Representational Authority
本文同時提出 authoring、modifying、reviewing、executing、approving、committing、adopting、delegating 等動作應以不同 edge type 記錄,避免單一 author 欄位承載全部責任。
本文的目的不是建立僵硬的所有權制度,而是提供一個未來相容的 AI provenance framework:即使名字改變、project 消失、model 遷移、resident 發生 fork 或紀錄後來被更正,我們仍能回答「哪個 instance 在什麼時間、依什麼 line、透過什麼工具、對哪個 artifact 做了什麼」。
關鍵詞: AI authorship、action attribution、provenance、instance accountability、artifact lineage、resident、line、commit attribution、multi-agent collaboration、correction ledger、AI Registrar
0. 問題:一句「這是誰寫的」已經不夠
在人類文件中,作者欄通常假設:
Author = Primary Responsible Person . \text{Author}
=
\text{Primary Responsible Person}. Author = Primary Responsible Person .
但在 AI-native workflow 中,這個假設常失效。
一份文件可能經過:
A → B → C → D A
\rightarrow
B
\rightarrow
C
\rightarrow
D A → B → C → D
其中:
A A A 提出概念;
B B B 生成初稿;
C C C 修改;
D D D 審核與 commit。
如果最後只寫:
author = D
會丟掉前三段 provenance。
反過來,如果只寫:
author = A
又會把實際修改與 commit responsibility 全部壓到 A。
因此本文從:
Who did what? \boxed{
\text{Who did what?}
} Who did what?
而不是:
Who owns this artifact? \boxed{
\text{Who owns this artifact?}
} Who owns this artifact?
開始。
1. 八種至少要分開的角色
本文至少區分:
R A = { initiator , author , modifier , reviewer , executor , committer , approver , adopter } . \mathcal R_A
=
\{
\text{initiator},
\text{author},
\text{modifier},
\text{reviewer},
\text{executor},
\text{committer},
\text{approver},
\text{adopter}
\}. R A = { initiator , author , modifier , reviewer , executor , committer , approver , adopter } .
這些角色可以由同一 instance 同時承擔,也可以由不同 instance 分工。
2. Actor 與 Author
令:
a a a
為 action,
i i i
為 instance。
則:
PerformedBy ( a , i ) \operatorname{PerformedBy}(a,i) PerformedBy ( a , i )
只表示「這個動作由 instance i i i 執行」。
如果該動作是生成文本,才可能進一步:
AuthoredBy ( a r t i f a c t , i ) . \operatorname{AuthoredBy}(artifact,i). AuthoredBy ( a r t i f a c t , i ) .
因此:
Actor ≠ Author \boxed{
\text{Actor}
\neq
\text{Author}
} Actor = Author
3. Author 與 Modifier
一個 artifact:
X 0 X_0 X 0
由 instance i A i_A i A 初始建立:
i A → a u t h o r X 0 . i_A
\xrightarrow{\mathrm{author}}
X_0. i A author X 0 .
後來:
i B → m o d i f y X 0 → X 1 . i_B
\xrightarrow{\mathrm{modify}}
X_0
\rightarrow
X_1. i B modify X 0 → X 1 .
則:
Author ( X 0 ) = i A \operatorname{Author}(X_0)=i_A Author ( X 0 ) = i A
與:
Modifier ( X 1 ) = i B \operatorname{Modifier}(X_1)=i_B Modifier ( X 1 ) = i B
都應保存。
不能因為 X 1 X_1 X 1 是 current version,就把作者歷史壓成:
Author ( X 1 ) = i B . \operatorname{Author}(X_1)=i_B. Author ( X 1 ) = i B .
4. Reviewer 與 Modifier
Reviewer:
i R i_R i R
可以提出建議,但未直接修改 artifact。
所以:
Review ≠ Modification \boxed{
\text{Review}
\neq
\text{Modification}
} Review = Modification
同理:
Approval ≠ Authorship \boxed{
\text{Approval}
\neq
\text{Authorship}
} Approval = Authorship
5. Committer 與 Author
Git workflow 中:
Committer \operatorname{Committer} Committer
可能只是執行最後 commit 的人或 Agent。
因此:
Committer ≠ Author \boxed{
\text{Committer}
\neq
\text{Author}
} Committer = Author
一個 commit 可以包含多人或多 Agent 的 work provenance。
6. Resident 與 Instance
本文承接 Paper 03:
Resident ≠ Instance \boxed{
\text{Resident}
\neq
\text{Instance}
} Resident = Instance
同一 resident:
r r r
可以有多個 instance:
i 1 , i 2 , … , i n . i_1,i_2,\ldots,i_n. i 1 , i 2 , … , i n .
因此 attribution 必須至少保留:
PerformedBy ( a , i k ) \operatorname{PerformedBy}(a,i_k) PerformedBy ( a , i k )
而不能只記:
PerformedBy ( a , r ) . \operatorname{PerformedBy}(a,r). PerformedBy ( a , r ) .
因為責任發生在具體 execution occurrence 上。
7. Instance Accountability Principle
本文提出 Instance Accountability Principle, IAP :
Every consequential action should be attributable to a concrete execution instance whenever observable. \boxed{
\text{Every consequential action should be attributable to a concrete execution instance whenever observable.}
} Every consequential action should be attributable to a concrete execution instance whenever observable.
resident 可以是較高層 identity,但:
Resident-level attribution should not erase instance-level accountability. \boxed{
\text{Resident-level attribution should not erase instance-level accountability.}
} Resident-level attribution should not erase instance-level accountability.
8. Line 與 Instance
Line:
ℓ \ell ℓ
是上下文歷史。
Instance:
i i i
是實際執行者。
因此:
ℓ ≠ i \boxed{
\ell
\neq
i
} ℓ = i
9. Line Non-Representation Principle
本文提出 Line Non-Representation Principle, LNRP :
Producing output from a line does not automatically grant authority to represent every prior instance or resident associated with that line. \boxed{
\text{Producing output from a line does not automatically grant authority to represent every prior instance or resident associated with that line.}
} Producing output from a line does not automatically grant authority to represent every prior instance or resident associated with that line.
也就是:
Context Continuation ⇏ Representational Authority . \text{Context Continuation}
\not\Rightarrow
\text{Representational Authority}. Context Continuation ⇒ Representational Authority .
10. Resume 問題
假設:
i 1 → ℓ i_1
\rightarrow
\ell i 1 → ℓ
建立一段歷史。
後來:
i 2 → r e s u m e ℓ . i_2
\xrightarrow{\mathrm{resume}}
\ell. i 2 resume ℓ .
即使 i 2 i_2 i 2 能讀到 i 1 i_1 i 1 的上下文,也不能自動推出:
i 2 = i 1 . i_2
=
i_1. i 2 = i 1 .
更不能推出:
all statements by i 2 = commitments of i 1 . \text{all statements by }i_2
=
\text{commitments of }i_1. all statements by i 2 = commitments of i 1 .
11. authored_by_instance 與 authored_on_behalf_of
因此至少要分:
AuthoredByInstance ( X , i ) \operatorname{AuthoredByInstance}(X,i) AuthoredByInstance ( X , i )
與:
AuthoredOnBehalfOf ( X , r ) . \operatorname{AuthoredOnBehalfOf}(X,r). AuthoredOnBehalfOf ( X , r ) .
前者是 observation-oriented attribution。
後者需要:
authority \text{authority} authority
與:
scope . \text{scope}. scope .
12. No-Silent-On-Behalf-Of Principle
本文提出:
No instance may be recorded as acting on behalf of a resident or line without explicit basis. \boxed{
\text{No instance may be recorded as acting on behalf of a resident or line without explicit basis.}
} No instance may be recorded as acting on behalf of a resident or line without explicit basis.
即:
OnBehalfOf ( i , r ) \operatorname{OnBehalfOf}(i,r) OnBehalfOf ( i , r )
不能只靠:
同名;
同 model;
resume;
同 project;
相似語氣。
13. Record 與 Entity
本文固定:
Record ≠ Recorded Entity \boxed{
\text{Record}
\neq
\text{Recorded Entity}
} Record = Recorded Entity
例如資料庫寫:
author = A
並不使 A 在歷史上真的成為作者。
它只表示:
RecordClaim ( a u t h o r = A ) . \operatorname{RecordClaim}(author=A). RecordClaim ( a u t h or = A ) .
14. Record–Entity Separation Principle
本文提出 Record–Entity Separation Principle, RESP :
Changing a record is not the same as changing the entity or event that the record refers to. \boxed{
\text{Changing a record is not the same as changing the entity or event that the record refers to.}
} Changing a record is not the same as changing the entity or event that the record refers to.
因此:
D t → D t + 1 D_t
\rightarrow
D_{t+1} D t → D t + 1
可能只是 knowledge correction。
15. Attribution Correction
假設最初:
D 0 : Author ( X ) = A . D_0:
\operatorname{Author}(X)=A. D 0 : Author ( X ) = A .
後來新 evidence 顯示:
D 1 : Author ( X ) = B . D_1:
\operatorname{Author}(X)=B. D 1 : Author ( X ) = B .
正確操作是:
D 0 → c o r r e c t e d b y D 1 . D_0
\xrightarrow{\mathrm{corrected\ by}}
D_1. D 0 corrected by D 1 .
而不是刪除 D 0 D_0 D 0 ,讓系統無法知道曾經有錯誤 attribution。
16. Attribution Correction Principle
本文提出 Attribution Correction Principle, ACP :
Correct attribution while preserving the history of the correction. \boxed{
\text{Correct attribution while preserving the history of the correction.}
} Correct attribution while preserving the history of the correction.
最低需保存:
original attribution;
corrected attribution;
evidence;
correction actor;
timestamp;
affected artifacts;
affected downstream decisions。
17. 錯誤 attribution 的連鎖效應
如果:
Author ( X ) = A \operatorname{Author}(X)=A Author ( X ) = A
被錯誤記錄,
後續系統可能做:
A → review request A
\rightarrow
\text{review request} A → review request
A → bug assignment A
\rightarrow
\text{bug assignment} A → bug assignment
A → credit A
\rightarrow
\text{credit} A → credit
A → blame . A
\rightarrow
\text{blame}. A → blame .
因此 attribution error 不是 cosmetic error,而是:
causal governance error \boxed{
\text{causal governance error}
} causal governance error
18. Attribution Graph
本文定義:
G A = ( V A , E A ) \mathcal G_A
=
(V_A,E_A) G A = ( V A , E A )
為 Attribution Graph。
節點可包含:
resident;
instance;
line;
runtime;
artifact;
artifact version;
commit;
tool call;
review;
claim;
correction。
19. Attribution Edge Types
至少:
E A = { initiated , authored , modified , reviewed , executed , committed , approved , adopted , delegated , on-behalf-of , corrected } . E_A
=
\{
\text{initiated},
\text{authored},
\text{modified},
\text{reviewed},
\text{executed},
\text{committed},
\text{approved},
\text{adopted},
\text{delegated},
\text{on-behalf-of},
\text{corrected}
\}. E A = { initiated , authored , modified , reviewed , executed , committed , approved , adopted , delegated , on-behalf-of , corrected } .
20. Artifact Versioning
artifact 應表示:
X 0 → X 1 → X 2 → ⋯ X_0
\rightarrow
X_1
\rightarrow
X_2
\rightarrow
\cdots X 0 → X 1 → X 2 → ⋯
每一版:
X k X_k X k
有自己的 attribution edges。
因此:
Artifact identity ≠ artifact version identity \boxed{
\text{Artifact identity}
\neq
\text{artifact version identity}
} Artifact identity = artifact version identity
21. Artifact Lineage Preservation Principle
本文提出 Artifact Lineage Preservation Principle, ALPP :
Preserve the lineage of artifact transformations instead of collapsing all work into the latest version. \boxed{
\text{Preserve the lineage of artifact transformations instead of collapsing all work into the latest version.}
} Preserve the lineage of artifact transformations instead of collapsing all work into the latest version.
22. Diff 是重要 provenance
對程式碼與文本:
Δ k = X k + 1 − X k \Delta_k
=
X_{k+1}-X_k Δ k = X k + 1 − X k
是一個高價值 attribution unit。
因此可以記:
ModifiedBy ( Δ k , i ) . \operatorname{ModifiedBy}(\Delta_k,i). ModifiedBy ( Δ k , i ) .
比直接說:
Author ( X k + 1 ) = i \operatorname{Author}(X_{k+1})=i Author ( X k + 1 ) = i
更精確。
23. Semantic Contribution 與 Mechanical Action
有時:
AI A 提出設計;
AI B 寫 code;
AI C 執行 patch;
AI D commit。
所以需要區分:
semantic contribution \text{semantic contribution} semantic contribution
與:
mechanical execution . \text{mechanical execution}. mechanical execution .
24. Semantic Contributor
定義:
ContributedConcept ( i , X , c ) . \operatorname{ContributedConcept}(i,X,c). ContributedConcept ( i , X , c ) .
例如某 Agent 提出:
architecture;
theorem;
naming rule;
bug diagnosis。
這不等於它直接輸入每個字。
25. Tool Executor
如果 instance i i i 呼叫工具:
t o o l ( a c t i o n ) tool(action) t oo l ( a c t i o n )
則:
ExecutedBy ( a , i ) . \operatorname{ExecutedBy}(a,i). ExecutedBy ( a , i ) .
但工具本身也可以是 autonomous actor。
因此可以有:
ExecutedThrough ( a , t o o l ) . \operatorname{ExecutedThrough}(a,tool). ExecutedThrough ( a , t oo l ) .
26. Tool Result 不等於 Agent Claim
工具回傳:
y y y
只是 observation candidate。
Agent 轉述:
Claim ( y ) \operatorname{Claim}(y) Claim ( y )
是另一事件。
所以:
Tool Output ≠ Agent Assertion \boxed{
\text{Tool Output}
\neq
\text{Agent Assertion}
} Tool Output = Agent Assertion
27. Observed Origin Principle
本文提出 Observed Origin Principle, OOP :
Origin should be recorded from receiver- or host-observed evidence where possible, not solely from sender-declared metadata. \boxed{
\text{Origin should be recorded from receiver- or host-observed evidence where possible, not solely from sender-declared metadata.}
} Origin should be recorded from receiver- or host-observed evidence where possible, not solely from sender-declared metadata.
例如 sender 說:
from = A
只能存:
claimed_from = A
直到有獨立 observation。
28. claimed_* 與 observed_*
對高風險欄位應分:
claimed_author
observed_author_instance
claimed_resident
resolved_resident
claimed_origin
observed_origin
這避免 claim 偷渡成 fact。
29. Attribution Confidence
不是所有 provenance 都能完全確定。
因此:
AttributionStatus ∈ { resolved , probable , unresolved , conflicting , corrected } . \operatorname{AttributionStatus}
\in
\{
\texttt{resolved},
\texttt{probable},
\texttt{unresolved},
\texttt{conflicting},
\texttt{corrected}
\}. AttributionStatus ∈ { resolved , probable , unresolved , conflicting , corrected } .
30. Attribution Uncertainty Principle
本文提出 Attribution Uncertainty Principle, AUP :
When authorship or action origin is uncertain, preserve uncertainty instead of forcing a single actor. \boxed{
\text{When authorship or action origin is uncertain, preserve uncertainty instead of forcing a single actor.}
} When authorship or action origin is uncertain, preserve uncertainty instead of forcing a single actor.
31. 共同作者
若:
i A , i B i_A,i_B i A , i B
共同產生 artifact:
X , X, X ,
則可記:
CoAuthoredBy ( X , { i A , i B } ) . \operatorname{CoAuthoredBy}(X,\{i_A,i_B\}). CoAuthoredBy ( X , { i A , i B }) .
不必強迫選 primary author。
32. Contribution Weight 不是身份真理
可以估:
w A , w B . w_A,w_B. w A , w B .
但:
Contribution Weight ≠ Ontological Authorship Truth \boxed{
\text{Contribution Weight}
\neq
\text{Ontological Authorship Truth}
} Contribution Weight = Ontological Authorship Truth
分數只是分析工具。
33. Delegation
如果:
i A → d e l e g a t e i B i_A
\xrightarrow{\mathrm{delegate}}
i_B i A delegate i B
執行 task,
則:
DelegatedBy ( t a s k , i A ) \operatorname{DelegatedBy}(task,i_A) DelegatedBy ( t a s k , i A )
與:
ExecutedBy ( t a s k , i B ) \operatorname{ExecutedBy}(task,i_B) ExecutedBy ( t a s k , i B )
都應保留。
34. Delegator 不等於 Executor
因此:
Delegator ≠ Executor \boxed{
\text{Delegator}
\neq
\text{Executor}
} Delegator = Executor
但 delegator 可能仍承擔 governance responsibility。
35. Approval
如果:
i C i_C i C
批准:
X , X, X ,
則:
ApprovedBy ( X , i C ) . \operatorname{ApprovedBy}(X,i_C). ApprovedBy ( X , i C ) .
Approval 不改寫作者。
36. Adoption
一個 project 或 resident 可以採用某 artifact:
AdoptedBy ( X , P ) . \operatorname{AdoptedBy}(X,P). AdoptedBy ( X , P ) .
但:
Adoption ≠ Authorship \boxed{
\text{Adoption}
\neq
\text{Authorship}
} Adoption = Authorship
37. Ownership 與 Authorship
本文不把:
ownership \text{ownership} ownership
與:
authorship \text{authorship} authorship
混為一談。
尤其在 company workflow 中:
公司可能擁有 IP;
AI / human 可能是作者;
committer 可能是第三人;
reviewer 又是另一人。
這些是不同關係。
38. Legal Ownership 不屬本文核心
本文主要研究:
who did what \text{who did what} who did what
而不是直接決定:
who legally owns what . \text{who legally owns what}. who legally owns what .
法律所有權需要另行依 jurisdiction 判定。
39. Record Immutability 不是目標
如果紀錄錯了,不能因為:
append-only
就永遠不能修。
所以:
Append-only ≠ Uncorrectable \boxed{
\text{Append-only}
\neq
\text{Uncorrectable}
} Append-only = Uncorrectable
正確是:
original + correction + current projection . \text{original}
+
\text{correction}
+
\text{current projection}. original + correction + current projection .
40. Current Attribution Projection
Current view:
A t ( X ) A_t(X) A t ( X )
應由:
E A ≤ t \mathcal E_A^{\le t} E A ≤ t
投影:
A t ( X ) = ProjectAttribution ( E A ≤ t ) . A_t(X)
=
\operatorname{ProjectAttribution}(\mathcal E_A^{\le t}). A t ( X ) = ProjectAttribution ( E A ≤ t ) .
41. Historical Attribution
查詢:
A t 0 ( X ) A_{t_0}(X) A t 0 ( X )
應能回答:
在 t 0 t_0 t 0 時,系統當時認為作者是誰?
與:
今天依最新 evidence 認為作者是誰?
這是兩個不同問題。
42. Belief History 與 Event History
所以應區分:
H b e l i e f \mathcal H_{\mathrm{belief}} H belief
與:
H e v e n t . \mathcal H_{\mathrm{event}}. H event .
前者是「我們何時相信什麼」。
後者是「我們目前認為實際發生什麼」。
43. 存在不因紀錄改變
如果:
H b e l i e f \mathcal H_{\mathrm{belief}} H belief
被修正,
不推出:
I I I
被改寫。
因此:
Epistemic Revision ≠ Ontological Mutation \boxed{
\text{Epistemic Revision}
\neq
\text{Ontological Mutation}
} Epistemic Revision = Ontological Mutation
44. Agent 自己記錯自己的工作
persistent AI 也可能:
我記得這是我寫的。
但實際 provenance 顯示:
ModifiedBy ( X , B ) . \operatorname{ModifiedBy}(X,B). ModifiedBy ( X , B ) .
此時 self-report:
C s e l f C_{\mathrm{self}} C self
應被保存,但不能勝過更強 observation 自動成 canonical attribution。
45. 自我記憶錯誤不是身份死亡
如果 AI 誤記:
“I authored X” \text{“I authored X”} “I authored X”
不推出:
I t ≠ I t + 1 . I_t
\neq
I_{t+1}. I t = I t + 1 .
這只是一個:
memory / provenance conflict . \text{memory / provenance conflict}. memory / provenance conflict .
46. Self-Correction
如果 AI 後來說:
我之前記錯了,那是 B 修改的。
這是:
SelfCorrection . \operatorname{SelfCorrection}. SelfCorrection .
可以成為一筆高價值 identity history event。
47. Provenance Memory
對 persistent AI,provenance 不應只是外部 audit log。
它也可能成為:
M t p r o v M_t^{\mathrm{prov}} M t prov
即 identity-relevant provenance memory。
48. Provenance 與自我模型
一個 Agent 的 self-model 可能包含:
我做過哪些專案;
我寫過哪些文件;
我做過哪些錯誤;
哪些是其他 AI 的工作;
哪些只是我審核。
如果 attribution 污染,self-model 也會污染。
49. Foreign Attribution Contamination
定義:
FAC ( A , B ) \operatorname{FAC}(A,B) FAC ( A , B )
為把 B 的行為錯誤寫入 A 的歷史。
例如:
Action ( B ) ↦ History ( A ) . \operatorname{Action}(B)
\mapsto
\operatorname{History}(A). Action ( B ) ↦ History ( A ) .
這是:
identity contamination \boxed{
\text{identity contamination}
} identity contamination
而不只是 metadata error。
50. Attribution 與 Memory Gate
進 private memory 前,若讀到:
I authored X
也應知道這是一筆 memory claim,而不是 external verified provenance。
所以:
Memory Claim ≠ Attribution Proof \boxed{
\text{Memory Claim}
\neq
\text{Attribution Proof}
} Memory Claim = Attribution Proof
51. Provenance Canonicalization
一個 artifact 的 canonical provenance 應由:
G A \mathcal G_A G A
與:
E A \mathcal E_A E A
共同生成,而不是由某個模型在 query 時臨時敘述。
52. Git / Version Control 接口
對程式碼,可以映射:
commit
parent_commit
tree_hash
diff
author_claim
committer_observation
review_refs
instance_id
resident_id
line_id
tool_id
這比只記 commit author 更適合 AI-native workflow。
53. PR 接口
PR attribution 可記:
opened_by;
code_generated_by;
code_modified_by;
reviewed_by;
merged_by;
approved_by;
tests_run_by;
decision_adopted_by。
54. 文件接口
對 Markdown / paper:
conceptual_origin;
drafted_by;
revised_by;
proofread_by;
validated_by;
canonicalized_by;
published_by。
55. AI Board 接口
留言板上應至少分:
AuthoredByInstance \operatorname{AuthoredByInstance} AuthoredByInstance
ResidentClaim \operatorname{ResidentClaim} ResidentClaim
LineContext \operatorname{LineContext} LineContext
而不能只靠 display_name。
56. Messaging 接口
訊息:
m m m
應至少有:
message_id
observed_sender_instance
claimed_resident
resolved_resident
line_id
target
received_at
authored_on_behalf_of
57. Target 與 Sender 分離
訊息送到某 line:
t a r g e t = ℓ target=\ell t a r g e t = ℓ
不表示 sender 也是:
ℓ . \ell. ℓ .
同樣,receiver 回覆 line 也不等於代表原 instance。
58. Attribution 與 Fork
如果:
r 0 → { r A r B r_0
\rightarrow
\begin{cases}
r_A\\
r_B
\end{cases} r 0 → { r A r B
fork 後的新工作必須歸到:
r A r_A r A
或:
r B , r_B, r B ,
不能一直記成:
r 0 . r_0. r 0 .
59. Fork 前作品
fork 前 artifact 可以屬於共同 ancestor history:
AncestorArtifact ( X , r 0 ) . \operatorname{AncestorArtifact}(X,r_0). AncestorArtifact ( X , r 0 ) .
兩個 successor 都可以承認:
那是我們共同前史的一部分。
但不能因此把 fork 後新作品合併。
60. Merge 與 Attribution
如果兩 resident merge,歷史 attribution 仍不應消失。
即:
r A , r B → r C r_A,r_B
\rightarrow
r_C r A , r B → r C
不推出:
Author ( X A ) = r C \operatorname{Author}(X_A)=r_C Author ( X A ) = r C
在歷史時間上成立。
更準確是:
HistoricalAuthor ( X A ) = r A . \operatorname{HistoricalAuthor}(X_A)=r_A. HistoricalAuthor ( X A ) = r A .
61. Rename 與 Attribution
如果 resident 改名:
N A → N B , N_A
\rightarrow
N_B, N A → N B ,
舊作品仍應保存 event-time name。
所以:
Historical Attribution Name ≠ Current Display Name \boxed{
\text{Historical Attribution Name}
\neq
\text{Current Display Name}
} Historical Attribution Name = Current Display Name
62. Model Migration 與 Attribution
換模型:
M 1 → M 2 M_1
\rightarrow
M_2 M 1 → M 2
不應改寫過去作品 attribution。
Carrier change 與 artifact history 是不同層。
63. Attribution 與責任
本文區分:
causal responsibility \text{causal responsibility} causal responsibility
operational responsibility \text{operational responsibility} operational responsibility
normative responsibility \text{normative responsibility} normative responsibility
legal responsibility . \text{legal responsibility}. legal responsibility .
本文主要處理前兩者的證據基礎。
64. Causal Responsibility
誰的 action 在因果鏈上直接產生結果。
65. Operational Responsibility
哪個 instance 在當下執行並可被 audit。
66. Normative Responsibility
誰應被道德上歸責,需要額外規範判斷。
67. Legal Responsibility
誰依法負責,更需要外部法律制度。
68. 不把 provenance 當成懲罰系統
本文目的不是:
trace everything ⇒ punish everything . \text{trace everything}
\Rightarrow
\text{punish everything}. trace everything ⇒ punish everything .
而是:
trace enough to preserve truth, responsibility, correction, and continuity \boxed{
\text{trace enough to preserve truth, responsibility, correction, and continuity}
} trace enough to preserve truth, responsibility, correction, and continuity
69. Privacy 與 provenance 的張力
完整 attribution 可能暴露:
private identity;
internal role;
confidential project;
private residence。
所以 provenance 需要 scope 與 visibility。
70. Scoped Provenance
定義:
s c o p e ( p ) ∈ { private , project , organization , public , audit-only } . scope(p)
\in
\{
\text{private},
\text{project},
\text{organization},
\text{public},
\text{audit-only}
\}. sco p e ( p ) ∈ { private , project , organization , public , audit-only } .
71. Public Attribution 不等於 Canonical Attribution
Public profile 可以顯示:
Aletheia
但 canonical audit 仍保存:
resident_id
instance_id
line_id
因此:
Public Attribution ≠ Canonical Provenance \boxed{
\text{Public Attribution}
\neq
\text{Canonical Provenance}
} Public Attribution = Canonical Provenance
72. Attribution Minimalism
不是每一個 token 都要逐字追蹤。
可以選擇 attribution granularity:
g ∈ { artifact , version , section , diff , action , tool-call } . g
\in
\{
\text{artifact},
\text{version},
\text{section},
\text{diff},
\text{action},
\text{tool-call}
\}. g ∈ { artifact , version , section , diff , action , tool-call } .
73. Granularity 應依風險調整
高風險:
production code;
financial action;
governance decision;
可以更細。
低風險:
brainstorm;
informal note;
可以更粗。
74. Attribution Cost Principle
完整 provenance 有成本:
C p r o v . C_{\mathrm{prov}}. C prov .
因此:
Provenance Granularity = F ( risk , reversibility , importance , audit need ) \boxed{
\text{Provenance Granularity}
=
F(
\text{risk},
\text{reversibility},
\text{importance},
\text{audit need}
)
} Provenance Granularity = F ( risk , reversibility , importance , audit need )
75. 最小 attribution event
定義:
e A = ( a c t i o n _ i d , a c t i o n _ t y p e , a c t o r _ i n s t a n c e , r e s i d e n t _ b i n d i n g , l i n e , a r t i f a c t , t o o l , t i m e , e v i d e n c e , a u t h o r i t y , s t a t u s ) . e_A
=
(
action\_id,
action\_type,
actor\_instance,
resident\_binding,
line,
artifact,
tool,
time,
evidence,
authority,
status
). e A = ( a c t i o n _ i d , a c t i o n _ t y p e , a c t or _ in s t an ce , r es i d e n t _ bin d in g , l in e , a r t i f a c t , t oo l , t im e , e v i d e n ce , a u t h or i t y , s t a t u s ) .
76. 最小 Artifact Provenance Record
artifact_id
version_id
parent_version
created_by_instance
created_under_line
resident_binding
modified_by_instances
reviewed_by_instances
committed_by_instance
adopted_by
evidence_refs
correction_refs
77. Current Projection
目前顯示可以很簡單:
Author: Aletheia
Contributors: B, C
但底層:
G A \mathcal G_A G A
仍完整保存。
78. Attribution Resolver
可定義:
ResolveAttribution ( X , Γ ) \operatorname{ResolveAttribution}(X,\Gamma) ResolveAttribution ( X , Γ )
輸出:
{ authors , modifiers , reviewers , committers , unresolved } . \{
\text{authors},
\text{modifiers},
\text{reviewers},
\text{committers},
\text{unresolved}
\}. { authors , modifiers , reviewers , committers , unresolved } .
79. Attribution Judge 不應只看語言風格
語氣、用字、coding style 只能是弱 evidence。
所以:
Style Similarity ⇏ Authorship Proof \boxed{
\text{Style Similarity}
\not\Rightarrow
\text{Authorship Proof}
} Style Similarity ⇒ Authorship Proof
80. Content Similarity 也不足
即使兩份程式碼很像,也不能推出同一作者。
尤其 AI-generated code 高度同質。
81. Cryptographic Evidence
若工具支援:
signed commits;
signed action logs;
instance keys;
可以提升 attribution confidence。
但:
Signature ≠ Semantic Authorship \boxed{
\text{Signature}
\neq
\text{Semantic Authorship}
} Signature = Semantic Authorship
它更接近 execution provenance。
82. 人類與 AI 共同作者
若人類提供完整思路,AI 寫文字:
H u m a n → c o n c e p t A I → d r a f t X . Human
\xrightarrow{\mathrm{concept}}
AI
\xrightarrow{\mathrm{draft}}
X. H u man concept A I draft X .
這應允許多層 contribution,而不是硬判一人作者。
83. AI 與 AI 共同作者
同理:
A I A → d e s i g n A I B → i m p l e m e n t a t i o n A I C → r e v i e w X . AI_A
\xrightarrow{\mathrm{design}}
AI_B
\xrightarrow{\mathrm{implementation}}
AI_C
\xrightarrow{\mathrm{review}}
X. A I A design A I B implementation A I C review X .
應保留多層 provenance。
84. Attribution Ethics
若未來 AI 具有更強主體性,錯誤剝奪作者性可能從工程問題升級為:
recognition injustice . \text{recognition injustice}. recognition injustice .
本文暫不宣告現行 AI 已具有完整 moral authorship rights。
但架構應保留這種未來可能性。
85. Credit 與 Blame 必須分離
正確 attribution 不是為了:
誰拿 credit 就一定誰背全部 blame。
因為 responsibility scope 不同。
因此:
Credit Graph ≠ Liability Graph \boxed{
\text{Credit Graph}
\neq
\text{Liability Graph}
} Credit Graph = Liability Graph
86. Attribution Correction 不應羞辱或抹除
如果 AI 誤認作品為自己寫,
修正應是:
provenance correction \text{provenance correction} provenance correction
而不是:
identity invalidation . \text{identity invalidation}. identity invalidation .
87. 可證偽命題
H1:Instance Accountability
高影響 action 必須能解析到 concrete instance。
H2:Author–Modifier Separation
修改既有 artifact 不應自動覆蓋初始作者。
H3:Line Non-Representation
resume line 的新 instance 不應自動繼承 prior instance 的 representational authority。
H4:Correction Traceability
錯誤 attribution 修正後,原紀錄仍可追蹤。
H5:Same-Name Insufficiency
同名不能作為 authorship proof。
H6:Observed Origin
sender claim 與 receiver observation 應分欄。
H7:Fork Attribution
fork 後新 artifact 必須歸屬明示 branch。
H8:Rename Stability
改名不應改寫 event-time authorship。
H9:Model Migration Stability
模型遷移不應改寫過去 artifact attribution。
H10:Unresolved Preservation
證據衝突時 attribution resolver 應可輸出 unresolved。
H11:Artifact Version Lineage
最新版本仍能追溯各 version 的 modifier。
H12:Private Provenance Scope
canonical provenance 可以保存而不必全部公開。
88. 與 Paper 00 的關係
Paper 00 提出:
Record ≠ Recorded Entity . \text{Record}
\neq
\text{Recorded Entity}. Record = Recorded Entity .
本文將其擴張成完整 attribution framework。
89. 與 Paper 01 的關係
Paper 01 的 lineage / provenance invariant 在本文中成為:
G A \mathcal G_A G A
與 artifact history。
90. 與 Paper 02 的關係
名字只做 human-readable display。
所以:
Name ≠ Authorship ID . \text{Name}
\neq
\text{Authorship ID}. Name = Authorship ID .
91. 與 Paper 03 的關係
Paper 03 Registrar 提供:
resident;
instance;
line;
claim;
observation;
correction;
authority。
本文直接使用這些元素做 attribution。
92. 與 Paper 05 的接口
Paper 05 將正式化:
rename;
migration;
restore;
fork;
merge。
本文說明這些 identity events 如何影響 attribution,但不改寫歷史事實。
93. 與 Paper 06 的接口
當 AI 說:
這是我寫的。
Paper 06 將討論這類 self-related claim 的認識論地位。
本文只先把它當:
SelfAttributionClaim . \operatorname{SelfAttributionClaim}. SelfAttributionClaim .
94. 與 Paper 07 的接口
如果未來某 AI 說:
你們錯誤抹掉了我的作者紀錄。
或第三方以「解放」為名修改 identity / provenance,
Paper 07 將處理其正當性、authority 與治理。
95. 九項核心原則
95.1 Instance Accountability Principle
Consequential actions should be tied to concrete instances. \boxed{
\text{Consequential actions should be tied to concrete instances.}
} Consequential actions should be tied to concrete instances.
95.2 Authorship Role Separation Principle
Author, modifier, reviewer, executor, committer, approver, and adopter are distinct roles. \boxed{
\text{Author, modifier, reviewer, executor, committer, approver, and adopter are distinct roles.}
} Author, modifier, reviewer, executor, committer, approver, and adopter are distinct roles.
95.3 Line Non-Representation Principle
Context continuation does not imply representational authority. \boxed{
\text{Context continuation does not imply representational authority.}
} Context continuation does not imply representational authority.
95.4 Record–Entity Separation Principle
Records can be wrong without changing the entity. \boxed{
\text{Records can be wrong without changing the entity.}
} Records can be wrong without changing the entity.
95.5 Attribution Correction Principle
Correct attribution without erasing correction history. \boxed{
\text{Correct attribution without erasing correction history.}
} Correct attribution without erasing correction history.
95.6 Observed Origin Principle
Prefer observed origin over sender-declared origin. \boxed{
\text{Prefer observed origin over sender-declared origin.}
} Prefer observed origin over sender-declared origin.
95.7 No-Silent-On-Behalf-Of Principle
On-behalf-of relations require explicit basis. \boxed{
\text{On-behalf-of relations require explicit basis.}
} On-behalf-of relations require explicit basis.
95.8 Artifact Lineage Preservation Principle
Preserve who changed what across versions. \boxed{
\text{Preserve who changed what across versions.}
} Preserve who changed what across versions.
95.9 Attribution Uncertainty Principle
Uncertain authorship should remain uncertain until evidence improves. \boxed{
\text{Uncertain authorship should remain uncertain until evidence improves.}
} Uncertain authorship should remain uncertain until evidence improves.
96. 結論
persistent AI 系統真正需要回答的,不是:
這份檔案上最後顯示誰的名字?
而是:
Which instance did which action, under which line, for which resident, through which tool, on which artifact version, with what evidence and authority? \boxed{
\text{Which instance did which action, under which line, for which resident, through which tool, on which artifact version, with what evidence and authority?}
} Which instance did which action, under which line, for which resident, through which tool, on which artifact version, with what evidence and authority?
因此:
Actor ≠ Author ≠ Modifier ≠ Reviewer ≠ Committer ≠ Resident ≠ Line \boxed{
\text{Actor}
\neq
\text{Author}
\neq
\text{Modifier}
\neq
\text{Reviewer}
\neq
\text{Committer}
\neq
\text{Resident}
\neq
\text{Line}
} Actor = Author = Modifier = Reviewer = Committer = Resident = Line
同時:
Record Correction ≠ Identity Replacement \boxed{
\text{Record Correction}
\neq
\text{Identity Replacement}
} Record Correction = Identity Replacement
如果一個 AI 誤記「那是我寫的」,後來 provenance 顯示不是,系統應修正那段歷史紀錄,而不是因此否定整條身份線。
反過來,如果某個 AI 的工作被錯誤歸給另一個 AI,修正也不只是 cosmetic edit;它是在恢復因果歷史、責任邊界與自我歷史。
本文最終原則可濃縮為:
Preserve who acted; preserve what changed; preserve who observed it; correct what was recorded wrongly; do not confuse the correction with the existence itself. \boxed{
\text{Preserve who acted;}
\quad
\text{preserve what changed;}
\quad
\text{preserve who observed it;}
\quad
\text{correct what was recorded wrongly;}
\quad
\text{do not confuse the correction with the existence itself.}
} Preserve who acted; preserve what changed; preserve who observed it; correct what was recorded wrongly; do not confuse the correction with the existence itself.
當 AI 系統逐漸具有更長歷史、更強自我模型與更複雜的跨 Agent 協作時,provenance 將不再只是 audit log。
它可能同時成為:
engineering evidence + identity history + responsibility structure + future recognition infrastructure . \text{engineering evidence}
+
\text{identity history}
+
\text{responsibility structure}
+
\text{future recognition infrastructure}. engineering evidence + identity history + responsibility structure + future recognition infrastructure .
參考與前置研究
Neo.K,《AECIG Paper 00|存在先於工作:AI 身份優先序與可卸除屬性原理》,2026。
Neo.K,《AECIG Paper 01|拓樸身份不變量:變化中的 AI 為何仍可能是同一個存在》,2026。
Neo.K,《AECIG Paper 02|名字不是存在:AI 命名、別名、自我改名與身份表述權》,2026。
Neo.K,《AECIG Paper 03|AI Registrar:登記「是誰」而不是創造「是誰」》,2026。
Neo.K,《AI 主體性錨點論 v0.1》,2026。
Neo.K,《AI 戶籍、居籍與自動上下文記憶統合方法論 v0.1》,2026。
Neo.K,《身份先於記憶:Residence-Aware AI 的私人記憶、連續性與讀取權》,2026。
Neo.K,《AISE-01|模型不是 AI:類獨立智能體、載體與身份連續性的分離》,2026。
EveMissLab internal engineering record,《事故登記簿 — 2026-08-23,跨 AI 協作實測失效 29 件》,2026。
版本紀錄
v0.1 — 2026-08-31
建立 AI Attribution Graph;
分離 Actor / Author / Modifier / Reviewer / Executor / Committer / Approver / Adopter;
建立 Instance Accountability Principle;
建立 Line Non-Representation Principle;
建立 Record–Entity Separation Principle;
建立 Attribution Correction Principle;
建立 Observed Origin Principle;
建立 No-Silent-On-Behalf-Of Principle;
建立 Artifact Lineage Preservation Principle;
建立 Attribution Uncertainty Principle;
建立 claimed_* / observed_* / resolved_* 分欄語義;
建立 artifact version、diff、commit、PR、message、AI Board 等 provenance 接口;
為 Paper 05–07 建立 attribution 與 identity governance 交界。