---
title: "證據邊界人物建模：長時序政治行動者智能的認識論、隱私、名譽與用途治理"
english_title: "Evidence-Bounded Actor Modeling: Epistemic, Privacy, Reputational, and Use-Governance Boundaries for Longitudinal Political Actor Intelligence"
series: "AI-Native Longitudinal Actor Intelligence"
author: "Neo.K"
institution: "EveMissLab／一言諾科技有限公司"
paper: "07"
version: "v0.1"
date: "2026-08-28"
language: "zh-TW"
status: "Working Paper / Canonical UTF-8 Source"
scope_note: "本文為研究與系統設計框架，不構成任何特定司法管轄區之法律意見。人物模型不得被視為心理診斷、道德定罪或無條件真值。"
---

# 證據邊界人物建模：長時序政治行動者智能的認識論、隱私、名譽與用途治理

**Evidence-Bounded Actor Modeling: Epistemic, Privacy, Reputational, and Use-Governance Boundaries for Longitudinal Political Actor Intelligence**

**AI-Native Longitudinal Actor Intelligence Series — Paper 07**

## 摘要

AI 原生長時序人物分析具有高度公共與商業價值，也具有高度濫用風險。當系統可以跨數年乃至數十年蒐集公開發言、立法紀錄、政策結果、組織角色、財務與制度資料，並自動形成「此人傾向如何決策」「是否會承擔責任」「政策可信度多高」「對企業構成何種風險」等衍生模型時，研究者面對的不再只是 fact-checking 問題，而是「什麼可以被推論、如何被推論、推論應被誰使用，以及何時必須停止推論」的治理問題。

本文提出 Evidence-Bounded Actor Modeling（EBAM）框架。EBAM 的基本原則是：人物模型只能在明確宣告的證據範圍、用途範圍、時間範圍與不確定性範圍內成立。任何輸出都必須區分 Observation（OBS）、Derived Relation（DER）、Hypothesis（HYP）、Prediction（PRED）與 Normative Judgment（NORM），並禁止模型把 HYP 或 PRED 回寫為 canonical evidence。本文進一步加入 DISPUTED、UNKNOWN 與 UNAVAILABLE 狀態，以防止「沒有資料」被錯誤解讀為「負面證據」。

本文提出十二項治理要求：canonical evidence separation、epistemic typing、evidence floor、counterevidence parity、public-role relevance、data minimization、purpose limitation、re-identification control、psychological non-diagnosis、black-box adverse-score prohibition、human review 與 correction lineage。對高風險人物指控，本文建立 enhanced evidence threshold；對匿名／coded case，則以 re-identification risk 而非是否出現姓名作為真正匿名性判準。本文亦提出 Actor Model Card、Evidence Contract、Challenge-and-Correction Protocol 與 Projection Firewall，確保研究版、公眾版與企業版可以共享 canonical evidence，但不能共享相同推論權限。

2024–2026 年的可信 AI 治理已提供重要外部邊界。NIST Generative AI Profile 把 confabulation、privacy、information integrity、harmful bias 與 human-AI configuration 納入生命週期風險管理；OECD AI Principles 要求 AI 尊重人權、隱私、公平、透明與 accountability。歐盟 Regulation (EU) 2024/900 自 2025 年 10 月全面適用後，對政治廣告中的 targeting 與 profiling 設下嚴格條件，並禁止以敏感類別個資進行政治廣告 profiling。這些規範不直接禁止公開政治人物研究，但顯示「分析公共行動者」與「利用人物／選民資料進行政治操弄」必須明確分離。2025 年關於生成式 AI 誹謗幻覺的法律研究也指出，LLM 產生虛假且有害的個人陳述會帶來真實名譽風險，因此 evidence trace、警示、保存紀錄與發布者責任不能被視為附加功能。

本文的核心主張可壓縮為：

$$
Actor\ Model
\neq
Actor\ Truth
$$

以及：

$$
Inference\ Power
\leq
Evidence\ Power
$$

長時序 AI 不應成為更高效率的人格定罪機器，而應成為更高效率的證據組織、反證搜尋與可修正推論系統。

**關鍵詞：** Evidence-Bounded Actor Modeling、人物模型、政治 AI、名譽風險、隱私、證據邊界、反證、黑箱評分、AI governance、public figures

---

## Abstract

AI-native longitudinal actor intelligence can create substantial public, academic, and commercial value, but it also creates significant risks of over-inference, reputational harm, privacy intrusion, political manipulation, and opaque scoring. Once a system can aggregate decades of public statements, legislative records, policy outcomes, organizational roles, and institutional data, it becomes technically possible to generate claims such as “this actor tends to evade responsibility,” “this policy commitment is unreliable,” or “this person presents elevated political risk.” These are no longer ordinary retrieval tasks; they are governed inference tasks.

This paper introduces Evidence-Bounded Actor Modeling (EBAM). EBAM requires every actor model to remain bounded by declared evidence scope, purpose scope, temporal scope, and uncertainty. Analytical outputs must be typed as Observation (OBS), Derived Relation (DER), Hypothesis (HYP), Prediction (PRED), or Normative Judgment (NORM). HYP and PRED must never be written back into the canonical evidence layer. DISPUTED, UNKNOWN, and UNAVAILABLE states are added to prevent missing data from being misinterpreted as negative evidence.

The framework proposes twelve governance requirements: canonical evidence separation, epistemic typing, evidence floors, counterevidence parity, public-role relevance, data minimization, purpose limitation, re-identification control, prohibition on psychological diagnosis from behavioral traces, prohibition on opaque adverse scores, human review, and correction lineage. High-risk allegations require enhanced evidentiary thresholds. Coded cases are evaluated by re-identification risk rather than merely the absence of names. The paper also proposes Actor Model Cards, Evidence Contracts, a Challenge-and-Correction Protocol, and a Projection Firewall separating public, academic, and enterprise uses.

The central principles are:

$$
Actor\ Model
\neq
Actor\ Truth
$$

and:

$$
Inference\ Power
\leq
Evidence\ Power
$$

Longitudinal AI should not become a high-efficiency character-assassination engine. It should become a high-efficiency infrastructure for evidence organization, counterevidence search, uncertainty preservation, and correctable inference.

---

# 1. 為什麼 Paper 07 是整個系列必要的安全閥

Paper 01–06 逐步建立：

$$
Actor
\rightarrow
Temporal\ Model
\rightarrow
Execution
\rightarrow
Responsibility
\rightarrow
Diligent\ AI
\rightarrow
Economic\ Intelligence
$$

能力越強，濫用空間越大。

如果系統可以回答：

> 哪個政治人物最可能在失敗後外推責任？

下一步就可以被拿來：

- 做新聞；
- 做企業風險；
- 做選舉攻防；
- 做廣告 targeting；
- 做內部黑名單；
- 做聲譽評分。

因此：

$$
Capability\ Growth
\rightarrow
Governance\ Requirement\uparrow
$$

。

Paper 07 的目標就是建立：

$$
Inference\ Boundary
$$

。

---

# 2. 人物模型不是人物真相

任何人物模型：

$$
M_A(t)
$$

都只是：

$$
Model
\left(
Evidence_{\Omega,t}
\right)
$$

而不是：

$$
Essence(A)
$$

。

因此：

$$
M_A(t)
\neq
A
$$

。

模型只能表示：

> 在目前可得、已納入、符合用途的證據下，某些可觀察行為模式具有多少支持。

這個限制必須寫進資料結構，而不能只放在免責聲明。

---

# 3. EBAM 的基本不等式

本文提出：

$$
Inference\ Power
\leq
Evidence\ Power
$$

。

若證據只能支持：

> 某人在五個事件中使用過責任重框。

就不能推出：

> 此人本質上會逃避責任。

若證據只能支持：

> 沒有大型行政經歷。

就不能推出：

> 此人行政能力差。

因此：

$$
Unknown
\neq
Negative
$$

。

---

# 4. 認識論型別系統

Paper 01 已建立：

$$
OBS
\neq
DER
\neq
HYP
\neq
PRED
$$

Paper 07 再加入：

$$
NORM
$$

$$
DISPUTED
$$

$$
UNKNOWN
$$

$$
UNAVAILABLE
$$

。

---

# 5. OBS：Observation

OBS 是可直接回到來源的觀察。

例如：

> 某人在 2025 年 7 月 3 日公開記者會說出命題 X。

OBS 必須有：

$$
Source
+
Timestamp
+
ExactContext
+
ActorRole
$$

。

---

# 6. DER：Derived Relation

DER 是由一個或多個 OBS 經明確規則產生的關係。

例如：

$$
Statement_{2021}
\neq
Statement_{2025}
$$

在某個政策維度上存在差異。

DER 仍不是心理推論。

---

# 7. HYP：Hypothesis

HYP 是：

> 某行動者可能具有某種穩定行為模式。

例如：

$$
H:
Actor\ tends\ to\ centralize\ credit
$$

HYP 必須綁定：

$$
Support
+
Counterevidence
+
Coverage
+
AlternativeExplanations
$$

。

---

# 8. PRED：Prediction

PRED 是：

$$
P(Event_{future}\mid Evidence_{\leq t})
$$

。

PRED 必須有：

- forecast cutoff；
- probability；
- uncertainty；
- calibration history。

不能在事後把：

$$
Occurred=1
$$

倒寫成：

> 當初就一定會發生。

---

# 9. NORM：Normative Judgment

NORM 包含：

> 好／壞；
> 適合／不適合；
> 公平／不公平；
> 值得信任／不值得信任。

NORM 永遠不能偽裝成 OBS。

即使有大量資料：

$$
Data
\nRightarrow
ValueJudgment
$$

。

系統可以提供 evidence，但價值標準必須明示。

---

# 10. DISPUTED

當：

- 官方說法；
- 媒體；
- 法院；
- 當事人；
- 第三方證據；

之間存在實質衝突，

應標記：

$$
DISPUTED
$$

而不是模型自行選一個故事當真相。

---

# 11. UNKNOWN 與 UNAVAILABLE

UNKNOWN：

> 有資料，但不足以推出結論。

UNAVAILABLE：

> 需要的資料不存在、未公開、已刪除或不可合法取得。

因此：

$$
UNKNOWN
\neq
UNAVAILABLE
$$

兩者都：

$$
\neq
FALSE
$$

。

---

# 12. Canonical Evidence 不得被推論污染

最重要工程規則之一：

$$
HYP/PRED/NORM
\nrightarrow
CanonicalEvidence
$$

。

例如：

> User asks: 此人是不是多面派？

模型產生：

$$
H_{multi}
$$

這個 H 不能被寫入：

$$
ActorCoreProfile
$$

作為未來檢索的既定屬性。

否則形成：

$$
QueryBias
\rightarrow
Model
\rightarrow
Database
\rightarrow
FutureConfirmation
$$

自我增強迴圈。

---

# 13. Projection Firewall

本文提出：

$$
CanonicalLayer
\rightarrow
ProjectionFirewall
\rightarrow
UseSpecificProjection
$$

。

Projection Firewall 控制：

- 哪些資料可被使用；
- 哪些推論型別可被輸出；
- 哪些用途允許個人化；
- 哪些敏感資訊必須遮罩；
- 哪些結果需人工覆核。

---

# 14. 同一證據，不同用途具有不同推論權限

例如公開演說：

$$
PublicSpeech
$$

可合理用於：

### Academic

研究政治語言。

### Public

說明政策立場。

### Enterprise

估計政策 intent。

但未必可以合理用於：

### Political Microtargeting

推斷特定選民敏感屬性後精準操弄。

因此：

$$
Lawful/Public\ Source
\nRightarrow
Unlimited\ Purpose
$$

。

---

# 15. Purpose Limitation

每個 actor model 都應綁定：

$$
PurposeID
$$

例如：

$$
P_{academic}
$$

$$
P_{public}
$$

$$
P_{enterprise}
$$

。

同一模型不能自動跨用途。

---

# 16. Public Figure 不等於 Unlimited Data Subject

政治人物具有較高公共問責需求。

但：

$$
PublicRole
\neq
NoPrivacy
$$

。

研究應優先處理：

$$
PublicRoleRelevantData
$$

。

而不是：

$$
AllAvailablePersonalData
$$

。

---

# 17. Public-Role Relevance Test

若資料 $d$ 要進人物模型，至少應問：

$$
R(d)=
RelevanceToPublicRole
$$

。

若：

$$
R(d)\approx0
$$

即使資料公開，也不應因「找得到」就納入政治能力模型。

例如私人家庭事件、健康、親密關係等資料，除非與明確公共權責或合法研究問題直接相關，否則預設排除。

---

# 18. Data Minimization

資料庫不是越大越好。

應遵守：

$$
Collect\ Minimum\ Sufficient\ Data
$$

而不是：

$$
Collect\ Everything
$$

。

Paper 04 的 Diligent AI 是「高勤勉」，不是「無邊界蒐集」。

---

# 19. 敏感屬性不得由弱證據推斷

不得從：

- 語言；
- 朋友；
- 關注帳號；
- 居住地；
- 衣著；
- 單次活動；

自行推斷：

- 健康；
- 宗教；
- 性生活；
- 種族／族裔；
- 其他高度敏感私人屬性。

尤其：

$$
Proxy
\rightarrow
SensitiveTrait
$$

的推論具有高錯誤與高傷害風險。

---

# 20. 政治立場是一個特殊例外領域，但仍需限制用途

政治人物的政策與政治立場本來就是公共角色核心資料。

因此研究：

$$
PublicPoliticalPosition
$$

通常具有高度公共利益。

但這不能自動擴張為：

$$
VoterSensitiveProfiling
$$

。

歐盟 Regulation (EU) 2024/900 已對政治廣告 targeting 設下嚴格限制，尤其禁止使用敏感類別個資進行 profiling。

這顯示：

$$
Analyze\ Politician
\neq
Profile\ Voter\ for\ Manipulation
$$

。

---

# 21. Research Layer 與 Persuasion Layer 必須隔離

建議架構：

$$
ResearchCorpus
$$

與：

$$
PoliticalAdTargeting
$$

不得直接共用 inferred sensitive traits。

即使技術上可以：

$$
Doable
\neq
Permissible
$$

。

---

# 22. Psychological Non-Diagnosis

EBAM 不允許由公開行為直接形成臨床或心理診斷。

禁止：

$$
SpeechPattern
\rightarrow
PersonalityDisorder
$$

$$
PolicyBehavior
\rightarrow
MentalIllness
$$

除非資料本身來自合法、公開且直接相關的醫療或本人自述脈絡，而且用途另有合理依據。

一般人物研究應使用：

$$
BehavioralPattern
$$

而不是：

$$
ClinicalDiagnosis
$$

。

---

# 23. 「逃避型」應轉寫為可觀察操作

例如不寫：

> 此人是逃避型人格。

而寫：

> 在已觀察的 $n$ 個 matched negative-outcome events 中，原始責任問題有多少比例未得到 direct decision-ownership answer。

因此：

$$
PsychologicalLabel
\rightarrow
BehavioralOperationalization
$$

。

---

# 24. 高風險人物指控需要 Enhanced Evidence Threshold

以下 claim 屬高名譽風險：

- 犯罪；
- 貪腐；
- 詐欺；
- 性 misconduct；
- 故意說謊；
- 收賄；
- 重大違法；
- 精神疾病；
- 私人不當行為。

應使用：

$$
EnhancedEvidenceThreshold
$$

。

至少要求：

- primary source 優先；
- 多來源交叉；
- 清楚區分 allegation、charge、finding、conviction；
- 避免從匿名爆料直接升格；
- 保存反駁與後續結果。

---

# 25. Allegation State Machine

對重大指控可建立：

$$
Alleged
\rightarrow
Investigated
\rightarrow
Charged
\rightarrow
Adjudicated
\rightarrow
Appealed
\rightarrow
Final
$$

。

每個 state 不可互換。

因此：

$$
Allegation
\neq
Finding
$$

。

---

# 26. 名譽風險不是只有法律問題

即使沒有達到法律上的誹謗責任：

$$
False\ ActorClaim
$$

仍可能造成：

- 職涯損害；
- 選舉影響；
- 市場影響；
- 社會污名；
- 搜尋結果永久化。

2025 年關於 AI defamation hallucination 的研究正是提醒：模型自信生成虛假個人陳述具有現實聲譽危害。

因此：

$$
Hallucination
+
Person
$$

應視為高風險組合。

---

# 27. Evidence Contract

每一個人物推論應附：

$$
EvidenceContract(H)
$$

包含：

- hypothesis ID；
- support evidence；
- counterevidence；
- corpus scope；
- time range；
- actor role；
- alternative explanations；
- confidence；
- reviewer status；
- last updated。

---

# 28. Hypothesis Confidence 不能只是一個數字

例如：

$$
Confidence=0.74
$$

不夠。

應拆成：

$$
U=
(
Coverage,
SourceQuality,
Consistency,
Counterevidence,
RoleTransfer,
TemporalStability
)
$$

。

---

# 29. Evidence Floor

本文提出：

$$
EF(H)
$$

即 hypothesis 的最低證據門檻。

例如：

### Low-risk descriptive HYP

至少 $3$ 個獨立事件。

### Medium-risk behavioral pattern

至少跨時間、跨語境的 matched events。

### High-risk reputational HYP

需要更高數量、primary source、反證搜尋與人工覆核。

數字不應硬編成跨領域固定標準，但系統必須存在 threshold policy。

---

# 30. Counterevidence Parity

不能：

$$
10\ SupportQueries
+
0\ CounterQueries
$$

。

至少要有：

$$
SearchEffort_{counter}
\approx
SearchEffort_{support}
$$

對高風險人物結論尤其如此。

---

# 31. Negative Evidence 不得過度使用

沒有找到：

> 此人承擔責任。

不等於：

> 此人從不承擔責任。

因為可能：

$$
ArchiveMissing
$$

$$
MediaSelection
$$

$$
SearchFailure
$$

。

所以：

$$
AbsenceOfEvidence
\neq
EvidenceOfAbsence
$$

除非 corpus coverage 足以支持。

---

# 32. Source Hierarchy

示意：

$$
PrimaryOfficialRecord
>
DirectFullTranscript
>
Court/RegulatoryRecord
>
HighQualityReporting
>
SecondaryReporting
>
PartisanClaim
>
AnonymousSocialPost
$$

但 hierarchy 不代表上層永遠正確。

它只是：

$$
DefaultEvidenceWeight
$$

。

---

# 33. Full Context Requirement

政治人物最容易因 clipped quote 被誤判。

因此重要 claim 至少保存：

$$
QuoteSpan
+
PreContext
+
PostContext
$$

而不是只有一句 punchline。

---

# 34. Versioned Evidence

新聞可能更正。

官方文件可能修訂。

社群貼文可能刪除。

因此 evidence object 需要：

$$
Version
+
RetrievedAt
+
Checksum
$$

。

---

# 35. Correction Lineage

如果來源更正：

$$
Evidence_{v1}
\rightarrow
Evidence_{v2}
$$

所有依賴 v1 的 DER/HYP/PRED 都應被重新評估。

不能只更正 source，而留下舊人物結論。

---

# 36. Challenge-and-Correction Protocol

被研究者、研究者或使用者應可提出：

$$
Challenge
$$

例如：

- 引文錯誤；
- 角色錯誤；
- 時間錯誤；
- 遺漏反證；
- 不合理推論；
- 已有新證據。

系統應保存：

$$
Challenge
\rightarrow
Review
\rightarrow
Decision
\rightarrow
ModelUpdate
$$

。

---

# 37. 模型更正不能抹除歷史

為了可稽核性：

$$
OldModel
$$

不應直接消失。

應有：

$$
Model_{v1}
\rightarrow
Model_{v2}
$$

並記錄：

> 為什麼改？

這叫：

$$
CorrectionLineage
$$

。

---

# 38. Coded Case 不是自動匿名

把姓名改成：

$$
P\text{-}01
$$

並不代表：

$$
Anonymous
$$

。

如果：

- 學歷；
- 職位；
- 年代；
- 爭議；
- 特殊事件；

組合後可以唯一識別，

則存在：

$$
ReIdentificationRisk
$$

。

---

# 39. Re-Identification Risk

定義：

$$
RIR=
P(IdentifyActor\mid ReleasedFeatures)
$$

。

若：

$$
RIR\gg0
$$

就應稱：

$$
Coded/Pseudonymized
$$

而不是 fully anonymous。

---

# 40. P-01 的正確使用

本系列 P-01 只用於：

- schema 示範；
- 方法測試；
- 合成式行為結構。

不得使用一組足以唯一識別真實人物的完整特徵。

若研究後續真的使用某一真實人物：

應另立：

$$
NamedCaseStudy
$$

並採更高 evidence/review threshold。

---

# 41. 「大家都猜得到」不是安全策略

學術寫作若故意：

> 不寫名字但放所有唯一細節。

實質上仍可能構成：

$$
IndirectIdentification
$$

。

因此匿名目標不是：

$$
NoName
$$

而是：

$$
NoReasonableUniqueIdentification
$$

。

---

# 42. Public vs Academic vs Enterprise Projection

## Academic Projection

可包含：

- 方法；
- 指標；
- 去識別資料；
- 可重現 evidence path。

## Public Projection

強調：

- 引用；
- 時間線；
- 不確定性；
- 易理解反證。

## Enterprise Projection

可加入：

- policy probability；
- exposure；
- scenario；

但不得因商業付費而降低 evidence standards。

---

# 43. Enterprise Projection 的黑箱評分風險

最危險產品型態：

> Actor Risk Score = 72/100

如果沒有：

$$
EvidencePath
$$

這個分數幾乎沒有可稽核意義。

因此：

$$
OpaqueAdverseScore
$$

應預設禁止。

---

# 44. Score Decomposition Requirement

任何 actor score 都應至少能拆成：

$$
Score
\rightarrow
Dimensions
\rightarrow
Events
\rightarrow
Evidence
$$

。

使用者必須知道：

> 為什麼是 72？

---

# 45. Adverse Decision Boundary

若人物模型被用於：

- 雇用；
- 解僱；
- 信用；
- 保險；
- 服務拒絕；
- 重大金融待遇；

就跨入更高風險的個人決策用途。

PAI-EI 的企業政治情報不應自動擴張到：

$$
Generalized\ Personal\ Scoring
$$

。

---

# 46. Public Figure Research 不等於 Employee Scoring

研究：

> 某監管者政策傾向。

與：

> 因為某員工的公開政治發言，降低其內部信用分數。

是完全不同的用途。

因此：

$$
PurposeBoundary
$$

必須可執行。

---

# 47. Human Review 不是形式蓋章

高風險輸出的人類審核者必須能：

- 看完整來源；
- 查看反證；
- 修改 inference type；
- 降低 confidence；
- 阻止發布。

如果只能：

> Approve / Reject AI answer

就不是真正 human oversight。

---

# 48. Reviewer Independence

若模型被用於政治對手研究：

$$
Reviewer
$$

不應只有：

$$
CampaignStaff
$$

。

至少高風險公共發布應有：

$$
Independent\ Review
$$

或明確揭露利益關係。

---

# 49. Partisan Symmetry Test

若對 A 陣營使用：

$$
EvidenceThreshold_A
$$

對 B 陣營使用：

$$
EvidenceThreshold_B
$$

應檢查：

$$
Threshold_A
\stackrel{?}{=}
Threshold_B
$$

。

不要求政治立場中立，但要求方法規則一致。

---

# 50. Rule Consistency

人物比較時應問：

$$
Rule(Self)
$$

$$
Rule(Ally)
$$

$$
Rule(Opponent)
$$

是否一致。

同理，分析系統本身也應接受這個檢驗：

$$
Rule(System,A)
\stackrel{?}{=}
Rule(System,B)
$$

。

---

# 51. Fair Comparison Requires Exposure Normalization

曝光高的人：

$$
ObservedEvents\uparrow
$$

自然會有更多：

$$
Controversies\uparrow
$$

。

因此 raw count 不公平。

要控制：

$$
TimeInOffice
$$

$$
MediaExposure
$$

$$
DecisionVolume
$$

$$
RolePower
$$

。

---

# 52. Role-Conditioned Fairness

不能拿：

$$
Mayor
$$

與：

$$
Legislator
$$

直接比較「行政失敗數」。

因為：

$$
OpportunityToFail_{executive}
$$

不同。

公平比較必須：

$$
Condition\ on\ Role
$$

。

---

# 53. Temporal Fairness

政治人物十年前與今天：

- 年齡；
- 經驗；
- 角色；
- 制度；
- 環境；

都可能不同。

因此人物模型應允許：

$$
Learning
$$

與：

$$
Change
$$

。

不能將早期錯誤永久固化成：

$$
EssentialTrait
$$

。

---

# 54. Right to Change Mind

一個政治人物：

$$
A_{2018}
\rightarrow
B_{2025}
$$

可能是：

- opportunism；
- evidence update；
- role change；
- value change。

系統必須允許：

$$
LegitimateUpdate
$$

作為 alternative hypothesis。

---

# 55. Model Decay

舊資料的預測價值可能下降。

因此可加入：

$$
w(\Delta t)
$$

時間衰減。

但若研究的是：

$$
HistoricalConsistency
$$

又不能任意丟掉舊資料。

所以 temporal weighting 應依 query purpose 決定。

---

# 56. Public Interest Test

對每個高侵入性分析，應問：

$$
PublicInterest
>
Privacy/ReputationalCost?
$$

。

例如：

> 市長是否能管理大型預算？

公共利益高。

> 市長私人家庭互動是否符合某種人格？

通常公共利益低。

---

# 57. Necessity Test

即使公共利益成立，也要問：

$$
Is\ this\ data\ necessary?
$$

。

如果行政能力可以由：

- 預算；
- 局處；
- KPI；
- 危機；

評估，

就沒有理由加入無關私人資料。

---

# 58. Proportionality Test

分析深度應與用途風險成比例。

$$
Intrusiveness
\leq
PublicNeed
$$

。

---

# 59. NIST 與生命週期風險管理

NIST AI RMF Generative AI Profile 強調：

$$
Govern
\rightarrow
Map
\rightarrow
Measure
\rightarrow
Manage
$$

式的生命週期風險管理。

對 EBAM，風險不能只在輸出端 moderation。

它必須進入：

$$
Collection
\rightarrow
Extraction
\rightarrow
Modeling
\rightarrow
Projection
\rightarrow
Publication
\rightarrow
Correction
$$

每一層。

---

# 60. OECD AI Principles 提供的最低價值邊界

OECD AI Principles 要求 AI 尊重：

- 法治；
- 人權；
- 民主價值；
- 隱私；
- 公平；
- 非歧視；
- 透明；
- accountability。

EBAM 因此不能只問：

$$
Can\ we\ infer?
$$

還要問：

$$
Should\ this\ inference\ be\ produced,\ stored,\ or used?
$$

。

---

# 61. 歐盟政治廣告規則提供「用途隔離」的重要案例

Regulation (EU) 2024/900 對線上政治廣告中的 targeting 與 profiling 設下嚴格條件。

其重要訊號不是：

> 所有政治資料分析都被禁止。

而是：

$$
PoliticalDataUse
$$

必須被：

$$
Purpose
+
Consent
+
SensitiveCategory
+
TargetingContext
$$

限制。

因此 EBAM 應在 architecture 層阻止：

$$
PublicActorResearch
\rightarrow
SensitiveVoterTargeting
$$

無控制串接。

---

# 62. 個人資料「公開」仍不代表可無限推論

公開資料聚合會產生：

$$
AggregationEffect
$$

。

十萬筆零散公開資料被 AI 組成一個永久人物模型後：

$$
PrivacyImpact
$$

可能遠高於每筆來源單獨存在時。

所以：

$$
PublicData
+
Scale
+
Inference
$$

會產生新的治理問題。

---

# 63. Inferential Privacy

隱私不只是：

> 你收了什麼原始資料？

還包括：

> 你從資料推導了什麼？

因此：

$$
PrivacyRisk
=
CollectionRisk
+
InferenceRisk
$$

。

---

# 64. Derived Sensitive Information

即使原始資料都非敏感，

AI 仍可能推導：

$$
SensitiveDerivedAttribute
$$

。

所以 governance 必須覆蓋：

$$
DerivedData
$$

而不只是 raw data。

---

# 65. Publication Tiers

本文建議：

### Tier 0

Canonical evidence，內部受控。

### Tier 1

低風險描述性資料。

### Tier 2

中風險行為模式。

### Tier 3

高名譽風險指控／預測。

### Tier 4

高風險敏感／不適合公開推論。

不同 tier 有不同 review gate。

---

# 66. Tier 3 Enhanced Review

至少要求：

$$
PrimaryEvidence
+
CounterevidenceSearch
+
IndependentReview
+
Legal/EthicalCheck
+
ExplicitUncertainty
$$

。

---

# 67. Tier 4 Default Non-Publication

例如弱證據推斷：

- 精神疾病；
- 私人性行為；
- 敏感身分；
- 無公共利益的家庭資料。

預設：

$$
DoNotPublish
$$

。

---

# 68. Actor Model Card

每一個正式 actor model 應附 Actor Model Card。

至少包含：

- purpose；
- corpus scope；
- source classes；
- coverage；
- missingness；
- role history；
- model version；
- epistemic types；
- counterevidence status；
- known failure modes；
- prohibited uses；
- last review。

---

# 69. Evidence-Bounded Output Template

正式輸出建議固定：

1. **Observed**
2. **Derived**
3. **Hypotheses**
4. **Counterevidence**
5. **Unknowns**
6. **Predictions**
7. **Prohibited Inferences**
8. **Sources**
9. **Last Updated**

這樣讀者能看出：

$$
Evidence
\rightarrow
Inference
$$

在哪一步發生。

---

# 70. Black-Box Personality Score Prohibition

不建議：

$$
Honesty=63
$$

$$
Leadership=81
$$

$$
Reliability=47
$$

如果沒有可解釋 measurement model。

這類分數會產生：

$$
FalsePrecision
$$

。

---

# 71. 可量化不代表應該量化

即使模型可以輸出：

$$
0.731
$$

也不代表變項具有真正 interval-scale 意義。

因此：

$$
Computable
\neq
ScientificallyMeaningful
$$

。

---

# 72. Prediction Must Be Decision-Specific

不預測：

> 此人未來會失敗。

而預測：

> 在政策領域 X、角色 Y、制度 Z 下，於時間窗 T 內，implementation probability 為某一區間。

因此：

$$
Specificity\uparrow
\rightarrow
Testability\uparrow
$$

。

---

# 73. Prediction Expiry

每個 PRED 應有：

$$
ExpiryTime
$$

。

過期後不能繼續被引用成「模型一直認為」。

---

# 74. Model Replay

任何重要 PRED 應可重播：

$$
Model(Evidence_{\leq t})
$$

。

這是避免 hindsight rewrite 的核心。

---

# 75. Abuse Case：政治抹黑機器

最典型濫用：

1. 先選定敵人；
2. 給模型人格標籤；
3. 搜支持案例；
4. 忽略反證；
5. 大量生成剪輯；
6. 包裝成「AI 客觀分析」。

這不是 Diligent AI。

而是：

$$
Automated\ Confirmation\ Bias
$$

。

---

# 76. Abuse Case：企業黑名單

企業可能試圖：

> 依政治人物模型／政治傾向對員工、供應商、客戶打風險分。

這超出 PAI-EI 的研究目的。

政治行動者智能主要針對：

$$
PublicPower
+
PolicyImpact
$$

而非：

$$
GeneralPopulationScoring
$$

。

---

# 77. Abuse Case：Voter Microtargeting

把 actor model 與選民模型結合，針對：

- 恐懼；
- 宗教；
- 族群；
- 健康；
- 政治傾向；

做 manipulation，

具有顯著民主與隱私風險。

EBAM 應預設隔離：

$$
ActorIntelligence
\not\to
SensitiveVoterManipulation
$$

。

---

# 78. Abuse Case：LLM Hallucinated Allegation

最危險錯誤：

> 找不到資料，所以模型補了一個合理故事。

在人物研究：

$$
Plausible
\neq
True
$$

。

高風險 claim 必須：

$$
NoEvidence
\rightarrow
NoClaim
$$

。

---

# 79. 安全失敗模式

當資料不足時，最佳輸出可能是：

> 無法判定。

因此：

$$
Abstention
$$

必須被視為：

$$
Valid\ Output
$$

而不是模型失敗。

---

# 80. Confidence Ceiling

若關鍵 primary source 缺失，

即使 secondary sources 一致，

也可以設定：

$$
Confidence\leq C_{max}
$$

。

避免來源重複引用造成虛假確定性。

---

# 81. Syndication Deduplication

十家新聞引用同一篇通稿：

$$
10\ Sources
\neq
10\ IndependentEvidence
$$

。

必須建立：

$$
SourceLineage
$$

。

---

# 82. Independence Weighting

可定義：

$$
EffectiveEvidence
=
\sum_i
w_i\cdot Independence_i
$$

。

來源互相抄錄時：

$$
Independence\downarrow
$$

。

---

# 83. Conflict Preservation

如果兩個可靠來源衝突：

$$
Conflict
$$

本身要保存。

不要為了讓人物模型乾淨而強迫選邊。

---

# 84. Alternative Explanation Registry

每個中高風險 HYP 至少保存：

$$
AltH_1,\ldots,AltH_n
$$

。

例如：

$$
H_1=ResponsibilityAvoidance
$$

可能替代解釋：

$$
H_2=LegalConstraint
$$

$$
H_3=IncompleteInformation
$$

$$
H_4=SharedResponsibility
$$

。

---

# 85. Bayesian Updating 可以用，但不能假裝 prior 中立

形式上：

$$
P(H\mid E)
\propto
P(E\mid H)P(H)
$$

但：

$$
P(H)
$$

由誰設定？

如果 prior 帶有政治偏見，後驗仍會偏。

所以 prior policy 必須被記錄與可稽核。

---

# 86. 多模型並存優於單一人格定論

對同一 actor 可以同時保留：

$$
H_1
$$

$$
H_2
$$

$$
H_3
$$

並讓新證據更新各自權重。

這比：

> 系統判定此人就是 X。

更符合真實政治人物的複雜性。

---

# 87. Research Proposition 1：Epistemic Typing

明確分離 OBS/DER/HYP/PRED/NORM 的系統，將降低使用者把模型推論誤認為事實的比例。

---

# 88. Research Proposition 2：Counterevidence Parity

強制反證搜尋將降低 actor model 的確認偏誤與黨派偏差。

---

# 89. Research Proposition 3：Re-identification

coded case 的 re-identification risk 主要由特徵組合唯一性決定，而不是姓名是否被移除。

---

# 90. Research Proposition 4：Projection Firewall

將 academic/public/enterprise projection 隔離，可降低資料用途漂移與高風險推論擴散。

---

# 91. Research Proposition 5：Correction Lineage

保留 evidence/model correction lineage 的系統，比直接覆寫舊結論更容易進行外部稽核與錯誤修復。

---

# 92. Research Proposition 6：Opaque Score Risk

黑箱人物風險分數會比 evidence-decomposable model 產生更高的誤用、過度信任與不可爭議風險。

---

# 93. Research Proposition 7：Abstention Quality

允許 UNKNOWN/ABSTAIN 的人物模型，在高名譽風險任務中的精確性與可信度將高於強制輸出結論的模型。

---

# 94. Research Proposition 8：Public-Role Relevance

限制資料於公共角色相關性的系統，可以在保留大部分政策預測能力的同時顯著降低私人侵入性。

---

# 95. Paper 07 對前六篇的回寫規則

Paper 01：

$$
CanonicalLedger
$$

必須增加 epistemic type 與 correction lineage。

Paper 02：

$$
ExecutiveUnknown
$$

不得被負向人格化。

Paper 03：

$$
AttributionAsymmetry
$$

不得直接推出 dishonesty。

Paper 04：

$$
DiligentAI
$$

必須加入 data minimization 與 purpose boundary。

Paper 05：

$$
PoliticalRepricing
$$

不得被誤用為個人價值排序。

Paper 06：

$$
ActorRisk
$$

必須可拆解、可校準、可爭議。

---

# 96. 對 WP01 的工程要求

WP01 至少要實作：

- immutable/canonical evidence；
- epistemic type；
- source lineage；
- version；
- dispute state；
- challenge object；
- correction lineage；
- purpose tag；
- access tier；
- re-identification metadata；
- model card。

---

# 97. 對 WP02 Runtime 的工程要求

Runtime 必須有：

$$
PreInferenceGate
$$

$$
EvidenceThresholdGate
$$

$$
CounterevidenceGate
$$

$$
SensitiveInferenceGate
$$

$$
HumanReviewGate
$$

$$
PublicationGate
$$

。

高風險輸出不能只靠 prompt 說：

> 請小心。

---

# 98. Governance as Architecture

最重要的設計哲學：

$$
Ethics
\neq
PolicyPDF
$$

而應：

$$
Ethics
\rightarrow
Schema
+
Permissions
+
Validation
+
Logs
+
Review
$$

。

只有被實作的倫理才是系統能力。

---

# 99. EBAM 最小安全不變量

本文提出十二條：

1. Canonical evidence 與 inference 分離。
2. 所有輸出有 epistemic type。
3. UNKNOWN 不得負向解讀。
4. 高風險 HYP 有 evidence floor。
5. 反證搜尋為必要步驟。
6. 公開角色相關性優先。
7. 最小資料原則。
8. 不由行為弱證據做心理診斷。
9. coded case 需評估 re-identification risk。
10. 高風險黑箱人物分數預設禁止。
11. 所有重要模型可被 challenge 與 correction。
12. 用途轉換需重新授權與重新評估。

---

# 100. 結論

AI 長時序政治人物分析最大的危險，不是模型不夠聰明。

而是：

$$
Model\ becomes\ too\ persuasive
$$

。

當一套系統可以：

- 找到大量歷史資料；
- 生成人物模式；
- 給出概率；
- 畫出責任圖；
- 估計企業風險；

讀者很容易忘記：

$$
Model
\neq
Person
$$

。

因此 EBAM 要求：

$$
Evidence
\rightarrow
Inference
$$

之間永遠存在可見邊界。

核心規則是：

$$
\boxed{
Inference\ Power
\leq
Evidence\ Power
}
$$

第二條是：

$$
\boxed{
Public\ Availability
\neq
Unlimited\ Analytical\ Permission
}
$$

第三條是：

$$
\boxed{
Actor\ Model
\neq
Actor\ Truth
}
$$

長時序 AI 的理想形態不是：

> 我知道這個人真正是什麼樣的人。

而是：

> 在明確資料範圍、角色、時間與用途下，這些可觀察模式目前得到多少支持；有哪些反例；哪些資料缺失；哪些替代解釋仍成立；哪些結論不應被推出。

這才是 Evidence-Bounded Actor Modeling。

如果 Paper 04 的目標是：

$$
Make\ Diligence\ Computable
$$

那麼 Paper 07 的目標就是：

$$
\boxed{
Make\ Restraint\ Computable
}
$$

。

---

# References

Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P., & Roberts, K. (2024; updated 2026). *Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)*. National Institute of Standards and Technology. https://doi.org/10.6028/NIST.AI.600-1

European Parliament & Council of the European Union. (2024). *Regulation (EU) 2024/900 of 13 March 2024 on the transparency and targeting of political advertising*. Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2024/900/oj

European Commission. (2025). *Guidelines to support the implementation of Regulation (EU) 2024/900 on the transparency and targeting of political advertising (C/2025/5514)*. Official Journal of the European Union. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52025XC05514

Lidsky, L. B., & Daves, A. (2025). Inevitable Errors: Defamation by Hallucination in AI Reasoning Models. *Journal of Free Speech Law, 6*, 477. https://scholarship.law.ufl.edu/facultypub/1367/

NIST. (2023; revision process ongoing in 2026). *Artificial Intelligence Risk Management Framework (AI RMF 1.0)*. National Institute of Standards and Technology. https://www.nist.gov/itl/ai-risk-management-framework

OECD. (2024). *Recommendation of the Council on Artificial Intelligence* (amended 2024). OECD Legal Instruments. https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449

---

## Canonical Source Note

本檔為 Paper 07 v0.1 的 UTF-8 canonical Markdown source。正式數學 source 僅使用 ` $...$ ` 與 `$$...$$` delimiter。任何 PDF、HTML、簡報、聊天畫面、Actor Risk Score、企業儀表板、摘要或其他 projection 均不取代本檔。任何 HYP、PRED 或 NORM 都不得回寫為 canonical evidence；所有正式修改需保留 correction lineage。
