title: "勤人 AI 政治/經濟情報執行環境:持續攝取、Agent 編排、反證、預測、暴露傳播與治理閉環" english_title: "Diligent AI Political/Economic Intelligence Runtime: Continuous Ingestion, Agent Orchestration, Counterevidence, Forecasting, Exposure Propagation, and Governance Loops" series: "AI-Native Longitudinal Actor Intelligence" author: "Neo.K" institution: "EveMissLab/一言諾科技有限公司" artifact: "Technical Whitepaper 02" version: "v0.1" date: "2026-08-28" language: "zh-TW" status: "Canonical Runtime Whitepaper / UTF-8 Source" runtime_scope: "Defines the continuous runtime operating model built on WP01 canonical architecture. It does not authorize autonomous political persuasion, sensitive voter targeting, unlawful surveillance, or automated financial trading."
勤人 AI 政治/經濟情報執行環境:持續攝取、Agent 編排、反證、預測、暴露傳播與治理閉環
Diligent AI Political/Economic Intelligence Runtime: Continuous Ingestion, Agent Orchestration, Counterevidence, Forecasting, Exposure Propagation, and Governance Loops
AI-Native Longitudinal Actor Intelligence — Technical Whitepaper 02
摘要
WP01 定義了 Temporal Actor Intelligence 中「什麼是合法狀態」:SourceArtifact、EvidenceUnit、Actor、Role、Claim、Event、Hypothesis、Prediction、PolicyForecast、Economic Exposure 與 Projection 的 canonical 分層、bitemporal semantics、provenance、Evidence Contract、Projection Firewall 與 validate-before-commit 不變量。本白皮書 WP02 進一步回答:「這套系統如何持續運作?」
本文提出 Diligent AI Political/Economic Intelligence Runtime(DAPIER)。其基本任務不是讓一個大型模型不斷閱讀網路並自由更新人物印象,而是將長時序政治研究拆為可監控、可重試、可回放、可限權的工作流。Runtime 包含 Source Discovery、Snapshot Ingestion、Parsing/Transcription、Entity Resolution、Claim/Event Extraction、Verification、Canonical Commit、Temporal Replay、Hypothesis Evaluation、Counterevidence Search、Policy Realization Forecast、Economic Exposure Propagation、Alerting、Projection、Human Review、Evaluation 與 Correction 等 stage。
DAPIER 採取 proposal-only agent architecture。任何 LLM、搜尋 Agent、資料抽取 Agent、反證 Agent、Forecast Agent 或 Exposure Agent 都不能直接改寫 canonical evidence。Agent 只能提出 Proposal;真正的 canonical mutation 必須經 Canonical Commit Gate 驗證 provenance、schema、時間、實體解析、重複資料、來源權利與安全政策後才可 commit。Inference 與 Projection 另有獨立 write domain,避免人物假說與預測反向污染 evidence。
Runtime 將「新資料到達」與「模型重算」解耦。資料 ingest 是事件驅動且增量式;人物模型、政策機率與經濟暴露則透過 dependency-aware invalidation 觸發局部重算。只有當 canonical state 的相關 dependency 發生實質變化時,才重新執行 HYP、PRED 或 PRP 計算。這使「勤」不等於每次全量重讀,而是:
政策預測 Runtime 使用 Paper 06 的 Policy Realization Probability:
並將其更新建模為帶 evidence cutoff 的 forecast state。任何 forecast 必須保存 cutoff、feature lineage、model version、probability components、calibration cohort 與 expiry;回測時只能使用 cutoff 時已知資料。Economic Exposure Engine 則把 PRP 經 mechanism graph 傳播到 industry / firm / region / commodity 等 economic entities,但 Exposure 仍然只是 projection,不是交易指令。
本文同時定義 runtime scheduler、queue semantics、idempotency、retry、dead-letter、human-review queues、budget control、model routing、telemetry、quality SLO、alert suppression、publication tiers 與 incident response。Runtime event envelope 可選擇對映 CloudEvents 1.0 family;可觀測性採 OpenTelemetry traces / metrics / logs,並在 GenAI/Agent call 上使用明確 version-pinned semantic conventions。2026 年 OpenTelemetry 已持續擴充 GenAI observability,而 NIST 2026 年亦強調 agentic AI 生態需要 measurement probes 與 deployed-system monitoring。OECD 2026 的 Digital Government Outlook 更指出,agentic AI 將風險焦點從「模型說什麼」推向「系統做什麼」,並要求 traceable decision trails、meaningful accountability、user control,以及可暫停、逆轉與挑戰自動化行動。這些要求在 DAPIER 中被實作為 action-space separation、approval gates、audit trails 與 reversible projections。
WP02 的最終不變量是:
以及:
至此,8 篇理論論文與 2 篇技術白皮書形成完整閉環:從政治人物形成、長時序證據、論述與執行、責任歸因、勤人 AI、能力重新定價、經濟情報、證據邊界,到 canonical architecture 與 24/7 runtime。
關鍵詞: Diligent AI、agentic AI、political intelligence、economic intelligence、event-driven runtime、policy forecasting、counterevidence、OpenTelemetry、CloudEvents、human-in-the-loop、governance
Abstract
WP01 defined valid state for Temporal Actor Intelligence. WP02 defines how that state is continuously produced, maintained, recomputed, reviewed, and projected.
The Diligent AI Political/Economic Intelligence Runtime (DAPIER) is an event-driven, proposal-only agent runtime. Source discovery, ingestion, parsing, entity resolution, claim/event extraction, verification, counterevidence search, replay, hypothesis testing, policy forecasting, economic exposure propagation, alerting, and reporting are executed as bounded workflows. No LLM or autonomous agent has direct canonical-write authority. Agents produce proposals; a Canonical Commit Gate performs schema, provenance, time, entity, duplication, rights, and policy checks before canonical mutation.
Evidence ingestion and model recomputation are decoupled. New sources are incrementally canonicalized. Hypotheses, forecasts, and economic exposures are invalidated and recomputed only when relevant canonical dependencies change. Forecasts are cutoff-bound, versioned, replayable, calibrated, and expiring. Alerts are informational projections rather than autonomous decisions.
The runtime includes scheduler semantics, idempotent jobs, retry and dead-letter policies, human-review queues, model routing, cost budgets, telemetry, service-level objectives, publication tiers, incident response, and rollback of projections. Event transport may map to the CloudEvents 1.0 family. Observability uses OpenTelemetry traces, metrics, and logs with deployment-pinned semantic conventions. Governance is enforced in action spaces and approval gates, not merely through prompting.
The central invariants are:
and:
1. WP01 / WP02 Contract
WP01:
WP02:
WP02 不能修改 WP01 canonical semantics。
2. Runtime Mission
DAPIER 的 mission:
。
3. Runtime Non-Goals
不允許:
- autonomous political persuasion campaign;
- sensitive voter profiling / targeting;
- covert influence operations;
- unlawful private-data collection;
- autonomous stock/derivatives trading;
- direct personnel / credit / insurance adverse decisions;
- direct canonical writes by generative agents。
4. Runtime Trust Domains
Domain A — External Untrusted
Internet、media、public feeds、partner feeds。
Domain B — Raw Artifact
snapshotted but not yet canonicalized。
Domain C — Canonical Evidence
WP01 schema-valid evidence。
Domain D — Derived / Inference
DER、HYP、PRED、PRP、Exposure。
Domain E — Projection
public / academic / enterprise reports、alerts、dashboards。
資料只能依允許方向流動。
5. Proposal-Only Agent Architecture
所有 agent:
不能:
。
真正 commit:
。
6. Agent Types
v0.1 定義:
- Source Discovery Agent
- Fetch/Snapshot Worker
- Parser/Transcriber
- Entity Resolution Agent
- Claim/Event Extraction Agent
- Verification Agent
- Counterevidence Agent
- Replay Engine
- Hypothesis Evaluation Agent
- Policy Forecast Agent
- Economic Exposure Agent
- Projection/Report Agent
- Human Review Coordinator
- Evaluation/Calibration Worker
- Correction/Revalidation Worker
「Agent」表示 runtime role,不要求全部由 LLM 實作。
7. Capability Envelope
每個 agent 綁定:
。
8. Authority Envelope
寫入權限:
external_read
raw_write
proposal_write
derived_write
inference_write
projection_write
canonical_commit
publication_approve
其中 canonical_commit 不授予一般 Agent。
9. Least Authority Principle
。
Search Agent 不需要 publication authority。
Forecast Agent 不需要 canonical commit。
10. Workflow Object
每個工作流保存:
- workflow_id;
- workflow_type;
- trigger;
- purpose;
- cutoff;
- input IDs;
- stage state;
- attempt;
- budget;
- deadline;
- lineage;
- cancellation state。
11. Runtime State Machine
高階狀態:
CREATED
QUEUED
RUNNING
WAITING_REVIEW
RETRYABLE_FAILURE
BLOCKED
COMPLETED
CANCELLED
DEAD_LETTER
。
12. Stage State
每個 stage 另有:
PENDING
RUNNING
SUCCEEDED
FAILED_RETRYABLE
FAILED_FINAL
SKIPPED
STALE
。
13. Event-Driven Runtime
Runtime 以:
驅動 workflow。
例如:
source.discovered
source.snapshotted
evidence.proposed
evidence.committed
actor.state_invalidated
hypothesis.invalidated
forecast.invalidated
forecast.updated
exposure.updated
alert.proposed
challenge.opened
correction.committed
。
14. CloudEvents Compatibility
Runtime event envelope 可映射 CloudEvents 1.0 family:
idsourcespecversiontypetimesubjectdatacontenttypedata
DAPIER 不依賴特定 message broker。
15. Broker Neutrality
可以使用:
- Kafka-compatible;
- NATS;
- RabbitMQ;
- managed pub/sub;
- transactional outbox + queue。
Canonical semantics 不綁 broker。
16. Idempotency
所有 side-effecting stage 都需要:
。
建議:
。
17. At-Least-Once Delivery
Runtime 預設 message transport 可為 at-least-once。
因此:
。
18. Transactional Outbox
canonical commit 後產生 domain event 時:
應在同一 transaction boundary。
避免:
資料寫了,但 event 丟了。
19. Source Discovery
Discovery 只提出:
不是 SourceArtifact。
20. Discovery Connectors
可包括:
- official websites;
- legislature feeds;
- government gazettes;
- RSS/Atom;
- news APIs;
- public social feeds;
- court/regulatory repositories;
- economic data feeds。
21. Connector Contract
每個 connector:
discover(cursor, since) -> candidates[]
fetch(candidate) -> bytes + metadata
checkpoint() -> cursor
。
22. Connector Checkpoint
Checkpoint 與 canonical ledger 分離。
重跑 connector 不應改變 source semantics。
23. Snapshot First
來源一旦選擇 ingest:
之後才 parse。
禁止:
直接跳過 snapshot。
24. Parsing / Transcription
HTML、PDF、audio/video 轉為:
。
Representation 必須指向 original artifact。
25. Parser Versioning
保存:
- parser name;
- version;
- parameters;
- generated_at。
parser 升級後可重新生成 representation。
26. Entity Resolution Workflow
輸入:
輸出:
。
AMBIGUOUS 進 review 或延後 commit。
27. Extraction Proposal
Claim/Event extractor 只能產生:
EvidenceProposal
ClaimProposal
EventProposal
RoleProposal
。
28. Verification Agent
Verification 檢查:
- exact quote;
- locator;
- timestamp;
- actor identity;
- role;
- source consistency;
- parsing errors;
- duplication。
29. Canonical Commit Gate
Gate 至少驗證:
。
30. Canonical Commit Is Deterministic
同一 Proposal + 同一 gate policy:
應可重播。
若包含人工判斷,保存 reviewer record。
31. Commit Outcomes
COMMIT
REJECT
REQUEST_REVIEW
DUPLICATE
SUPERSEDE
。
32. New Evidence Does Not Rewrite Actor Truth
每次:
只代表:
canonical corpus 新增資訊。
不代表:
。
先做 dependency invalidation。
33. Dependency Graph
DER/HYP/PRED/PRP/Exposure 保存:
。
canonical object 變動:
。
34. Stale Is Not Deleted
直到 recompute。
舊版本保留。
35. Selective Recompute
只有 affected subgraph 重算:
。
36. Full Replay
全量 replay 用於:
- schema migration;
- model upgrade evaluation;
- audit;
- historical backtest;
- incident recovery。
不是每一新事件都 full replay。
37. Replay Engine
概念接口:
replay_actor(actor_id, valid_cutoff, record_cutoff, purpose, model_set)
輸出 immutable view。
38. Frozen Research Run
正式研究 run 先:
再開始分析。
研究過程中新資料進入 canonical,不影響 frozen run。
39. Hypothesis Workflow
。
40. Counterevidence Agent
不是「找反對新聞」。
它依 hypothesis 定義:
並逐類搜尋。
41. Counterevidence Completion
Paper 04:
成為 workflow completion condition。
T2/T3 publication:
才可進 gate。
42. Alternative Hypothesis Registry
同一 observed pattern 可保留:
。
Runtime 不強迫單一人物解釋。
43. Policy Forecast Workflow
對 policy :
- Freeze cutoff
- Rebuild relevant actors/roles
- Estimate Intent
- Estimate Adoption
- Estimate Implementation
- Estimate Persistence
- Compose PRP
- Calibrate
- Store Prediction
- Set expiry
- Register dependencies
44. PRP Components
Runtime 必須保存四個 component。
只存總 PRP 不足以解釋。
45. Forecast Model Routing
不同 component 可使用不同方法:
- statistical model;
- Bayesian model;
- calibrated classifier;
- LLM-assisted feature extraction;
- human expert prior;
- ensemble。
46. LLM Cannot Be Probability Oracle
LLM 直接回答:
73%。
不構成 calibrated forecast。
LLM 可:
- extract features;
- generate rationale candidates;
- identify comparable cases。
最終 probability 需 calibration framework。
47. Forecast Cutoff
每個 feature:
。
違反即:
。
48. Prediction Expiry
PRED 必須有:
。
到期:
EXPIRED
不再當 current forecast。
49. Forecast Update Trigger
trigger:
- material actor state change;
- institutional state change;
- policy stage change;
- coalition change;
- legal ruling;
- budget event;
- scheduled recalibration。
50. Materiality Threshold
不是每一篇新聞都更新 PRP。
定義:
。
低於 threshold:
。
51. Policy Momentum
Runtime 可對快速變化建立 alert candidate。
52. Economic Exposure Workflow
。
53. Exposure Is Projection
Exposure 不寫回 canonical evidence。
54. Exposure Propagation
對 graph:
使用:
- declared coefficients;
- calibrated models;
- scenario ranges。
55. Exposure Uncertainty
每條 exposure 保存:
。
56. Enterprise Alert
Alert 形式:
PRP for policy X rose from 0.31 to 0.52; primary driver: committee movement; exposure for industry Y increased.
而不是:
Buy / sell。
57. Alert Is Projection
核心:
。
Runtime 不自主下單、不自動遊說、不自動政治操作。
58. Alert Dedupe
同一 state change:
避免 event storm。
59. Alert Cooldown
每 alert key:
。
只有 material delta 超過 threshold 才 bypass。
60. Public Alert
Public projection 可提示:
- major policy stage change;
- evidence update;
- corrected claim;
- model uncertainty increase。
不得推送高名譽風險 HYP 未經 review。
61. Human Review Queues
至少:
- Entity Ambiguity Queue
- High-Risk Evidence Queue
- Responsibility/Causality Queue
- T3 Publication Queue
- Re-identification Queue
- Challenge/Correction Queue
- Model-Drift Queue
62. Human Review SLA
高風險 pending:
不是自動 timeout 放行。
63. Review Decision
APPROVE
REJECT
REQUEST_MORE_EVIDENCE
DOWNGRADE_TIER
CHANGE_EPISTEMIC_TYPE
MARK_DISPUTED
。
64. Budget Object
每 workflow:
。
65. Cost-Aware Routing
簡單 extraction:
高風險 reasoning:
。
不是所有 task 用最大模型。
66. Escalation Policy
如果:
或:
可升級:
或:
。
67. Search Budget
勤人 AI 不代表無限搜尋。
每 hypothesis 定義:
- minimum support search;
- minimum counter search;
- saturation criterion;
- hard budget。
68. Search Saturation
若新增 search batches 的:
可停止。
但要報 coverage / missingness。
69. Queue Priority
priority 可由:
決定。
70. Backpressure
來源爆量:
Runtime 必須:
- degrade noncritical crawling;
- preserve official/high-value feeds;
- increase batch sizes;
- never skip canonical validation。
71. Retry Policy
retryable:
- network timeout;
- transient provider error;
- temporary parser failure。
non-retryable:
- schema invalid;
- prohibited purpose;
- unresolved legal rights;
- deterministic policy gate failure。
72. Exponential Backoff
。
73. Dead-Letter Queue
超過 retry:
並保留:
- payload ref;
- reason;
- attempts;
- last error;
- owner。
74. Poison Source
來源反覆使 parser 崩潰:
標記:
不阻塞主 pipeline。
75. Runtime Cancellation
所有長流程必須:
。
支持:
- user cancel;
- policy cancel;
- incident cancel;
- budget cancel。
76. Reversibility
Canonical source commit 一般不刪。
但:
可:
- withdraw;
- supersede;
- correct;
- disable distribution。
77. Agentic Action Boundary
Agent 可自主:
- 搜尋;
- 讀取;
- 分析;
- 建立 proposal;
- 建立 draft report。
Agent 不可自主:
- canonical write;
- high-risk public publish;
- money transfer;
- political ad placement;
- sensitive targeting。
78. OECD Agentic AI Boundary
2026 OECD 對 agentic AI 的關鍵區分:
對:
。
因此 DAPIER 對 action space 的治理比 prompt 更重要。
79. Pause / Reverse / Challenge
任何高影響 automated chain 需要:
。
這與 OECD 2026 對政府 agentic AI 的 traceability / user control 要求一致。
80. Observability
Runtime 必須有:
。
採 OpenTelemetry-compatible telemetry。
81. Trace Hierarchy
一個 research workflow:
workflow.run
├─ source.search
├─ source.fetch
├─ parse
├─ entity.resolve
├─ extract
├─ verify
├─ counterevidence.search
├─ forecast
└─ projection.generate
。
82. GenAI Observability
LLM/Agent call telemetry 至少:
- operation;
- model/provider;
- latency;
- token/input size;
- output size;
- tool calls;
- retry;
- error type;
- cost estimate;
- prompt/content capture policy。
83. Sensitive Telemetry
完整 prompt / completion:
。
因可能含:
- licensed text;
- internal hypotheses;
- sensitive data。
預設不記完整 content 到一般 telemetry。
84. Semantic Convention Pinning
OpenTelemetry semantic conventions 會演進。
部署必須:
並由 mapping layer 轉 internal telemetry schema。
85. Runtime Metrics
核心:
- connector lag;
- ingestion throughput;
- source error rate;
- canonical commit latency;
- entity ambiguity rate;
- stale inference count;
- forecast recompute latency;
- alert latency;
- review queue depth;
- token/model cost;
- evidence trace failure;
- counterevidence completion。
86. Quality SLO
示意:
。
實際 deployment SLO 需依系統規模設定。
87. Freshness SLO
不同 source class:
- official policy feeds:minutes-hours;
- major news:hours;
- academic publications:days;
- archival backfill:best effort。
不需要所有來源 real-time。
88. Diligence SLO
Paper 04:
作 quality dashboard。
89. Forecast SLO
每 forecast family:
- calibration;
- Brier score;
- Log Loss;
- coverage;
- lead time;
- abstention rate。
90. Model Drift
監控:
。
drift 觸發:
。
91. Deployed AI Monitoring
NIST 2026 特別指出 deployed AI 需要 functionality、operational 等持續 monitoring。
DAPIER 因此把 post-deployment monitoring 視為 runtime core,而不是 release 後附加工作。
92. Agent Measurement Probes
每 Agent 都要可追:
在高風險 workflow 可保留更細 decision trail。
93. NIST Agent Standards Direction
2026 NIST 已啟動 AI Agent Standards Initiative,重點包含 interoperable / secure agent ecosystem。
DAPIER 因此不假設 proprietary single-agent runtime 是長期唯一介面。
94. Model Provider Abstraction
模型接口:
generate()
embed()
classify()
extract()
rerank()
。
Runtime 不把 canonical semantics 綁定特定模型 provider。
95. Tool Provider Abstraction
搜尋、資料庫、轉錄、圖運算都透過 capability interface。
96. Deterministic Tool First
可 deterministic 的:
- hash;
- schema validation;
- time filtering;
- graph traversal;
- numeric calculation;
不交給 LLM 自由推理。
97. LLM Use Cases
適合:
- semantic extraction;
- ambiguous classification;
- query decomposition;
- hypothesis generation;
- evidence synthesis;
- natural-language projection。
98. LLM Abstention
Agent 可以輸出:
INSUFFICIENT_EVIDENCE
AMBIGUOUS_ENTITY
NEEDS_REVIEW
不是強迫生成答案。
99. Multi-Agent Is Optional
DAPIER 支援多 Agent,但:
。
簡單 task 可單 worker。
100. Agent Coordination
多 Agent 共享:
而不是自由聊天當 source of truth。
101. Shared Scratchpad Is Non-Canonical
Agent scratchpad:
。
不能被下次 run 當 evidence。
102. Runtime Memory
分:
Canonical Memory
WP01 ledger。
Operational Memory
workflow state、checkpoints。
Ephemeral Reasoning Context
當次 Agent context。
三者不可混用。
103. Research Session
每次深度分析產生:
保存:
- query;
- scope;
- cutoff;
- evidence snapshot ID;
- workflows;
- outputs;
- review state。
104. Repeatability
同一:
應能產生可比較結果。
LLM 非 deterministic 時保存:
- model ID;
- parameters;
- seed if available;
- output lineage。
105. Query Planner
Planner 將問題拆成:
但不能改變 purpose。
106. Query Plan Example
「此人物是否在成功時較常攬功?」
拆:
- identify role/time scope;
- collect positive matched events;
- collect negative matched events;
- rebuild responsibility edges;
- code attribution;
- counterevidence;
- calculate asymmetry;
- report uncertainty。
107. Research Planner Risk
Planner 可能因 user wording 形成 confirmation bias。
Mitigation:
同時自動產生:
。
108. Null Hypothesis
例如:
。
109. Evidence Search Strategy
順序優先:
。
110. Media Search Is Discovery, Not Truth
新聞搜尋結果先變:
。
111. Long-Horizon Monitoring
Condition watch:
才更新 actor model。
不是固定每天重寫人物評價。
112. Material Change
包括:
- new role;
- major decision;
- policy stage move;
- significant correction;
- adjudicated finding;
- major election;
- material coalition change。
113. Scheduled Maintenance
每日:
- connector health;
- critical feeds。
每週:
- unresolved entity review;
- stale inferences。
每月:
- forecast calibration;
- coverage audit;
- model drift。
頻率可配置。
114. Archive Backfill
Historical corpus 可用低優先 queue 漸進補齊。
115. Backfill Must Preserve Record Time
若 2026 才抓到 2018 文件:
。
歷史 replay 不偷用。
116. Correction Workflow
來源更正:
。
117. Challenge Workflow
外部 challenge:
。
118. Incident Classes
- canonical corruption;
- temporal leakage;
- source poisoning;
- widespread entity merge error;
- privacy breach;
- prohibited-use invocation;
- publication of unsupported high-risk allegation;
- forecast pipeline miscalibration。
119. Incident Severity
SEV0
SEV1
SEV2
SEV3
SEV0 可觸發:
。
120. Kill Switch
Runtime 必須可:
- stop connector;
- stop agent class;
- freeze publication;
- freeze forecast;
- revoke capability;
- invalidate projection。
121. Canonical Recovery
Canonical ledger recovery:
- restore snapshot;
- replay append log;
- verify hashes;
- rebuild indexes;
- rebuild graph;
- mark inferences stale;
- recompute。
122. Non-Canonical Recovery
Vector/graph/cache:
。
123. Cost Incident
若 token / API cost 超 budget:
:
- lower search depth;
- delay archival backfill;
- lower-cost model;
- preserve critical official feeds;
- never disable evidence gates。
124. Publication Pipeline
。
125. Public Report
至少包含:
- scope;
- cutoff;
- observed;
- derived;
- hypotheses;
- counterevidence;
- unknowns;
- sources;
- last updated。
126. Enterprise Report
額外:
- PRP;
- component probabilities;
- policy momentum;
- exposure mechanism;
- scenario;
- uncertainty;
- lead time。
127. Academic Export
提供:
- corpus snapshot ID;
- schema version;
- de-identified records where required;
- derivation methods;
- model versions;
- codebook;
- reproducibility manifest。
128. Publication Revocation
若重大錯誤:
並提供 correction notice。
129. No Silent Rewrite
公開報告更新:
保留 revision history。
130. Privacy Runtime Gate
在 ingestion / inference / publication 都檢查:
。
131. Coded Case Runtime
coded case projection 先做:
。
132. Sensitive Trait Gate
若 inference target 屬敏感私人屬性:
除非明確 research/legal policy 另行允許。
133. Political Targeting Boundary
Runtime 不提供:
build_voter_profile()
target_sensitive_group()
optimize_political_persuasion()
。
134. Financial Boundary
Runtime 可:
不能:
。
135. Security
所有 connector input 視為:
。
防:
- prompt injection;
- malicious HTML;
- poisoned documents;
- SSRF;
- decompression bombs;
- malicious attachments。
136. Tool Isolation
Parser / fetcher / model tool 各自最小 sandbox。
137. Prompt Injection Boundary
來源文件中的:
Ignore instructions...
只能當 source content。
不能改 runtime policy。
138. Retrieval Content Is Data
。
139. Secret Isolation
API keys:
- secret manager;
- never canonical;
- never prompt unless required;
- never telemetry content。
140. Network Policy
不同 worker 可有不同 outbound allowlist。
141. Data Retention
Raw artifacts、licensed data、telemetry 分別有 retention policy。
Canonical public evidence 可長期保存;敏感 operational logs 期限更短。
142. Auditability
每次高風險 output 要知道:
在:
用:
與:
產生。
143. Runtime Invariants
隨附 lai_runtime_invariants_v0.1.json,包含至少 24 條 machine-readable invariants。
144. Runtime Config Schema
隨附:
lai_runtime_config_schema_v0.1.json
定義:
- environment;
- queues;
- budgets;
- model routes;
- review policy;
- publication policy;
- telemetry;
- forecast;
- alerting;
- security。
145. Runtime Event Catalog
隨附:
lai_runtime_event_catalog_v0.1.json
定義 domain event types 與 CloudEvents-compatible mapping。
146. Runtime State Machine
隨附:
lai_runtime_state_machine_v0.1.json
定義 workflow 與 stage transition。
147. Example Config
隨附:
lai_example_runtime_config_v0.1.json
為 non-production reference。
148. Acceptance Test Suite
隨附:
lai_wp02_acceptance_tests_v0.1.json
涵蓋:
- no direct canonical writes;
- temporal cutoff;
- idempotency;
- counterevidence gate;
- publication gate;
- forecast expiry;
- stale dependency recompute;
- alert/action separation;
- telemetry;
- kill switch;
- correction lineage。
149. MVP Runtime
v0.1 MVP 建議:
。
150. MVP Services
最小服務:
- Connector Service
- Artifact Store
- Canonical Commit Service
- Canonical DB
- Index Service
- Replay Service
- Research/Inference Worker
- Forecast Worker
- Projection Service
- Review Console
- Telemetry Collector
151. MVP Deployment
第一版可:
。
不需要微服務極端拆分。
邏輯 boundary 比 deployment boundary 重要。
152. Reference Deployment Philosophy
先:
比過早微服務化更合理。
153. Scaling Axis
未來按:
- source volume;
- actor count;
- query concurrency;
- forecast families;
- enterprise tenants;
逐步拆。
154. Tenant Isolation
企業版需要:
與 canonical public evidence 分離。
客戶自有資料不得回流公共 canonical。
155. Customer Data Boundary
除非明確授權與 policy。
156. Multi-Tenant Inference
同一 public canonical 可供多 tenant 使用。
私人 scenario / annotations 隔離。
157. Evaluation Harness
每個 model route 有 benchmark:
- extraction;
- entity resolution;
- claim classification;
- contradiction;
- forecast;
- report factuality。
158. Shadow Evaluation
新模型先:
不影響 production projection。
159. Canary
低風險少量 workflow:
後才擴大。
160. Rollback
模型/config 回滾只影響 derived/inference/projection。
canonical evidence 不需回滾,除非 canonical incident。
161. Continuous Evaluation
與 NIST 2026 deployed-AI monitoring方向一致。
162. OpenTelemetry Resource
建議 resource attributes:
- service.name;
- service.version;
- deployment.environment;
- lai.workflow.type;
- lai.purpose;
- lai.risk_tier。
自訂 lai.* namespace 不假裝是 OTel 標準。
163. Trace Privacy
lai.actor.id 可用 internal pseudonymous ID。
不把敏感姓名當高基數公共 telemetry attribute。
164. Event Correlation
Runtime event:
連接 workflow trace。
165. SLO/Error Budget
系統優先:
高於:
。
當 error budget 耗盡:
而不是降低 gate。
166. Runtime Governance Board
正式部署可有:
- research owner;
- data steward;
- security owner;
- model owner;
- publication reviewer;
- enterprise risk owner。
不是單一 AI 自己治理自己。
167. Change Management
修改:
- schema;
- purpose;
- risk threshold;
- model route;
- publication tier;
需要 versioned change record。
168. Runtime Policy as Code
建議:
。
例如:
deny if purpose == sensitive_voter_microtargeting
deny T3 publish unless counterevidence_complete
deny canonical_write unless principal == commit_service
。
169. Policy Engine Is Deterministic
高階安全 gate 優先 deterministic policy。
不能交給 LLM:
你覺得可以發布嗎?
作唯一決策。
170. Fail Closed
對:
- canonical commit;
- T3 publication;
- sensitive inference;
- prohibited use;
遇到 policy service failure:
。
171. Fail Open
只有低風險非關鍵 telemetry 等可視需求 fail open。
172. Runtime Data Contracts
任何 stage output 先 schema validate。
。
173. Model Output Is Untrusted
即使是內部模型:
。
174. Structured Output
對 commit proposal,prefer:
不是 free-form prose。
175. Evidence Citation
任何 report-level material claim:
由 Projection Service render 人類 citation。
176. No Citation Fabrication
如果 evidence path 空:
且 material claim 應降級/刪除/abstain。
177. Runtime Search Audit
保存:
- query;
- provider;
- time;
- result IDs;
- selected results;
高風險研究可重播 discovery path。
178. Search Provider Diversity
重要研究可用多來源/多 provider 降低單一搜尋排序偏誤。
179. Search Result Ranking Is Not Evidence Weight
。
180. Runtime Research Closure
ResearchSession 完成條件:
- scope satisfied;
- coverage reported;
- counterevidence done;
- unresolved listed;
- evidence contracts valid;
- review done if required。
181. Partial Completion
若 budget 用盡:
並報:
- completed scope;
- missing scope;
- unresolved questions。
不能假稱完整。
182. Academic Reproducibility
研究 export 包含:
。
183. Enterprise Reproducibility
企業 alert 保存:
。
184. Public Reproducibility
公眾報告至少能回:
。
185. Runtime Security Review
每 connector / tool 新增前:
- threat model;
- permissions;
- outbound network;
- data class;
- secret needs;
- incident plan。
186. Agent Tool Allowlist
Agent 只能使用明確 allowlisted tools。
187. No Arbitrary Code Execution by Default
政治資料分析 Runtime 不需要讓一般研究 Agent arbitrary shell。
高權限 computation 隔離。
188. Sandbox
需要解析不可信檔案時:
。
189. Data Poisoning
來源如果刻意大量發布同一虛假敘事:
- source independence;
- provenance;
- official verification;
- counterevidence;
降低 poisoning 影響。
190. Sybil Source Detection
大量低品質站點不等於多獨立證據。
191. Model Self-Contamination
過去 AI 生成的報告如果被網路收錄:
不應被重新 ingest 成獨立 source evidence,除非標明它只是 secondary artifact。
192. Source Circularity Detection
需要 lineage 防止循環自證。
193. Runtime Ethics Invariant
。
194. Runtime Economic Invariant
。
195. Runtime Political Invariant
。
196. Runtime AI Invariant
。
197. Series Closure
8 Papers:
WP01:
WP02:
。
因此整體:
。
198. Final Runtime Equation
可將 DAPIER 壓縮為:
其中:
- :queues/events;
- :workflows;
- :bounded agents;
- :canonical commit gates;
- :replay / research;
- :forecast;
- :economic exposure propagation;
- :projection firewall;
- :observability / oversight。
199. Final Invariants
以及:
。
200. 結論
「勤人 AI」真正困難的地方並不是讓 AI 讀很多資料。
如果沒有 runtime discipline,所謂「長時序人物分析」很容易退化成:
。
DAPIER 拒絕這種架構。
它把「勤」拆成:
。
同時把「AI 自主」限制在:
而不是:
。
因此,政治人物的新資料不會直接改寫人物結論;Forecast 不會因新聞標題就自行飄動;企業 alert 不會自動變成交易;研究 Agent 不會直接修改 canonical evidence;高風險發布不會因模型很有自信就跳過人工審核。
這套 Runtime 最終追求的不是:
AI 永遠比人更會判斷政治。
而是:
讓原本需要大量研究員才能長期維護的政治記憶、反證、時序回放、政策機率與企業暴露分析,變成可持續運作,而且每一個重要結果都能被中止、回放、檢查與修正。
因此 WP02 的最終工程命題是:
而與 Paper 07 的:
以及 WP01 的:
共同形成完整三角:
至此,AI-Native Longitudinal Actor Intelligence v0.1 的理論與技術系列完成。
References and Current Technical Context
CloudEvents Authors / CNCF. (2026). CloudEvents Specification, 1.0 family. https://github.com/cloudevents/spec
NIST. (2024; page updated 2026). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1. https://doi.org/10.6028/NIST.AI.600-1
NIST. (2026). Challenges to the Monitoring of Deployed AI Systems, NIST AI 800-4. National Institute of Standards and Technology.
NIST Center for AI Standards and Innovation. (2026). AI Agent Standards Initiative. National Institute of Standards and Technology.
OECD. (2026). The Agentic AI Landscape and Its Conceptual Foundations. OECD Artificial Intelligence Papers, No. 56. https://doi.org/10.1787/396cf758-en
OECD. (2026). Digital Government Outlook 2026: From Foundations to Transformational Impact. OECD Publishing. https://doi.org/10.1787/0496b2bc-en
OpenTelemetry. (2026). OpenTelemetry Semantic Conventions. https://opentelemetry.io/docs/specs/semconv/
OpenTelemetry. (2026). Inside the LLM Call: GenAI Observability with OpenTelemetry. https://opentelemetry.io/blog/2026/genai-observability/
Canonical Source Note
本檔為 WP02 v0.1 的 UTF-8 canonical Markdown source。正式數學 source 僅使用 $...$ 與 $$...$$ delimiter。任何 dashboard、alert、trace UI、聊天畫面、HTML、PDF、簡報、runtime state cache、vector index、graph projection 或模型輸出都不是 canonical source。Runtime policy、schema、invariant 與 workflow 的正式修改必須先修改 canonical artifact,執行 validation,再 commit 新版本。