類全域 AI 世界—計算—觀察統合系列(Paper 10)
驗證層:多世界、多投影、多模型為何仍不等於真理
The Verification Layer: Why Many Worlds, Projections, Models, and Verifiers Still Do Not Equal Truth
作者: Neo.K
AI 協作: Aletheia(GPT-5.6 Sol)
機構: EveMissLab/一言諾科技有限公司
系列: 類全域 AI 世界—計算—觀察統合系列
英文系列名: Global-Like AI World–Computation–Observation Synthesis Series
篇次: Paper 10 / 12
版本: v0.1
日期: 2026-09-09
研究定位: WDC Cross-World Evidence × SWFR × DEST × MWT Global Legality × WCO Domain Layer × WCO Execution Layer × Model/Simulation Credibility × Dependence-Aware Ensemble Verification × Counterexample Search
前篇: Paper 09《執行層:方法選擇、MSSP、RDR、WFS 與 Runtime Routing》
狀態: WCO 驗證層母規格/evidence-verification architecture;不宣稱存在跨所有學科通用的 evidence aggregation 公式,也不宣稱任何有限驗證流程可保證世界層絕對真理
摘要
Paper 01–09 已經使類全域 AI 可以:
- 維持受治理世界族;
- 在世界族中執行異質計算;
- 選擇觀察算子;
- 維持 epistemic qualification;
- 自適應選擇 projection;
- 將 projection 落到 physical carrier;
- 重建長期記憶;
- 將 intent 編譯成 durable execution plan。
但系統越強,越容易出現一種新的危險:
很多東西都同意,所以看起來一定是真的。
這種「一致性幻覺」可能來自:
- 100 個 worlds 共享同一 simulator;
- 20 個 foundation models 共享高度重疊的訓練資料;
- 10 個 observers 實際讀取同一 sensor;
- 5 種 projections 只是同一 observation 的不同 carrier view;
- 3 個 verifiers 共用同一 specification bug;
- 多個 agents 在討論後彼此污染上下文;
- 多個 runs 只是改 random seed;
- simulation ensemble 從未接受 real-world calibration。
因此本文建立 WCO 的正式 Verification Layer。
本文延續 WDC-05 的第一原則:
World Count=Independent Evidence Count=Truth.
並將其擴張為:
World Count=Model Count=Observer Count=Projection Count=Verifier Count=Independent Evidence Count.
更進一步:
Agreement=Independence=Validity=Transport=Truth.
本文提出 WCO Evidence–Verification Fabric(EVF):
VtWCO=⟨Etpkt,Gtdep,Gtprov,Gtcert,Ctclaim,Xtcounter,Tttransport,Utunknown,Dtverify,Htverify⟩.
其中:
- Etpkt:typed evidence packets;
- Gtdep:evidence-dependence graph / hypergraph;
- Gtprov:provenance graph;
- Gtcert:certificate graph;
- Ctclaim:claim-type and acceptance contracts;
- Xtcounter:counterexample / falsification state;
- Tttransport:world-to-target / model-to-reality transport contracts;
- Utunknown:unknown-world / omitted-mechanism state;
- Dtverify:verification and transport debts;
- Htverify:verification history / reopening state。
本文定義一般化 Evidence Packet:
ei(q)=⟨EvidenceId,ClaimId,SourceClass,WorldId,RunId,Model,Data,Assumptions,Observer,ObservationOperator,Projection,Evaluator,Verifier,Provider,Communication,Outcome,Validity,Uncertainty,Transport,Certificates⟩.
證據聚合的輸入因此不是:
{y1,…,yn},
而是:
{e1,…,en}.
本文將 evidence dependence 拆成多維向量:
dij=(dL,dB,dM,dD,dA,dO,dP,dE,dV,dR,dC).
其中可分別表示:
- lineage dependence;
- backend dependence;
- model / architecture dependence;
- data dependence;
- assumption dependence;
- observer / sensing dependence;
- projection dependence;
- evaluator dependence;
- verifier / checker dependence;
- provider / infrastructure dependence;
- communication / context dependence。
這使 common-mode error 不再被簡化成「outputs correlation」。
本文定義 dependence graph:
Gtdep=(VE,ED,ωD).
也允許使用 hyperedge 表示多個 evidence packets 共同依賴同一 hidden source:
h:{e1,…,ek}→zshared.
因此:
Different Processes=Different Error Sources.
本文保留概念性 Effective Evidence Count:
Neff=Neff(Gdep,ClaimType,AggregationContract),
但不提出 universal formula。任何報告若宣稱:
500 worlds / models / verifiers agree
至少應同時報告:
- total runs;
- evidence families;
- estimated dependence;
- unresolved dependence;
- counterexamples;
- target transport debt。
本文進一步建立五級 verification interface:
Eexec→Eint→Ecross→Etrans→Eext.
但箭頭不是邏輯蘊含:
Eexec⇒Eint⇒Ecross⇒Etrans⇒Eext.
Eexec — Execution Integrity
world / model / verifier 是否依自己的 specification 正確執行?
Eint — Internal Validation
結果是否在同一 contract 中穩定、可重現、通過 known cases?
Ecross — Cross-Family Robustness
不同 backend、model、data、assumption、observer、evaluator、verifier family 下是否仍成立?
Etrans — Target Transport
目前 evidence 能否合法搬到 target domain / reality?
Eext — External Resolution
prospective real-world data、independent experiment、deployment observation 或 trusted external resolution 是否支持?
本文再次強調:
Model Executes Correctly=Model Represents Target Adequately=Result Supports Target Decision.
這與模型/模擬 credibility 工程的 verification、validation、uncertainty 與 acceptance criteria 分離相容。
本文再建立 Claim-Type-Aware Verification。不同 claim 類型不能使用同一 aggregation rule:
ClaimType(q)∈{Universal,Existential,Probabilistic,Causal,Comparative,Forecast,Optimization,Safety}.
例如:
- universal claim 可能被一個有效反例擊穿;
- existential claim 可以被一個有效 witness 支持;
- probabilistic claim 不能被單一 rare counterexample 自動否定;
- causal claim 需要干預/識別條件;
- forecast claim 需要 prospective resolution;
- safety claim 可能對 tail / worst-case evidence 特別敏感。
因此:
Evidence Aggregation=f(EvidencePackets,Dependence,ClaimType,Counterexamples,Transport,AcceptanceContract).
而不是:
Vote(y1,…,yn).
本文保留 WDC 的 Counterexample Preservation Principle,並將其擴張為 WCO 的 Counterexample Pressure:
X¬qadv={W,Model,Observer,Projection,Verifier,Dataset,Assumption}¬q.
驗證系統不應只生成更多支持性 variants,而應主動尋找:
- weakest assumptions;
- high-risk regions;
- alternative causal mechanisms;
- rare valid branches;
- adversarial projections;
- independent data;
- different verifier implementation;
- specification mismatch;
- transport-breaking regimes。
因此:
Consensus+Independent Counterevidence Search>Consensus Alone
只作方法論偏序,不作 universal theorem。
本文亦將 Paper 06 的 projection 納入 verification。若多個 projections:
Π1,…,Πk
都展示相同 pattern,它們不自動形成 k 份 evidence。它們可能只是:
one observation→many views.
但 adversarial reprojection 可以作為 verification probe:如果 claim 只在某個 view 中看似成立,換 projection 後立即暴露 aliasing、folding、scale loss 或 uncertainty suppression,則原結論存在 projection sensitivity。
本文定義:
Dproj-sens(q)=VarΠ∈PvalidJ(q∣Π).
這不是 truth metric,而是 representation robustness diagnostic。
本文也把 verifier 本身納入驗證對象。
對 formal / software / scientific verification:
SpecificationCorrectness=ImplementationCorrectness=CheckerCorrectness.
兩個 checker 若共享:
- 同一 formal statement;
- 同一 parser;
- 同一 compiler;
- 同一 test oracle;
- 同一 hidden assumption;
其 agreement 不能被當成 fully independent verification。
因此本文定義 Verifier Diversity Profile:
VD=(Vspec,Vimpl,Vlogic,Vtoolchain,Vdata,Vinstitution,Vexecution).
本文再建立 Certificate Graph:
GC=(VC,EC).
certificate node 可以包括:
- execution integrity certificate;
- formal proof;
- test certificate;
- calibration certificate;
- bridge certificate;
- projection certificate;
- transport certificate;
- physical realization certificate;
- external replication record。
但是:
Local Certificate Validity⇒Global Certificate Validity.
因為長鏈可能仍有:
- incompatible scopes;
- version mismatch;
- circular support;
- conflicting assumptions;
- semantic drift;
- transport gap;
- history inconsistency;
- noncomposable bridges。
因此 certificate composition 必須經:
ComposeCert(C1,…,Cn,Γ,H)→{Pass,Fail,Undetermined,Conflicted}.
本文把 Paper 05 的 qualification 與 verification 接合:
Dverify=Dglobal=WorldLevelTruth.
一個命題即使在某 formal system 中 verified:
T⊢q,
仍只表示相對 T 的 formal result。若 q 同時聲稱物理世界狀態,仍需 empirical / transport obligations。
本文也正式保留 Unknown-World Mass:
UW
代表:
當前 world/model/observer/verifier ensemble 可能沒有包含真正關鍵的 mechanism / ontology / failure family。
即使:
Neff
很大,也不能自動得到:
UW=0.
本文定義 Ensemble Closure Error:
Eclosure=treating current evidence-generating family as exhaustive.
降低方法包括:
- new backend;
- new ontology;
- independent source;
- adversarial world;
- alternative observer;
- new measurement channel;
- different verifier;
- real anomaly;
- prospective test。
不是:
just more seeds.
本文進一步建立 Verification Debt Vector:
DV(q)=(Dexec,Dint,Ddep,Dcounter,Dspec,Dcert,Dglobal,Dtransport,Dexternal,Dunknown).
其中:
- Dexec:執行完整性債務;
- Dint:internal validation debt;
- Ddep:independence / dependence debt;
- Dcounter:反例搜尋債務;
- Dspec:specification debt;
- Dcert:certificate debt;
- Dglobal:local-to-global composition debt;
- Dtransport:target transport debt;
- Dexternal:external resolution debt;
- Dunknown:ensemble closure / unknown-world debt。
因此 Verification Layer 的輸出不應只是:
PASS.
而應是 Verification Capsule(VCap):
VCap(q)=⟨Claim,ClaimType,QCap,EvidencePackets,DependenceProfile,EvidenceFamilies,Counterexamples,VerifierProfile,CertificateGraph,VerificationLayer,TransportScope,Debt,Unknowns,DecisionStatus,ReopenTriggers⟩.
本文最終提出 Verification Closure:
VClose(q,τ,ρ)
表示:
- task-required verification layer 已到達;
- hard blockers 已處理;
- counterexample search 達到預設 budget / policy;
- dependence 已充分揭露;
- required certificates 可組合;
- transport debt 在 task threshold 內;
- unknowns 被顯式保留;
- reopen triggers 已記錄。
但:
VClose=Final Truth.
Verification closure 是 task-relative、risk-relative、time-relative 的 operational closure。
新 evidence、model failure、real-world contradiction、certificate revocation、new ontology 或 hidden common-mode source 都可以:
Reopen(q).
因此本文最終命題是:
Verification is not the accumulation of agreeing outputs.
而是:
the structured reduction of alternative error explanations through validity checks, dependence analysis, counterexample pressure, certificate composition, target transport, and external resolution.
關鍵詞: Verification Fabric、Evidence Independence、Common-Mode Failure、Cross-World Evidence、Counterexample、Verifier Diversity、Certificate Graph、Transport Debt、Simulation Credibility、Unknown World、Global AI
0. Paper 09 留下的 Verification 問題
Paper 09 已建立:
Execute→Observe→Verify/Reconcile→Commit/Stop/Replan.
但:
Verify 到底是什麼?
仍需獨立一篇回答。
1. Agreement 不是 Verification
Agreement=Verification.
2. Repetition 也不是 Independence
RepeatedAgreement=IndependentReplication.
3. World Count 不等 Evidence Count
NW=NEind.
4. Model Count 不等 Independent Model Count
NM=NMind.
5. Observer Count 不等 Independent Observation Count
NO=NOind.
6. Projection Count 不等 Evidence Count
NΠ=NEind.
7. Verifier Count 不等 Independent Verification Count
NV=NVind.
8. Evidence Packet
ei(q)=⟨EvidenceId,ClaimId,SourceClass,WorldId,RunId,Model,Data,Assumptions,Observer,ObservationOperator,Projection,Evaluator,Verifier,Provider,Communication,Outcome,Validity,Uncertainty,Transport,Certificates⟩.
9. Outcome 只是 Packet 的一個欄位
Yi
不能代表整份 evidence。
10. Evidence Source Class
可分:
Source∈{Real,External,World,Synthetic,Derived,Unknown}.
11. Source Class 不等 Truth Rank
real observation 也可能錯。
simulation 也可能非常有用。
但 source 不可以消失。
12. Provenance 是 Verification Input
EvidenceWithoutProvenance=HighVerificationDebt.
13. Evidence Dependence Graph
Gdep=(VE,ED,ωD).
14. Pairwise Dependence 只是簡化
真正 common-mode source 可能是 hyperedge。
15. Dependence Dimensions
dij=(dL,dB,dM,dD,dA,dO,dP,dE,dV,dR,dC).
16. Lineage Dependence
兩個 worlds 來自同 root。
17. Backend Dependence
使用同 simulator / runtime。
18. Model Dependence
不同 model name 可能共享 architecture、weights ancestry 或 training regime。
19. Data Dependence
不同 systems 可能引用相同 dataset / source chain。
20. Assumption Dependence
不同模型可能都假設同一錯誤機制。
21. Observer Dependence
不同 agents 可能讀同 sensor 或同 upstream observation。
22. Projection Dependence
不同 view 可能都是同 observation 的 transform。
23. Evaluator Dependence
所有 worlds 由同一 judge 評分。
24. Verifier Dependence
多 checker 可能共享同 specification / parser / oracle。
25. Provider Dependence
不同 service names 可能實際跑同 infrastructure / model。
26. Communication Dependence
多 agents 先互相討論後再「獨立」回答。
27. Output Correlation 不等 Dependence
OutputCorrelation=CausalDependence.
28. Output Difference 也不證明 Independence
DifferentOutcome=IndependentEvidence.
29. Evidence Family
共享主要 root / backend / data / assumptions / evaluator 的 worlds:
Fk.
30. Family 內 Replication 有價值
它可提升:
- stochastic precision;
- numerical stability;
- tail estimation;
- within-contract robustness。
31. 但 Family 內 Run 不等跨 Family Replication
WithinFamilyReplication=CrossFamilyReplication.
32. Replication Ladder
本文沿用:
R0<R1<R2<R3<R4.
33. R0 — Re-run
同 spec / backend / evaluator。
34. R1 — Branch
不同 seed / controlled intervention。
35. R2 — Backend
不同 simulator / model family。
36. R3 — Assumption / Evaluator
不同 assumptions / data path / evaluator family。
37. R4 — External Target
真實 experiment / prospective resolution。
38. 這不是固定 Numerical Weight
只表示 diversification depth。
39. Effective Evidence Count
Neff=Neff(Gdep,ClaimType,AggregationContract).
40. 不提供 Universal Formula
不同 claim / domain 的 dependence semantics 不同。
41. 報告必須分開
至少:
Nrun,Nworld,Nfamily,Neffestimated.
42. 無法估計也是合法結果
Neff=Unknown
優於假精確。
43. Five Verification Layers
Eexec,Eint,Ecross,Etrans,Eext.
44. Execution Integrity
問:
系統是否照 specification 跑?
45. Internal Validation
問:
同 contract 下是否穩定 / known-case correct?
46. Cross-Family Robustness
問:
換 backend / assumptions / evaluator / observer 是否仍成立?
47. Target Transport
問:
evidence 能不能搬到 target domain?
48. External Resolution
問:
real data / independent experiment 支不支持?
49. 五層不能自動跳級
Eexec⇒Eint⇒Ecross⇒Etrans⇒Eext.
50. Verification 與 Validation 分離
一個 implementation 正確:
⇒
model adequacy。
51. Model Adequacy 也不直接等於 Decision Adequacy
CorrectExecution=AdequateModel=AdequateDecisionEvidence.
52. Claim Type First
ClaimType(q)
必須在 aggregate 前明確。
53. Universal Claim
一個 valid in-scope counterexample 可能足以推翻 universal claim。
54. Existential Claim
一個 valid witness 可能足以支持 existence。
55. Probabilistic Claim
需要 distributional evidence。
single rare event 不可直接當 universal refutation。
56. Causal Claim
需要 causal identification / intervention assumptions。
57. Comparative Claim
需要 matched comparison contract。
58. Forecast Claim
需要 future prospective resolution。
59. Safety Claim
可能需 tail / worst-case / failure evidence。
60. No Universal Voting Rule
OneWorldOneVote
不是預設。
61. Pre-Registered Claim Equivalence
若要說不同 world outputs 支持同一命題,先定義:
ϕ(y),dϕ,ϵq.
62. Consensus Stretching
看完結果後才放寬 equivalence:
ConsensusStretching=PostHocEvidenceInflation.
63. Pre-Register / Version Changes
若 equivalence contract 改變,要留下版本與原因。
64. Counterexample Set
CE(q)={ei:Yi materially conflicts with q}.
65. Counterexample 不能被 Majority 直接刪除
100:1
不自動代表 1 無效。
66. Counterexample Validation Questions
至少問:
- bug?
- contract mismatch?
- rare but valid?
- different ontology?
- transport irrelevant?
- shared blind spot?
67. Counterexample Escalation
若:
Validity↑,Independence↑,Transport↑,
就提高 diagnostic / replication budget。
68. Counterexample Pressure
X¬qadv={World,Model,Observer,Projection,Verifier,Dataset,Assumption}¬q.
69. 不只 spawn supporting worlds
驗證器應主動嘗試破壞結論。
70. Adversarial World Search
找:
- weak assumptions;
- failure region;
- alternative dynamics;
- tail cases。
71. Adversarial Model Search
換:
- architecture;
- training family;
- formalism;
- priors。
72. Adversarial Observer Search
換 observation channel / scale / sensor。
73. Adversarial Projection Search
用不同 valid projection 看是否暴露 hidden distinction。
74. Adversarial Verifier Search
換 checker implementation / logic / toolchain。
75. Consensus + Falsification Pressure
Consensus+IndependentCounterevidenceSearch>ConsensusAlone.
76. Projection Agreement 不是 Evidence Independence
Π1(Y)=support∧Π2(Y)=support⇒2 EvidenceUnits.
77. One Observation → Many Views
Y→Π1(Y),…,Πk(Y).
通常仍是一份 source evidence。
78. Projection Sensitivity
Dproj-sens(q)=VarΠ∈PvalidJ(q∣Π).
79. Projection Sensitivity 不是 Truth Metric
它是 representation robustness diagnostic。
80. High Sensitivity
可能表示:
- aliasing;
- folding loss;
- scale sensitivity;
- uncertainty suppression;
- layout-induced bias。
81. Low Sensitivity 也不代表 True
shared source 仍可能錯。
82. Observer Agreement 不是 Truth
ObserverAgreement=Truth.
83. Multi-Agent Consensus 可能有 Communication Contamination
若 agents 先分享 reasoning / outputs:
dC↑.
84. Independent First Pass
某些 verification protocol 可先:
IndependentPass→Compare.
85. 之後再 Deliberation
Deliberation 可改善 solution,但不能回溯宣稱 initial evidence independent。
86. Model Agreement 不是 Independent Replication
ModelAgreement=IndependentReplication.
87. Different Model Names 也可能 Shared Ancestry
需要 provenance。
88. Verifier 也是被驗證對象
Verifier=Oracle.
89. Specification Correctness
spec 是否表示真正 intended claim?
90. Implementation Correctness
solver / model 是否照 spec?
91. Checker Correctness
checker 是否正確判定 output?
92. 三者分離
SpecificationCorrectness=ImplementationCorrectness=CheckerCorrectness.
93. Checker Agreement 也可能 Common-Mode
例如共享同 parser bug。
94. Verifier Diversity Profile
VD=(Vspec,Vimpl,Vlogic,Vtoolchain,Vdata,Vinstitution,Vexecution).
95. Independent Verification 不一定要求不同 Institution
institutional diversity 只是 dependence axis之一。
96. Formal Proof Verification
kernel check 可以強化:
Eexec/Eint.
97. 但 Formal Spec 仍可能 Wrong Target
ProofValid=SpecificationFaithful.
98. Software Tests
tests passing:
⇒
uncovered states不存在。
99. Test Oracle 也是依賴源
如果所有 tests 用同 oracle,就有 common-mode risk。
100. Certificate Graph
GC=(VC,EC).
101. Certificate Node Types
可包含:
- execution;
- proof;
- test;
- calibration;
- bridge;
- projection;
- physical realization;
- transport;
- external replication。
102. Certificate Dependency
certificate 也會依賴:
- version;
- source;
- spec;
- environment;
- assumptions。
103. Certificate Staleness
dependency 更新:
Ci→STALE.
104. Certificate Revocation
若 checker bug 發現:
Revoke(Ci).
105. Local Certificate 不等 Global Certificate
LocalCerts⇒GlobalCert.
106. 原因之一:Scope Incompatibility
每個 cert scope 不同。
107. Version Mismatch
cert A 對 v1,cert B 對 v2。
108. Circular Support
C1→C2→C1.
不是獨立支撐。
109. Semantic Drift
不同 cert 實際驗證了不同 claim。
110. Bridge Noncomposability
local bridges 都 legal,長鏈仍可能不 legal。
111. Certificate Composition
ComposeCert(C1,…,Cn,Γ,H)→{Pass,Fail,Undetermined,Conflicted}.
112. Hard Blocker 非補償
一個 critical type / permission / semantic gate fail:
不能用 100 個 soft supports 補過。
113. Local Verification 與 Global Verification
Paper 05:
Dverify=Dglobal.
本文保留。
114. Verified 仍不等 World-Level Truth
VerifiedΓ,T=AbsoluteTruth.
115. Formal Relative Truth
T⊢q
只表示 relative to T。
116. Empirical Claim 仍需 Reality Evidence
model proof 不能取代 measurement。
117. Simulation Verification 不能洗白 Reality Claim
Verifiedsim⇒Verifiedreal.
118. Evidence Transport Graph
GE=(VE,EE,τE).
119. Evidence Nodes
可以是:
- world outcome;
- aggregate;
- dataset;
- experiment;
- claim;
- certificate。
120. Edge Types
- replicate;
- calibrate;
- compare;
- validate;
- transport;
- contradict;
- aggregate。
121. Transport Debt
DT(q).
122. Transport Debt 可以包括
- calibration;
- causal justification;
- scale matching;
- distribution matching;
- external replication;
- measurement compatibility。
123. Cross-World Robustness 高,Transport Debt 仍可高
Ecross↑
不推出:
DT↓.
124. Real-World Calibration
known real cases 可以提升 transport evidence。
125. 但 Calibration 不等 Prospective Validation
用同資料調參再評分會過度樂觀。
126. Calibration / Test 分離
CalibrationData=ProspectiveTestData.
127. Prospective External Resolution
在看到 future outcome 前固定:
- claim;
- metric;
- tolerance;
- evidence rule。
128. Unknown-World Mass
UW.
129. 代表 Candidate Family 可能不完備
真 mechanism 可能沒被建模。
130. 即使所有 Current Worlds 獨立,也不消除 UW
Independence=Exhaustiveness.
131. Ensemble Closure Error
Eclosure=treating current ensemble as exhaustive.
132. 降低 Closure Error
可以加入:
- new ontology;
- new backend;
- new measurement;
- adversarial model;
- real anomaly;
- independent source。
133. 不是只加 Seeds
MoreSeeds=MoreOntologyCoverage.
134. Family Ablation
將 evidence 按 family 分組:
E=k⋃Fk.
135. Leave-One-Family-Out
依次移除 family。
136. Robustness
若移除任一 dominant family 仍成立:
LOFO Robust.
137. Family Sensitive
若去掉某 family 結論消失:
FamilySensitive.
138. Backend Ablation
移除 backend family。
139. Evaluator Ablation
移除 evaluator family。
140. Data Ablation
移除 source family。
141. Assumption Ablation
移除 assumption family。
142. Dependency Root Discovery
若多個 results 同時對某 upstream node 敏感,應標:
CommonModeRoot.
143. Evidence Diversity Budget
本文提出:
BV=(Bsupport,Bcounter,Breplicate,Bindependence,Btransport,Bexternal).
144. Confirmation Pressure
若:
Bsupport≫Bcounter,
consensus 可以被工程化。
145. Balanced 不代表平均
不同 claim type 應有不同 budget。
146. Safety Claim 可偏 Counterexample Budget
Bcounter↑.
147. Exploratory Claim 可偏 Diversity Budget
Bindependence↑.
148. Verification Debt
DV(q)=(Dexec,Dint,Ddep,Dcounter,Dspec,Dcert,Dglobal,Dtransport,Dexternal,Dunknown).
149. Dexec
execution integrity 未完成。
150. Dint
internal validation 不足。
151. Ddep
common-mode / independence 尚不清楚。
152. Dcounter
反例搜尋不足。
153. Dspec
spec / claim identity 不清楚。
154. Dcert
certificate 不完整或 stale。
155. Dglobal
local results 未閉合。
156. Dtransport
從 model/world 到 target 仍有 gap。
157. Dexternal
尚無 prospective / independent resolution。
158. Dunknown
未建模 mechanism family 風險。
159. Verification Capsule
VCap(q)=⟨Claim,ClaimType,QCap,EvidencePackets,DependenceProfile,EvidenceFamilies,Counterexamples,VerifierProfile,CertificateGraph,VerificationLayer,TransportScope,Debt,Unknowns,DecisionStatus,ReopenTriggers⟩.
160. VCap 不等單一 Confidence
VerificationCapsule=ScalarConfidence.
161. Decision Status
可包括:
{Unsupported,Candidate,InternallyValidated,CrossFamilyRobust,TransportQualified,ExternallySupported,Refuted,Conflicted,Undetermined}.
162. Externally Supported 仍不等 Eternal Truth
新的 evidence 仍可重開。
163. Verification Closure
VClose(q,τ,ρ).
164. Closure Conditions
至少:
- required layer reached;
- hard blockers addressed;
- dependence disclosed;
- counterexample policy satisfied;
- certificates composable;
- transport debt acceptable;
- unknowns explicit;
- reopen triggers stored。
165. VClose 不等 Final Truth
VClose=FinalTruth.
166. Reopen Triggers
- new evidence;
- external contradiction;
- new model family;
- checker bug;
- certificate revocation;
- ontology change;
- transport failure;
- regime shift。
167. Verification History
Htverify.
記錄:
- evidence added;
- family classification;
- counterexamples;
- ablations;
- cert changes;
- closure;
- reopen。
168. Verification Learning 不等 Truth Learning
系統可學會更有效 verification route。
但:
VerificationStrategyPreference=CanonicalTruth.
169. World Ensemble Learning Gate
world outcomes 不自動成 training target。
170. Generated Experience 不等 Real Experience
LearningFromGeneratedWorlds=LearningFromReality.
171. Recursive Verification Risk
若 AI 用自己生成的 worlds:
Generate→Verify→Train→Generate
可能自我封閉。
172. External Anchors
需要保留:
- independent sources;
- real measurements;
- externally maintained standards;
- adversarial evaluation;
- prospective tests。
173. Verification Layer 與 World Layer
WorldId / family / lineage 是 dependence basis。
174. Verification Layer 與 Computation Layer
backend / route / precision / approximation 進 evidence packet。
175. Verification Layer 與 Observation Layer
observer / sensor / operator dependence 必須保留。
176. Verification Layer 與 Domain Layer
VCap 更新:
Dverify,Dglobal.
177. Verification Layer 與 Projection Layer
projection sensitivity 是 diagnostic,不是新 evidence count。
178. Verification Layer 與 Physical Layer
physical calibration / measurement certificate 可作 external evidence。
179. Verification Layer 與 Memory Layer
所有 evidence / certificate / counterexample / invalidation 必須可重建。
180. Verification Layer 與 Execution Layer
Execution Plan 必須依:
VCap
決定:
- commit;
- stop;
- replan;
- escalate;
- external test。
181. WCO Verification State
正式:
VtWCO=⟨Etpkt,Gtdep,Gtprov,Gtcert,Ctclaim,Xtcounter,Tttransport,Utunknown,Dtverify,Htverify⟩.
182. WCO 現在形成八個能力/Runtime State
WtG,CtWF,OtG,DtWCO,PtG,PRSt,MtWCO,EOSt
加上:
VtWCO.
183. Verification 不是末端
驗證結果會回饋:
- world generation;
- computation routing;
- observation;
- domain debt;
- memory;
- plan。
184. Verification Feedback Loop
VCapt→Replan→World/Compute/Observe→Evidencet+1→VCapt+1.
185. MVP:Dependence-Aware Verification Runtime
沿用:
WA,WB,WC,WN.
186. Shared-Backend Family
先讓四 worlds 共用 backend。
187. Cross-Backend Family
新增:
WD
使用不同 simulator / model。
188. Observer Diversity
一組使用 same observation source。
另一組使用 independent measurement / parser。
189. Projection Diversity
同 evidence 生成:
不能增加 evidence count。
190. Verifier Diversity
使用:
- deterministic checker;
- property test;
- alternative implementation;
- human review。
191. Evidence Packet Store
每個 result 生成:
ei.
192. Dependence Graph
至少追:
- family;
- backend;
- data;
- evaluator;
- verifier。
193. Counterexample Injection
新增 high-validity rare failure world。
194. Majority Vote Baseline
比較:
Vote.
195. Dependence-Aware Baseline
比較:
Aggregate(Packets,Dependence,ClaimType,Counterexamples,Transport).
196. Family Ablation Test
leave-one-family-out。
197. Projection Laundering Test
同 source 投影三次。
系統必須仍計為同一 source family。
198. Verifier Common-Mode Test
兩 checker 共用故意錯誤 spec。
第三 checker 使用 independent spec translation。
199. Transport Test
simulation consensus 很高,但 external calibration fail。
不得升到 external support。
200. Unknown-World Test
加入 omitted mechanism。
確認:
UW
與 reopen。
201. Experiment 1 — Vote vs Dependence-Aware Aggregation
測錯誤 consensus。
202. Experiment 2 — More Seeds vs New Backend
比較 evidence diversification。
203. Experiment 3 — Model Names vs Provenance Families
測假多樣性。
204. Experiment 4 — Independent First Pass vs Deliberated Agents
測 communication contamination。
205. Experiment 5 — Projection Robustness
比較 view-sensitive claims。
206. Experiment 6 — Counterexample Preservation
確認 rare counterexample 不被 majority prune。
207. Experiment 7 — Checker Diversity
共同 spec bug vs independent specification audit。
208. Experiment 8 — Certificate Composition
local certs 全 pass,但製造 version mismatch。
global composition 應 fail。
209. Experiment 9 — Transport Debt
cross-world robustness 高,但 real calibration 缺失。
210. Experiment 10 — Verification Reopen
closure 後加入 real contradiction。
必須 reopen。
211. 可反駁性
本文會被削弱,如果:
- dependence-aware verification 在 shared-error tasks 中沒有比 majority / naive ensemble 更好的 calibration;
- evidence-family tracking 無法改善 common-mode failure diagnosis;
- counterexample preservation 不降低 false closure;
- projection sensitivity 對 representation-induced error 沒有 diagnostic value;
- verifier diversity profile 無法改善 checker common-mode detection;
- certificate composition layer 沒有比 local certificates 更高 correctness;
- transport debt 無法預測 simulation-to-reality failure;
- unknown-world / ensemble-closure state 沒有任何決策價值;
- simpler scalar confidence 在代表性 high-risk tasks 中完全等效。
212. 外部研究接口
Climate-model ensemble research 已明確處理 model performance 與 model interdependence,並指出模型數量增加不代表模型可視為等權獨立樣本。這是 dependence-aware multi-model aggregation 的成熟工程類比。
Bayesian predictive stacking 則提醒:在 M-open setting 中,真實資料生成機制可能不在候選模型集合內;因此驗證問題不應只被壓成「哪一個 candidate model 最真」。
NASA-STD-7009B 與 NASA-HDBK-7009B 將 modeling / simulation credibility、verification、validation、uncertainty 與 acceptance criteria 納入正式工程治理,也支持本文把「model 執行正確」「model 對 target 適切」「result 足以支撐 decision」分開處理。
本文不宣稱氣候模型 weighting、Bayesian stacking 或 NASA M&S standards 等同 WCO Verification Fabric。它們是方法學鄰近接口。
213. 本文不主張什麼
本文不主張:
- 多數決永遠錯;
- ensemble 沒有價值;
- 同 family replication 沒有價值;
- 不同 foundation model 一定相依;
- 不同 foundation model 一定獨立;
- Neff 有 universal estimator;
- 一個 counterexample 一定推翻 probabilistic claim;
- 所有 claim 都需要 R4 external resolution;
- projection sensitivity 等於 truth;
- checker diversity 保證 checker 正確;
- formal proof 等於 specification faithful;
- external observation 無 error;
- real data 永遠比 simulation 高品質;
- independence 等於 exhaustiveness;
- adversarial search 可以枚舉所有 failure;
- unknown-world mass 可以精確計算;
- VClose 等於 final truth;
- verification 一定單調;
- certificate 永久有效;
- WCO Verification Fabric 已完成 production implementation。
214. 核心非同一性
WorldCount=EvidenceCount=Truth.
ModelCount=IndependentModelCount.
ObserverAgreement=Truth.
ProjectionAgreement=IndependentEvidence.
VerifierAgreement=IndependentVerification.
Agreement=Independence=Validity=Transport.
CorrectExecution=AdequateModel=AdequateDecisionEvidence.
SpecificationCorrectness=ImplementationCorrectness=CheckerCorrectness.
LocalCert⇒GlobalCert.
VClose=FinalTruth.
215. 核心母式一:Evidence Packet
ei(q)=⟨EvidenceId,ClaimId,SourceClass,WorldId,RunId,Model,Data,Assumptions,Observer,ObservationOperator,Projection,Evaluator,Verifier,Provider,Communication,Outcome,Validity,Uncertainty,Transport,Certificates⟩.
216. 核心母式二:Dependence Vector
dij=(dL,dB,dM,dD,dA,dO,dP,dE,dV,dR,dC).
217. 核心母式三:Verification Ladder
Eexec→Eint→Ecross→Etrans→Eext.
且:
Eexec⇒Eint⇒Ecross⇒Etrans⇒Eext.
218. 核心母式四:Evidence Aggregation
Aggregate(EvidencePackets,Dependence,ClaimType,Counterexamples,Transport,AcceptanceContract).
219. 核心母式五:Verification Debt
DV(q)=(Dexec,Dint,Ddep,Dcounter,Dspec,Dcert,Dglobal,Dtransport,Dexternal,Dunknown).
220. 核心母式六:Verification Capsule
VCap(q)=⟨Claim,ClaimType,QCap,EvidencePackets,DependenceProfile,EvidenceFamilies,Counterexamples,VerifierProfile,CertificateGraph,VerificationLayer,TransportScope,Debt,Unknowns,DecisionStatus,ReopenTriggers⟩.
221. 核心母式七:WCO Verification State
VtWCO=⟨Etpkt,Gtdep,Gtprov,Gtcert,Ctclaim,Xtcounter,Tttransport,Utunknown,Dtverify,Htverify⟩.
222. 結論:驗證不是「同意的人變多」,而是「可替代的錯誤解釋變少」
如果:
100
個 worlds 同意,
類全域 AI 第一個問題不應只是:
100 比 0?
而應問:
這 100 個 worlds 來自幾個 evidence families?
共享哪些 backend、data、assumption、evaluator、observer 與 verifier?
如果拿掉 dominant family,結論還在嗎?
有沒有一個高-validity counterexample?
是否主動搜尋過反例?
換 projection 後結論是否仍穩定?
checker 是否共享 spec bug?
evidence 只在 simulation 內穩健,還是能 transport 到 target reality?
current ensemble 是否漏掉一整個 mechanism family?
所以:
Consensus without independence can be repetition, not evidence diversification.
而反過來:
Disagreement among sufficiently independent evidence families can be more informative than agreement among highly dependent ones.
真正成熟的 Verification Layer,不會把驗證理解成:
more votes.
而是理解成:
fewer surviving plausible explanations for why the conclusion could still be wrong.
這也是本文對「驗證」最核心的重新定義:
Verification=structured elimination of credible error explanations,
但永遠保留:
Unknown,Scope,TransportDebt,Reopenability.
因此:
Verification is not the accumulation of agreeing outputs.
而是:
the structured reduction of alternative error explanations through validity checks, dependence analysis, counterexample pressure, certificate composition, target transport, and external resolution.
至此,WCO 已具有:
- World Layer;
- Computation Layer;
- Observation Layer;
- Domain Layer;
- Projection Layer;
- Physical Layer;
- Memory Layer;
- Execution Layer;
- Verification Layer。
下一篇將進入:
Paper 11
治理層:全域認知不等於全域控制
正式回答:
當一個 AI 能看得更廣、算得更多、記得更久、驗證得更強時,為什麼它仍然不應因此自動取得更大的 authority?
223. 下一篇接口
Paper 11 將處理:
- cognition vs authority;
- capability vs permission;
- observation vs control;
- semantic authority;
- effect authority;
- delegated authority;
- authority graph;
- least authority;
- federated governance;
- human / AI / organization governance;
- approval / veto;
- revocation;
- authority inheritance;
- cross-world authority;
- bounded autonomy;
- responsibility / receipts;
- global cognition without global sovereignty。
參考文獻與內部前置研究
EveMissLab / Neo.K
- Neo.K × Aletheia,《WDC-05:跨世界證據:一致、反例、獨立性與證據轉移》,2026。
- Neo.K × Aletheia,《SWFR Paper 02:模擬世界族執行時》,2026。
- Neo.K × Aletheia,《WDC-07:世界集合學習》,2026。
- Neo.K × Aletheia,《DEST-01:多域知識判定論》,2026。
- Neo.K × Aletheia,《MWT-02:Global Legality Calculus》,2026。
- Neo.K × Aletheia,《WCO Paper 01–09》,2026。
- Neo.K × Aletheia,《投影計算論》,2026。
- Neo.K × Aletheia,《Global Observer Series C》,2026。
External Research Interfaces
- Knutti, R., Sedláček, J., Sanderson, B. M., Lorenz, R., Fischer, E. M., & Eyring, V. (2017). A climate model projection weighting scheme accounting for performance and interdependence. Geophysical Research Letters, 44, 1909–1918.
- Sanderson, B. M., Wehner, M., & Knutti, R. (2017). Skill and independence weighting for multi-model assessments. Geoscientific Model Development, 10, 2379–2395.
- Yao, Y., Vehtari, A., Simpson, D., & Gelman, A. (2018). Using Stacking to Average Bayesian Predictive Distributions. Bayesian Analysis, 13(3), 917–1007.
- NASA. NASA-STD-7009B — Standard for Models and Simulations. 2024.
- NASA. NASA-HDBK-7009B — NASA Handbook for Models and Simulations: An Implementation Guide for NASA-STD-7009B. 2026.
- Popper, K. (1959). The Logic of Scientific Discovery. Hutchinson.
- Mayo, D. G. (2018). Statistical Inference as Severe Testing. Cambridge University Press.
Paper 10 狀態:COMPLETE v0.1
下一篇:Paper 11 — 治理層:全域認知不等於全域控制
Canonical source:UTF-8 Markdown;數學 delimiter 僅使用 $...$ 與 $$...$$。