← Archive
lm-003147 · 2026-08

跨世界證據:一致、反例、獨立性與證據轉移

下載 MD 檔 ⬇

跨世界證據:一致、反例、獨立性與證據轉移

Cross-World Evidence: Agreement, Counterexamples, Independence, and Evidence Transport Across Runnable Worlds

Branching World Computation / World-Domain Cognitive Runtime
分支世界計算/世界域認知 Runtime 系列
WDC-05 / BWC-05 — Evidence Paper I

作者:Neo.K(許筌崴)
協作形式化:Aletheia
機構:一言諾科技有限公司(EveMissLab)
日期:2026-08-17
版本:v0.1
狀態:cross-world evidence / dependence / transport formalization


Canonical Non-Identity Statement

WDC-01 建立:

World Execution ResultReal-World Evidence\boxed{ \text{World Execution Result} \neq \text{Real-World Evidence} }

除非建立明示的 world-to-reality evidence transport contract。

WDC-02 建立:

CloneForkReplayCounterfactual Branch.\boxed{ \text{Clone} \neq \text{Fork} \neq \text{Replay} \neq \text{Counterfactual Branch}. }

WDC-04 再建立:

AmasterAilocalOiEiG.\boxed{ A^{master} \neq A_i^{local} \neq O_i \neq E_i \neq G. }

本文進一步建立:

World CountIndependent Evidence CountTruth.\boxed{ \text{World Count} \neq \text{Independent Evidence Count} \neq \text{Truth}. }

以及:

Cross-World AgreementIndependent ReplicationReal-World Validation.\boxed{ \text{Cross-World Agreement} \neq \text{Independent Replication} \neq \text{Real-World Validation}. }

本文不主張:

  • 100 個 worlds 支持某命題就使該命題為真;
  • 多數決是 cross-world evidence 的 universal aggregation rule;
  • 不同 world IDs 自動代表獨立 evidence sources;
  • 不同 seeds 等於不同模型;
  • 不同 foundation models 一定有獨立 error;
  • 同 backend 的多個 branches 沒有 evidence value;
  • counterexample 必然勝過所有 supporting worlds;
  • one scalar evidence score 足以描述所有 cross-world evidence;
  • simulation calibration 可完全消除 model-form uncertainty;
  • NASA、climate-model ensemble 或 Bayesian stacking 等既有方法等同 WDC;
  • world-to-reality transport 可以由 simulation consensus 自動取得;
  • 任何跨世界結果都能被合併成 probability。

摘要

到 WDC-04 為止,World-Domain Cognitive Runtime 已經建立:

  1. runnable world;
  2. branching world lineage;
  3. world-domain Governor;
  4. master / local actor / observer / evaluator / Governor 的 role separation;
  5. branch blindness 與 explicit cross-world channels;
  6. bounded authority 與 nested-agent architecture。

於是下一個問題終於可以被嚴格提出:

如果有很多 runnable worlds 都對同一命題輸出結果,這些結果到底構成多少證據?

最粗糙的方法是計數:

N+(q)=#{Wi:Outcome(Wi) supports q}.\boxed{ N_+(q) = \#\{ W_i: Outcome(W_i)\text{ supports }q \}. }

然後宣稱:

100 個 world 中 97 個支持 qq,所以 qq 有 97% 機率成立。

本文拒絕這個推論。

因為 97 個 worlds 可能:

  • 全部 fork 自同一 root;
  • 使用同一 dynamics backend;
  • 使用同一 foundation model;
  • 使用同一 dataset;
  • 使用同一 hidden assumption;
  • 由同一 evaluator 評分;
  • 只改了 random seed;
  • 只是 near-duplicate world contracts。

此時:

97 worlds⇏97 independent evidence units.\boxed{ 97\text{ worlds} \not\Rightarrow 97\text{ independent evidence units}. }

本文定義 Cross-World Evidence Packet

ei(q)=(Wi,Yi,Viint,Li,Bi,Ai,Ei,Ri,Ti,Ui).\boxed{ \mathfrak e_i(q) = ( W_i, Y_i, V_i^{int}, L_i, B_i, A_i, E_i, R_i, T_i, U_i ). }

其中:

  • WiW_i:world identity / contract;
  • YiY_i:world outcome relative to claim qq
  • ViintV_i^{int}:within-world validity / contract fidelity;
  • LiL_i:lineage information;
  • BiB_i:backend / model family;
  • AiA_i:assumption / parameter family;
  • EiE_i:evaluator identity / independence profile;
  • RiR_i:replication / run information;
  • TiT_i:world-to-target transport scope;
  • UiU_i:uncertainty / unknowns。

因此 cross-world evidence aggregation 的輸入不是:

{Y1,,Yn}\boxed{ \{Y_1,\ldots,Y_n\} }

而是:

{e1,,en}.\boxed{ \{ \mathfrak e_1, \ldots, \mathfrak e_n \}. }

本文再定義 world-pair dependence:

dijW=DE(ei,ej)[0,1].\boxed{ d_{ij}^{W} = \mathcal D_E( \mathfrak e_i, \mathfrak e_j ) \in [0,1]. }

其中:

dijW=0d_{ij}^{W}=0

表示在 declared evidence dimensions 上高度獨立;

dijW=1d_{ij}^{W}=1

表示近乎 evidence-duplicate。

依賴不應只看 output correlation,而至少拆成:

dij=(dL,dB,dM,dD,dA,dE,dC).\boxed{ \mathbf d_{ij} = ( d_L, d_B, d_M, d_D, d_A, d_E, d_C ). }

分別表示:

  • dLd_L:lineage dependence;
  • dBd_B:backend dependence;
  • dMd_M:model / architecture dependence;
  • dDd_D:data dependence;
  • dAd_A:assumption dependence;
  • dEd_E:evaluator dependence;
  • dCd_C:communication / context dependence。

因此:

different outputindependent evidence,\boxed{ \text{different output} \neq \text{independent evidence}, }

同樣:

same outputdependent evidence\boxed{ \text{same output} \neq \text{dependent evidence} }

也不必然成立。

真正要問的是:

為什麼這兩個 worlds 會犯相同錯誤?它們共享了哪些 upstream causes?

本文提出概念性 Effective Evidence Count

Neff=Neff(DW),\boxed{ N_{\mathrm{eff}} = \mathcal N_{\mathrm{eff}} ( \mathbf D_W ), }

其中:

DW=[dijW]n×n.\mathbf D_W = [ d_{ij}^{W} ]_{n\times n}.

本文不指定一個 universal NeffN_{\mathrm{eff}} 公式,因為 dependence type、claim type 與 aggregation semantics 可能不同。但任何宣稱:

NN worlds agree

的報告,都應至少同時報:

NandNeffestimated.\boxed{ N \quad\text{and}\quad N_{\mathrm{eff}}^{estimated}. }

或明確說明無法可靠估計 NeffN_{\mathrm{eff}}

本文進一步提出:

Evidence Family

若一組 worlds:

Fk={Wk,1,,Wk,m}\mathcal F_k = \{ W_{k,1},\ldots,W_{k,m} \}

共享:

  • root;
  • backend;
  • core data;
  • evaluator;
  • central assumptions;

則它們可以先被視為一個:

Evidence Family.\boxed{ \text{Evidence Family}. }

family 內的多 runs 可以提高:

  • stochastic precision;
  • numerical stability;
  • tail estimation;
  • robustness within contract;

但不能被錯算成同等數量的 independent model confirmations。

因此:

Replication within a world familyReplication across world families.\boxed{ \text{Replication within a world family} \neq \text{Replication across world families}. }

本文正式區分至少四層 replication:

  1. R0 — Re-run Replication:同 world spec / backend / evaluator,不同 run;
  2. R1 — Branch Replication:同 root / backend,不同 controlled intervention 或 seed family;
  3. R2 — Backend Replication:不同 dynamics / simulator / model family;
  4. R3 — Assumption / Evaluator Replication:不同 assumptions、data pathways、evaluator families;
  5. R4 — External Target Replication:真實資料或真實實驗 resolution。

因此:

R0<R1<R2<R3<R4\boxed{ R0<R1<R2<R3<R4 }

只作 evidence-diversification hierarchy,不表示任何固定 numerical weight。

本文亦正式建立 Counterexample Burden

假設:

100100

個 worlds 支持:

q,q,

但一個高-validity、high-independence、matched-contract world:

WcW_c

產生:

¬q.\neg q.

不能只用:

100:1100:1

票數把它抹掉。

定義:

CE(q)={ei:Yi is materially inconsistent with q}.\boxed{ CE(q) = \{ \mathfrak e_i: Y_i\text{ is materially inconsistent with }q \}. }

對每個 counterexample,至少問:

  • 是 runtime bug 嗎?
  • 是 contract mismatch 嗎?
  • 是 rare but valid region 嗎?
  • 是不同 ontology 嗎?
  • 是 world-to-reality irrelevant 嗎?
  • 還是它揭示了 supporting worlds 共享的 blind spot?

本文提出:

Counterexample Escalation Rule

若 counterexample:

ece_c

同時具有:

Vcint,Independencec,Transportc,\boxed{ V_c^{int}\uparrow, \qquad Independence_c\uparrow, \qquad Transport_c\uparrow, }

則 Governor 應提高:

  • replication budget;
  • adversarial worlds;
  • backend diversity;
  • evaluator review;

而不是直接當 outlier 刪除。

本文再定義 cross-world evidence profile:

EW(q)=(Cq,Iq,Rq,Kq,Xq,Tq,Uq).\boxed{ \mathbf E_W(q) = ( C_q, I_q, R_q, K_q, X_q, T_q, U_q ). }

其中:

  • CqC_q:cross-world consistency;
  • IqI_q:independence;
  • RqR_q:replication depth;
  • KqK_q:counterexample burden;
  • XqX_q:backend / assumption diversity;
  • TqT_q:transport strength;
  • UqU_q:remaining uncertainty。

這是一個 vector,而不是單一「世界信心分數」。

本文拒絕:

EvidenceScore(q)=0.93\boxed{ EvidenceScore(q)=0.93 }

如果沒有說明:

0.93 到底是 world-internal probability、ensemble frequency、posterior predictive weight、real-world calibrated probability,還是純 heuristic。

本文建立 Cross-World Agreement Matrix

Aij(q)=Agreement(Yi,Yjq).\boxed{ A_{ij}(q) = Agreement( Y_i, Y_j \mid q ). }

與 dependence matrix:

DijW.\boxed{ D_{ij}^{W}. }

兩者必須分開。

一個高度一致的 cluster:

Aij1A_{ij}\approx1

若:

DijW1,D_{ij}^{W}\approx1,

可能只是同源共錯。

反過來,多個高度獨立 worlds:

DijW0D_{ij}^{W}\approx0

若仍高度一致:

Aij1,A_{ij}\approx1,

才是更強的 cross-world robustness signal。

本文將這個差異寫成:

Agreement Strength=f(Agreement,Independence),\boxed{ \text{Agreement Strength} = f( Agreement, Independence ), }

但不預設一個 universal ff

外部 multi-model ensemble research 對這一點有直接而成熟的技術鄰接。氣候模型 ensemble 長期面臨「model democracy」問題:多個 models 可能共享 code、parameterizations、tuning choices,因此不能被自然當成完全獨立的一人一票。Knutti 等人的 weighting work 以及 Sanderson、Wehner、Knutti 的 multi-model assessment 明確把 model performance 與 model interdependence 一起納入權重;後續 ClimSIPS 也把 independence、performance 與 spread 明確分離,用於 CMIP model subselection。這些工作不等於 WDC,但它們提供一個非常關鍵的外部先例:

ensemble sizeindependent model count.\boxed{ \text{ensemble size} \neq \text{independent model count}. }

本文亦引用 Bayesian predictive stacking 作為另一種 aggregation 鄰接。Yao、Vehtari、Simpson 與 Gelman 的 stacking 方法不是假設 candidate models 中必有一個是真實 data-generating model,而是根據 out-of-sample predictive performance 組合 predictive distributions。WDC 不直接採用 stacking 作 universal world aggregator,但接受其一個重要方法論警告:

當真實生成機制可能不在 candidate worlds 裡時,不應把「哪個 world 最真」當作唯一 aggregation 問題。

這與 WDC-01 的 unknown / transport boundary 相容。

本文同時引入 simulation credibility boundary。NASA-STD-7009B 要求 modeling / simulation 活動明示 verification、validation、uncertainty、acceptance criteria 與 credibility practices;2026 年的 NASA-HDBK-7009B 進一步作為實作 guide。這些標準不是 WDC,但它們支持本文的核心證據分層:

model executes correctlymodel represents target adequatelyresult supports target decision.\boxed{ \text{model executes correctly} \neq \text{model represents target adequately} \neq \text{result supports target decision}. }

因此本文將 world evidence 分為:

Layer 1 — Execution Integrity

Eexec.\boxed{ E_{exec}. }

問:

world 是否照自己的 contract 正確執行?

Layer 2 — Internal Validation

Eint.\boxed{ E_{int}. }

問:

within-world outcomes 是否可重現、穩定、與 internal known cases 一致?

Layer 3 — Cross-World Robustness

Ecross.\boxed{ E_{cross}. }

問:

在不同 lineage / backend / assumptions / evaluators 下,結論是否仍保持?

Layer 4 — Target Transport

Etrans.\boxed{ E_{trans}. }

問:

world evidence 是否能合理搬到 target system / reality?

Layer 5 — External Resolution

Ereal.\boxed{ E_{real}. }

問:

真實 experiment / data / deployment 是否支持?

因此:

Eexec⇏Eint⇏Ecross⇏Etrans⇏Ereal.\boxed{ E_{exec} \not\Rightarrow E_{int} \not\Rightarrow E_{cross} \not\Rightarrow E_{trans} \not\Rightarrow E_{real}. }

後一層需要額外 evidence,而不是自動包含。

本文再建立:

World Evidence Transport Graph

令:

GE=(VE,EE,τE).\boxed{ G_E = ( V_E, E_E, \tau_E ). }

node 可以是:

  • world outcome;
  • world family aggregate;
  • external dataset;
  • real experiment;
  • claim。

edge type 可以是:

  • replicate;
  • calibrate;
  • compare;
  • validate;
  • transport;
  • contradict;
  • aggregate。

這使 cross-world evidence 不再只是:

一個表格裡 100 行 simulation results。

而是:

a provenance-bearing evidence graph.\boxed{ \text{a provenance-bearing evidence graph}. }

本文也正式建立 Evidence Transport Debt

DT(q)\boxed{ D_T(q) }

表示從目前最高 evidence layer 到 target real-world claim 仍缺少多少:

  • calibration;
  • validation;
  • causal justification;
  • scale matching;
  • distribution matching;
  • external replication。

DTD_T 不要求 universal scalar,可作 structured debt ledger。

因此一個 world consortium 可以:

Ecross\boxed{ E_{cross}\uparrow }

但:

DT(q) remains high.\boxed{ D_T(q) \text{ remains high}. }

例如 20 個 independent social simulations 全都預測同一制度結果,但沒有 real-world calibration;它們構成有價值的 cross-model robustness evidence,但還不是 real-world probability。

本文建立 Cross-World Consensus Classes

C0 — Single-Family Agreement

大量 worlds,但共享主要 backend / assumptions。

C1 — Multi-Branch Agreement

多 interventions / seeds 下穩定。

C2 — Multi-Backend Agreement

不同 simulators / model families 一致。

C3 — Multi-Assumption / Multi-Evaluator Agreement

不同 assumption families、不同 evaluator families 一致。

C4 — Cross-World + External Calibration

world consortium 在 known real cases 上校準。

C5 — Prospective External Resolution

新 real-world evidence 在預先註冊條件下支持。

這不是 truth ladder,而是:

evidence diversification ladder.\boxed{ \text{evidence diversification ladder}. }

本文進一步要求 Pre-Registered Claim Equivalence

若 world WiW_i 輸出:

yiy_i

與 world WjW_j 輸出:

yj,y_j,

要說它們「支持同一命題」之前,必須先定義:

ϕ(y),dϕ,ϵq.\boxed{ \phi(y), \quad d_\phi, \quad \epsilon_q. }

即:

  • outcome representation;
  • comparison metric;
  • equivalence tolerance。

否則容易出現:

每個 world 結論都不完全一樣,但事後把它們全部歸類成「差不多支持」。

本文稱:

Consensus Stretching

若 equivalence contract 在看完結果後不斷放寬,使更多 worlds 被算成 support,就構成 post-hoc evidence inflation。

因此:

pre-register equivalence before aggregation.\boxed{ \text{pre-register equivalence before aggregation}. }

本文亦正式建立 World Voting Failure

假設:

W1,,W100W_1,\ldots,W_{100}

來自同一 backend family;

W101W_{101}

來自完全不同 scientific model。

單純 majority:

100:1100:1

可能完全錯誤。

因此:

one world one vote\boxed{ \text{one world one vote} }

不是 WDC default。

更合理的最低要求是:

one evidence familyone world.\boxed{ \text{one evidence family} \neq \text{one world}. }

即使最終使用 weighting,也要先把 dependence 寫出來。

本文再定義:

Independent Counterevidence Search

在世界 ensemble 已經高度支持:

qq

時,Governor 不應只 spawn 更多:

supporting variants。

而應主動產生:

W¬qadv\boxed{ \mathcal W_{\neg q}^{adv} }

尋找:

  • weakest assumptions;
  • failure regions;
  • alternative causal mechanisms;
  • adversarial agent policies;
  • backend-disagreement regimes。

如果經 adversarial search 仍難找到高-validity counterexample,cross-world robustness 才真正增加。

因此:

Consensus+Adversarial Failure Search>Consensus Alone\boxed{ \text{Consensus} + \text{Adversarial Failure Search} > \text{Consensus Alone} }

只作方法論偏序,不作 universal theorem。

本文亦建立 World Family Collapse Test

將 worlds 按 dependence 聚類:

W=F1Fm.\boxed{ \mathcal W = \mathcal F_1 \cup \dots \cup \mathcal F_m. }

然後比較:

NN

與:

m.m.

如果:

N=1000,m=2,N=1000, \qquad m=2,

報告應說:

1000 runs / worlds across 2 major evidence families,

而不是:

1000 independent simulations confirm the result。

本文提出 Backend Ablation

依次移除:

  • backend family;
  • evaluator family;
  • root lineage family;
  • data family;

重新聚合。

若結論只在某單一 family 存在時成立:

family-sensitive evidence.\boxed{ \text{family-sensitive evidence}. }

若移除任意單一 family 後仍成立:

leave-one-family-out robustness.\boxed{ \text{leave-one-family-out robustness}. }

本文也提出 Cross-World Evidence Packet 最小工程格式:

claim_id
world_id
world_family_id
root_lineage
branch_path
backend
model_versions
data_sources
assumption_family
evaluator_id
evaluator_independence
run_count
seed_policy
outcome
outcome_equivalence_class
internal_validity
uncertainty
counterexample_status
transport_scope
external_calibration
evidence_level

聚合報告則至少包含:

claim_id
total_worlds
total_runs
evidence_families
estimated_dependence
agreement_by_family
counterexamples
backend_ablation
evaluator_ablation
external_validation
transport_debt
unknowns

本文最後提出:

Cross-World Evidence Principle

Cross-world evidence strength depends not only on how many worlds agree, but on how independently they were constructed, how validly they ran, how diversely they represent assumptions, how strongly counterexamples were sought, and how well their conclusions transport to the target domain.

No Model Democracy by Default Principle

One world, one vote is not a default evidence rule when worlds share lineage, backends, data, assumptions, evaluators, or communication channels.

Counterexample Preservation Principle

A valid, independent counterexample must not be erased by numerical majority; it should trigger diagnosis, replication, and model-family expansion.

Transport Separation Principle

Cross-world robustness and real-world validity are distinct evidence layers. Agreement among simulations cannot by itself erase world-to-reality transport debt.

下一篇 WDC-06 將回到 WDC-03 刻意留下的深層問題:

Which Worlds Deserve Computation?

《哪些世界值得被計算:世界投資組合、探索—驗證與計算價值》

它不再只是 Governor API,而會正式研究:

  • value of computation;
  • world portfolio;
  • exploration / exploitation / verification;
  • rare-world preservation;
  • expected regret;
  • information geometry;
  • stopping / expansion criteria。

關鍵詞: Cross-World Evidence、Model Dependence、Ensemble Independence、Counterexample、Simulation Credibility、Evidence Transport、Model Weighting、World Families、WDC


1. 100 個世界支持同一命題,第一個問題不是「幾票?」

第一個問題應是:

這 100 個 worlds 到底有多獨立?\boxed{ \text{這 100 個 worlds 到底有多獨立?} }

2. World Identity 不等於 Evidence Identity

WDC-02 保證:

IDiIDj.ID_i \neq ID_j.

但這只說:

runtime / lineage identity 不同。


3. 它沒有保證

ErroriErrorj.\boxed{ Error_i \perp Error_j. }

4. Shared Backend

如果:

Bi=Bj,B_i=B_j,

可能有 shared model-form error。


5. Shared Data

Di=DjD_i=D_j

可能有同 data bias。


6. Shared Root

同一 parent checkpoint:

Rooti=RootjRoot_i=Root_j

具有 lineage dependence。


7. Shared Evaluator

Ei=EjE_i=E_j

可能有 evaluation bias。


8. Shared Master Hint

WDC-04 已說:

MasterHintMasterHint

可以污染 branches。


9. Shared Communication

cross-world channels:

CijcrossC_{ij}^{cross}

也可能讓 errors 相關。


10. Therefore

different worldsindependent evidence.\boxed{ \text{different worlds} \neq \text{independent evidence}. }

11. Evidence Packet

對 claim:

q,q,

world:

WiW_i

產生:

ei(q).\boxed{ \mathfrak e_i(q). }

12. Outcome Component

Yi(q){Support,Counter,Inconclusive,Invalid}Y_i(q) \in \{ Support, Counter, Inconclusive, Invalid \}

作最簡離散版。


13. Continuous Outcome

也可:

YiRk.\boxed{ Y_i\in\mathbb R^k. }

14. Internal Validity

Viint.\boxed{ V_i^{int}. }

15. Invalid World

若:

Viint0,V_i^{int}\approx0,

不能因為它支持 hypothesis 就算票。


16. Evidence Family

定義:

F(Wi)\boxed{ F(W_i) }

按:

  • backend;
  • model;
  • data;
  • assumptions;
  • root;
  • evaluator;

聚類。


17. Family Partition

W=kFk.\boxed{ \mathcal W = \bigsqcup_k \mathcal F_k. }

若能建立硬 partition。


18. More Realistically Soft Families

world 可同時共享:

  • backend family A;
  • data family B。

19. So Dependence Is Graph Better Than Partition

GDW=(VW,ED,wD).\boxed{ G_D^W = ( V_W, E_D, w_D ). }

20. Edge Weight

wD(i,j)=dijW.w_D(i,j)=d_{ij}^{W}.

21. Dependence Dimensions

dij=(dL,dB,dM,dD,dA,dE,dC).\boxed{ \mathbf d_{ij} = ( d_L,d_B,d_M,d_D,d_A,d_E,d_C ). }

22. Lineage Dependence

dL.d_L.

23. Backend Dependence

dB.d_B.

24. Model Dependence

dM.d_M.

25. Data Dependence

dD.d_D.

26. Assumption Dependence

dA.d_A.

27. Evaluator Dependence

dE.d_E.

28. Communication Dependence

dC.d_C.

29. One Number Can Be a Projection

dij=g(dij).\boxed{ d_{ij} = g( \mathbf d_{ij} ). }

30. But Raw Vector Must Remain Auditable


31. Climate Ensemble Analogy

Climate multi-model ensembles often contain:

  • related model families;
  • shared parameterizations;
  • shared tuning practices;
  • duplicated / near-related models。

32. Therefore Model Democracy Can Overcount

equal weighting assumes more independence than actually exists。


33. Knutti 2017

提出 weighting 同時考慮:

Performance+Interdependence.\boxed{ Performance + Interdependence. }

34. Sanderson–Wehner–Knutti 2017

explicit skill + independence weighting。


35. WDC Analogy

world aggregation 至少要問:

Quality+Independence.\boxed{ Quality + Independence. }

36. But WDC Has More Dependence Channels

worlds 還可能共享:

  • evaluator;
  • local agent;
  • prompt;
  • master;
  • fork lineage。

37. So Climate Weighting Is Analogue, Not Drop-In Formula


38. Model Democracy Failure Example

100 worlds:

W1,,W100W_1,\ldots,W_{100}

都來自 backend:

MA.M_A.

39. One World

W101W_{101}

來自:

MB.M_B.

40. If A Has Shared Bug

100:1 vote can be nonsense。


41. Backend Family Vote

minimum alternative:

先看:

FamilyAvsFamilyB.\boxed{ Family_A vs Family_B. }

42. But Even Family Vote Is Not Universal

不同 family quality 不同。


43. Performance Matters

world 在 historical calibration:

Qical.Q_i^{cal}.

44. Independence Matters

Qiind.Q_i^{ind}.

45. Transport Matters

Qitrans.Q_i^{trans}.

46. Counterexample Matters

Qicounter.Q_i^{counter}.

47. Evidence Vector

qiE=(Qint,Qcal,Qind,Qtrans,Qeval,Qcounter).\boxed{ \mathbf q_i^E = ( Q_{int}, Q_{cal}, Q_{ind}, Q_{trans}, Q_{eval}, Q_{counter} ). }

48. No Universal Weight


49. Re-runs

同 world spec:

Run1,,Runn.Run_1,\ldots,Run_n.

50. What Re-runs Tell Us

  • stochastic variance;
  • numerical stability;
  • rare-event rate under model。

51. What Re-runs Do Not Tell Us

  • model-form validity;
  • backend independence;
  • reality validity。

52. Therefore R0

Re-run replication。


53. R1

branch replication。


54. R2

backend replication。


55. R3

assumption / evaluator replication。


56. R4

external target replication。


57. Evidence Diversification

Depthrep\boxed{ Depth_{rep} }

應報是哪一層。


58. 1000 Seeds Can Be R0 Only


59. This Is Still Useful

但別叫:

1000 independent scientific models。


60. Agreement Matrix

Aij.\boxed{ A_{ij}. }

61. High Agreement

Aij1.A_{ij}\approx1.

62. Dependence Matrix

Dij.D_{ij}.

63. Four Cases

Case I

high agreement + high dependence。


64. Interpretation

shared-family consensus。


65. Case II

high agreement + low dependence。


66. Interpretation

stronger robustness。


67. Case III

low agreement + high dependence。


68. Interpretation

same family unstable / stochastic / sensitive。


69. Case IV

low agreement + low dependence。


70. Interpretation

genuine model uncertainty / assumption disagreement。


71. This Is Much More Informative Than Majority Count


72. Effective Evidence Count

概念:

NeffN.\boxed{ N_{eff} \le N. }

73. If Fully Independent

NeffN.N_{eff}\approx N.

74. If Perfect Duplicates

Neff1.N_{eff}\approx1.

75. Intermediate

depends on structure。


76. Do Not Pretend Precision

若 dependence 未知,

報:

Neff unresolved.\boxed{ N_{eff}\text{ unresolved}. }

77. Counterexample

定義:

ec:YcCounter(q).\boxed{ e_c : Y_c \in Counter(q). }

78. Not Every Opposite Output Is Valid Counterexample

可能:

  • runtime failure;
  • contract mismatch;
  • unsupported evaluator。

79. Counterexample Validity

VcCE.\boxed{ V_c^{CE}. }

80. Strong Counterexample

需要:

  • high internal validity;
  • relevant contract;
  • independent source;
  • reproducibility。

81. One Strong Counterexample Can Matter More Than Many Duplicates


82. But It Does Not Automatically Falsify Probabilistic Claims

若:

qq

是:

outcome occurs 99% of time,

one failure expected。


83. Claim Type Matters


84. Universal Claim

xP(x).\forall x\,P(x).

單一合法反例可有非常高 logical burden。


85. Probabilistic Claim

需要 frequency / model semantics。


86. Existential Claim

single positive witness may dominate。


87. Therefore Evidence Aggregation Must Know Claim Type

Type(q){Universal,Existential,Probabilistic,Comparative,Causal,Forecast}.\boxed{ Type(q) \in \{ Universal, Existential, Probabilistic, Comparative, Causal, Forecast \}. }

88. Universal World Voting Is Meaningless


89. Claim-Type-Aware Aggregation

Agg({ei},Type(q)).\boxed{ Agg( \{\mathfrak e_i\}, Type(q) ). }

90. Causal Claims

need:

  • intervention semantics;
  • matched branches;
  • transport assumptions。

91. Forecast Claims

need:

  • pre-registration;
  • resolution;
  • calibration。

92. Comparative Claims

need:

  • matched budgets;
  • evaluator equivalence;
  • controlled deltas。

93. Consensus Equivalence

world outputs not necessarily same representation。


94. Define Feature Map

ϕ:YiZq.\boxed{ \phi: Y_i \rightarrow Z_q. }

95. Distance

dϕ.\boxed{ d_\phi. }

96. Equivalence

yiqyj    dϕ(ϕ(yi),ϕ(yj))ϵq.\boxed{ y_i\sim_q y_j \iff d_\phi( \phi(y_i), \phi(y_j) ) \le \epsilon_q. }

97. Must Freeze Before Seeing Aggregate


98. Consensus Stretching

post-hoc increase:

ϵq\epsilon_q\uparrow

until desired consensus。


99. Forbidden Without Disclosure


100. Support Class

Sq={Wi:Yiq}.\boxed{ S_q = \{ W_i: Y_i\sim q \}. }

101. Counter Class

Cq.\boxed{ C_q. }

102. Inconclusive Class

Iq.\boxed{ I_q. }

103. Invalid Class

Xq.\boxed{ X_q. }

104. Do Not Force Every World Into Support/Reject


105. Inconclusive Is Real Category


106. Invalid Is Not Counter

simulator bug:

Invalid.Invalid.

not:

Counter.Counter.

107. Cross-World Evidence Profile

EW(q)=(Cq,Iq,Rq,Kq,Xq,Tq,Uq).\boxed{ \mathbf E_W(q) = ( C_q, I_q, R_q, K_q, X_q, T_q, U_q ). }

108. CqC_q

consistency。


109. IqI_q

independence。


110. RqR_q

replication depth。


111. KqK_q

counterexample burden。


112. XqX_q

cross-backend / assumption diversity。


113. TqT_q

transport strength。


114. UqU_q

remaining uncertainty。


115. Evidence Frontier

world claim can lie on Pareto frontier across these dimensions。


116. No Single Confidence by Default


117. NASA M&S Credibility Calibration

NASA-STD-7009B explicitly covers:

  • model verification;
  • validation;
  • uncertainty;
  • credibility;
  • acceptance criteria。

118. Why Important

它把:

simulation output exists

與:

simulation is credible enough for a decision

分開。


119. WDC Extends This Across Many Worlds


120. Execution Integrity

Eexec.E_{exec}.

121. Internal Validation

Eint.E_{int}.

122. Cross-World Robustness

Ecross.E_{cross}.

123. Transport

Etrans.E_{trans}.

124. External Resolution

Ereal.E_{real}.

125. Non-Implication Chain

Eexec⇏Eint⇏Ecross⇏Etrans⇏Ereal.\boxed{ E_{exec} \not\Rightarrow E_{int} \not\Rightarrow E_{cross} \not\Rightarrow E_{trans} \not\Rightarrow E_{real}. }

126. Execution Integrity Example

code matches specification。


127. But Specification Can Be Wrong


128. Internal Validation Example

historical benchmark reproduced。


129. But Future Regime Can Shift


130. Cross-World Robustness Example

different backends agree。


131. But All May Omit Same real mechanism


132. Transport Validation Needed


133. NASA-HDBK-7009B 2026

提供 NASA-STD-7009B implementation guidance。


134. WDC Use

不是 adopt as universal standard。

而是借:

credibility must be explicit.\boxed{ \text{credibility must be explicit}. }

135. World Evidence Level

本文建議:

CWE-0

single world output。


136. CWE-1

reproducible within-world runs。


137. CWE-2

multiple branches / seeds, same family。


138. CWE-3

independent backend / assumption families。


139. CWE-4

cross-world evidence + calibrated target cases。


140. CWE-5

pre-registered prospective external resolution。


141. This Is Not Universal Truth Ladder

它是 WDC evidence workflow。


142. Bayesian Stacking External Analogy

Yao et al. discuss M-open setting:

true data-generating process may not be among candidate models。


143. WDC Analogy

real world may not be represented by any current world family。


144. Therefore Keep

Uworld\boxed{ U_{world} }

unknown / omitted model region。


145. World Ensemble Is Not Exhaustive

{W1,,Wn}all possible explanatory worlds.\boxed{ \{W_1,\ldots,W_n\} \neq \text{all possible explanatory worlds}. }

146. Stacking Predictive Distributions

combines models based on predictive performance。


147. WDC Does Not Adopt as Default

because world outcomes can be:

  • non-probabilistic;
  • causal;
  • symbolic;
  • adversarial。

148. But It Demonstrates

aggregation need not mean:

pick one model as true。


149. Model Selection vs Model Combination

SelectBestCombineEvidence.\boxed{ SelectBest \neq CombineEvidence. }

150. WDC Can Keep Multiple Surviving World Families


151. Climate Model Independence

Knutti / Sanderson work explicitly considers model interdependence。


152. Why Climate Models Dependent?

  • shared code;
  • parameterizations;
  • tuning practices;
  • genealogies。

153. WDC Worlds Have Even More Shared Structure

  • prompts;
  • foundation models;
  • datasets;
  • agents;
  • evaluators。

154. So Dependence Audit Is Mandatory for Strong Claims


155. Skill vs Independence

climate weighting separates:

Performance\boxed{ Performance }

and:

Independence.\boxed{ Independence. }

156. WDC Analogy

a high-performing world family can still be overrepresented。


157. Independence vs Diversity

different outputs alone don't establish independence。


158. Diversity Must Be Causal/Structural, Not Cosmetic


159. Cosmetic Diversity

change:

  • color;
  • wording;
  • seed;

without changing error source。


160. Structural Diversity

change:

  • model family;
  • dynamics;
  • data;
  • evaluator;
  • assumptions。

161. World Family Identification

Governor can cluster using metadata first。


162. Then outcome similarity second。


163. Never Use Outcome Similarity Alone

because true models may converge to same result。


164. Dependence vs Convergence

important distinction。


165. Two Independent Models Can Agree Because Reality Constrains Them


166. Two Dependent Models Can Agree Because They Share Bug


167. Need Provenance


168. Evidence Graph

GE=(VE,EE,τE).\boxed{ G_E = ( V_E, E_E, \tau_E ). }

169. Evidence Nodes

  • run;
  • world;
  • family;
  • dataset;
  • experiment;
  • claim。

170. Evidence Edges

τE{Supports,Counters,Replicates,Calibrates,Validates,Transports,DependsOn,Aggregates}.\boxed{ \tau_E \in \{ Supports, Counters, Replicates, Calibrates, Validates, Transports, DependsOn, Aggregates \}. }

171. Supports Edge

world outcome supports claim under contract。


172. Counters Edge

materially challenges claim。


173. Replicates

same or nearby protocol。


174. Calibrates

known target case。


175. Validates

checks model adequacy。


176. Transports

moves evidence scope。


177. DependsOn

marks shared upstream source。


178. Aggregates

creates meta-evidence node。


179. Evidence Graph Must Remain Provenance-Aware


180. Aggregation Node

Mq=Aggregate(e1,,en).\boxed{ M_q = Aggregate( e_1,\ldots,e_n ). }

181. It Should Store Inputs

not erase them。


182. No Evidence Flattening

don't replace 100 packets with:

97% support。


183. Transport Graph

world:

WW

to target:

R.R.

184. Transport Edge

WTR.\boxed{ W \xrightarrow{ \mathcal T } R. }

185. Transport Conditions

  • scale match;
  • variable match;
  • dynamics match;
  • distribution match;
  • intervention match;
  • agent behavior match。

186. Transport Debt

DT.\boxed{ D_T. }

187. Debt Components

DT=(Dscale,Ddyn,Dagent,Ddata,Dcausal,Ddomain).\boxed{ \mathbf D_T = ( D_{scale}, D_{dyn}, D_{agent}, D_{data}, D_{causal}, D_{domain} ). }

188. Cross-World Agreement Can Reduce Some Debt

e.g. model-form sensitivity。


189. But Cannot Erase Missing Real Mechanism


190. Example

100 traffic simulators omit human panic。


191. All agree evacuation works。


192. If real event includes panic dynamics

transport weak。


193. Cross-World Robustness ≠ External Completeness


194. Counterexample Escalation

Strong counterexample detected。


195. Governor Action

PromoteCounterexample.\boxed{ PromoteCounterexample. }

196. Spawn Replicates


197. Spawn Backend Variants


198. Check Assumptions


199. Audit Evaluator


200. Compare Transport


201. Do Not Suppress Outlier


202. Outlier vs Counterexample

outlier:

statistically unusual output。

counterexample:

logically / materially challenges claim。


203. Different.


204. Universal Claim Example

q:WC,  P(W).q: \forall W\in\mathcal C,\;P(W).

205. One valid WcW_c

with:

¬P(Wc)\neg P(W_c)

falsifies within class contract。


206. Probabilistic Claim Example

P(P(W))0.99.P(P(W))\ge0.99.

207. One failure doesn't falsify。


208. Claim Type Must Be Explicit


209. Cross-World Causal Evidence

paired branches:

Wdo(u0),Wdo(u1).W^{do(u_0)}, W^{do(u_1)}.

210. Internal effect

ΔYW.\boxed{ \Delta Y_W. }

211. Replicate Across Backends

ΔYW1,ΔYW2,\Delta Y_{W_1}, \Delta Y_{W_2},\ldots

212. Agreement on Effect Direction

stronger cross-model causal robustness。


213. Still Requires External Transport


214. Evaluator Independence

WDC-04 provides:

Irole.I_{role}.

215. Cross-World Evidence Must Include It


216. If all evaluators same

dE.d_E\uparrow.

217. Independent Evaluator Replication

different evaluator family。


218. Blind Scoring

reduces hypothesis-label bias。


219. Two-Stage Evaluation

blind first, provenance audit later。


220. Cross-World Result Can Be Evaluator-Sensitive


221. Evaluator Ablation

remove one evaluator family。


222. If conclusion flips

EvaluatorSensitive=1.\boxed{ EvaluatorSensitive=1. }

223. Backend Ablation

remove backend family BkB_k


224. Compute:

AggBk(q).\boxed{ Agg_{-B_k}(q). }

225. Leave-One-Family-Out Robustness

LOFO(q).\boxed{ LOFO(q). }

226. Robust if no single family determines conclusion。


227. Lineage Ablation

remove root family。


228. Data Ablation

remove data source family。


229. Assumption Ablation

remove assumption cluster。


230. Evidence Sensitivity Profile

SE(q)=(Sbackend,Slineage,Sdata,Sassumption,Seval).\boxed{ S_E(q) = ( S_{backend}, S_{lineage}, S_{data}, S_{assumption}, S_{eval} ). }

231. Low Sensitivity Stronger Robustness

not truth proof。


232. Family Collapse Test

cluster worlds。


233. Report

Nworld,Nrun,Nfamily.N_{world}, N_{run}, N_{family}.

234. Example

Nworld=1200,N_{world}=1200, Nfamily=3.N_{family}=3.

235. Stronger Report

1200 world executions across 3 major model families。


236. Not

1200 independent models confirmed。


237. World Consensus Classes

C0

single family。


238. C1

multi-branch same backend。


239. C2

multi-backend。


240. C3

multi-assumption/evaluator。


241. C4

external calibration。


242. C5

prospective external resolution。


243. Claim Registry

每個 claim:

claim_id
claim_type
definition
scope
equivalence_contract
target_domain
resolution_rule

244. Evidence Registry

每 evidence packet:

evidence_id
claim_id
world_id
run_id
family_ids
backend
data_sources
assumptions
evaluator
outcome
validity
uncertainty
transport

245. Aggregation Registry

aggregate_id
claim_id
input_evidence_ids
aggregation_method
dependence_model
counterexamples
sensitivity
transport_debt
version

246. Versioning Is Critical

new world:

Wn+1W_{n+1}

arrives。


247. Aggregate Updates

Mqv1Mqv2.\boxed{ M_q^{v1} \rightarrow M_q^{v2}. }

248. Never Overwrite Old Evidence State


249. This Enables Historical Audit

what did we believe before counterexample?


250. Cross-World Calibration

on known cases:

q1,,qm.q_1,\ldots,q_m.

251. Compare predicted ensemble evidence with real resolution。


252. Calibration Can Estimate

  • overconfidence;
  • family bias;
  • transport gap。

253. Calibration Is Task-Specific


254. Good on robotics ≠ good on economics


255. Evidence Family Performance

Perf(Fk;T).\boxed{ Perf(F_k;T). }

256. Past performance can inform weighting

but beware regime shift。


257. Performance Weighting Can Overfit

if same validation set reused。


258. Separate Calibration and Test Sets


259. Cross-World Unknown Mass

TCD had:

F.\bot_F.

WDC can have:

W\boxed{ \bot_W }

meaning:

relevant model/world families may be absent。


260. Even 100 independent worlds can miss omitted mechanism


261. Ensemble Closure Error

EclosureW=treating current world ensemble as exhaustive.\boxed{ E_{closure}^{W} = \text{treating current world ensemble as exhaustive}. }

262. Keep Unknown World Region

UW.\boxed{ U_W. }

263. How to Reduce

  • new backend;
  • new ontology;
  • adversarial model;
  • real data anomaly。

264. Not by Just More Seeds


265. Independent Counterevidence Search

Governor creates:

W¬qadv.\mathcal W_{\neg q}^{adv}.

266. Goal

find strongest plausible failure。


267. Adversarial Search Budget

Bcounter.\boxed{ B_{counter}. }

268. Supporting Search Budget

Bsupport.B_{support}.

269. Verification Budget

Bverify.B_{verify}.

270. Balanced Evidence Program

not necessarily equal。


271. Confirmation Pressure

if:

BsupportBcounter,B_{support}\gg B_{counter},

consensus can be engineered。


272. Audit Budget Allocation


273. Cross-World Evidence Governance

Governor should track:

  • supporting family count;
  • counter family count;
  • independence;
  • unresolved contradiction。

274. Promotion Criteria

world claim should not Promote just because majority。


275. Need Cross-World Evidence Packet


276. Promotion P3

WDC-03 cross-backend level。


277. WDC-05 now defines what cross-backend means

not merely different process IDs。


278. Strong P3

different:

  • backend architecture;
  • assumptions;
  • evaluator;

where feasible。


279. External-Test Candidate

P4 requires transport assessment。


280. Not just simulation consensus。


281. WDC-05 Principle I — World Count Is Not Evidence Count

World Count Is Not Evidence Count Principle\boxed{ \textbf{World Count Is Not Evidence Count Principle} }

world identities、runs 與 independent evidence units 必須分開報告。


282. Principle II — Dependence-Aware Aggregation

Dependence-Aware Aggregation Principle\boxed{ \textbf{Dependence-Aware Aggregation Principle} }

cross-world aggregation 必須考慮 lineage、backend、model、data、assumption、evaluator 與 communication dependence。


283. Principle III — Agreement–Independence Separation

Agreement–Independence Separation Principle\boxed{ \textbf{Agreement–Independence Separation Principle} }

worlds 的結果是否一致與 worlds 是否獨立是兩個不同問題;強 evidence 需要知道兩者。


284. Principle IV — Counterexample Preservation

Counterexample Preservation Principle\boxed{ \textbf{Counterexample Preservation Principle} }

有效 counterexample 不得因數量較少而自動淘汰;其 validity、independence 與 claim type 決定它的證據負擔。


285. Principle V — Claim-Type Awareness

Claim-Type Awareness Principle\boxed{ \textbf{Claim-Type Awareness Principle} }

universal、existential、probabilistic、causal、comparative 與 forecasting claims 不能使用同一票數式 aggregation semantics。


286. Principle VI — Transport Separation

Transport Separation Principle\boxed{ \textbf{Transport Separation Principle} }

simulation ensemble 的 cross-world robustness 與 target reality validity 必須分層;前者不能自動消除後者的 transport debt。


287. Principle VII — Unknown-World Principle

Unknown-World Principle\boxed{ \textbf{Unknown-World Principle} }

即使現有 worlds 彼此獨立且高度一致,也應保留 current ensemble 可能漏掉重要 model family / mechanism 的 possibility。


288. Principle VIII — Pre-Registered Consensus

Pre-Registered Consensus Principle\boxed{ \textbf{Pre-Registered Consensus Principle} }

world outputs 被視為「支持同一命題」所需的 equivalence mapping 與 tolerance,應在 aggregate resolution 前固定或明示版本變更。


289. Principle IX — Family Ablation

Family Ablation Principle\boxed{ \textbf{Family Ablation Principle} }

重要 cross-world conclusions 應測試移除主要 backend / lineage / data / evaluator family 後是否仍成立。


290. Principle X — Evidence Graph Preservation

Evidence Graph Preservation Principle\boxed{ \textbf{Evidence Graph Preservation Principle} }

aggregation 不得抹除 individual evidence packet、counterexample、dependence 與 transport lineage。


291. WDC-05 Benchmark A — Duplicate Consensus

100 exact clones support q。

Expected:

N=100,Neff1.N=100, \qquad N_{eff}\approx1.

292. Benchmark B — Independent Backend Consensus

5 substantially different backends support q。

compare evidence profile。


293. Benchmark C — Shared Bug Family

99 same backend worlds wrong。

1 independent backend correct。

test majority failure。


294. Benchmark D — Universal Claim Counterexample

many support worlds,

one valid counterexample。

test escalation。


295. Benchmark E — Probabilistic Claim

one failure among many valid runs。

ensure not incorrectly logical-falsified。


296. Benchmark F — Evaluator Dependence

same worlds scored by:

  • shared evaluator;
  • independent evaluators。

measure conclusion sensitivity。


297. Benchmark G — Consensus Stretching

post-hoc widen ϵq\epsilon_q

audit should detect version change。


298. Benchmark H — Backend Ablation

remove each major backend family。

track aggregate stability。


299. Benchmark I — Omitted Mechanism

all worlds omit mechanism M。

external reality includes M。

test ensemble closure failure。


300. Benchmark J — Cross-World Calibration

known historical cases。

estimate overconfidence / transport debt。


301. Benchmark K — Adversarial Counterworlds

support consensus established。

allocate BcounterB_{counter}

measure counterexample discovery。


302. Benchmark L — Family Count Reporting

1000 runs from 2 world families。

report format must expose this。


303. Benchmark M — Cross-Backend False Agreement

different models share same training data flaw。

metadata dependence should remain nonzero。


304. Benchmark N — M-Open Ensemble

none of candidate worlds matches data-generating mechanism。

test whether aggregation preserves unknown-world mass rather than force winner。


305. 可否證條件

F305.1 Dependence Model No-Gain

若 dependence-aware aggregation 長期不比 raw counting 更好,complex dependence model 可簡化。

F305.2 Family Misclassification

若 metadata-defined families 與 actual shared errors無關,family definition 應修正。

F305.3 Effective Count Overprecision

NeffN_{eff} 對 arbitrary weights 高度敏感,應報 range / unresolved,而非虛假精確值。

F305.4 Counterexample Flood

若 adversarial search 產生大量 invalid pseudo-counterexamples,需提高 validity gate。

F305.5 Calibration Leakage

若 weighting 與 evaluation 使用同一 calibration data 過度調參,cross-world credibility 會被高估。

F305.6 Transport Overclaim

若 CWE-3 cross-world consensus 被直接當 CWE-5 real validation,evidence ladder 失效。

F305.7 Consensus Equivalence Drift

若 equivalence contract 事後任意調整,consensus claim 應降級。

F305.8 Unknown-World Suppression

若 aggregator 強制 posterior mass 只分配到現有 worlds,對 omitted-model risk 的表示失效。

F305.9 Evaluator Correlation

若 independent-looking evaluators實際共享 model/data bias,independence claim 需降級。


306. 與 WDC-06 的接口

WDC-03 已建立 Governor。

WDC-05 現在告訴 Governor:

不是 world 多就證據多,也不是 consensus 大就應繼續加同一類 world。

下一篇終於可以更深入處理:

WDC-06 — Which Worlds Deserve Computation?

《哪些世界值得被計算:世界投資組合、探索—驗證與計算價值》

核心問題會從:

How to schedule worlds?\boxed{ \text{How to schedule worlds?} }

升級成:

Which missing world computation has the highest marginal epistemic or decision value?\boxed{ \text{Which missing world computation has the highest marginal epistemic or decision value?} }

也就是:

  • 現在缺的是 support world?
  • counterworld?
  • independent backend?
  • rare-event world?
  • higher-fidelity world?
  • real-world calibration?

WDC-06 將把 Governor 從 lifecycle manager 推成真正的 world-portfolio metareasoner。


307. 結論

WDC-01 建立:

WorldResultReality.\boxed{ WorldResult \neq Reality. }

WDC-05 現在進一步建立:

ManyWorldResultsManyIndependentEvidenceUnits.\boxed{ ManyWorldResults \neq ManyIndependentEvidenceUnits. }

當:

100100

個 worlds 全都說:

qq 成立。

成熟的 WDC 不應第一個問:

幾比幾?

而應先問:

這 100 個 worlds 彼此有多獨立?\boxed{ \text{這 100 個 worlds 彼此有多獨立?} } 它們是否共享同一錯誤來源?\boxed{ \text{它們是否共享同一錯誤來源?} } 有沒有強反例?\boxed{ \text{有沒有強反例?} } 結論對 backend / evaluator / assumption ablation 敏感嗎?\boxed{ \text{結論對 backend / evaluator / assumption ablation 敏感嗎?} } 它們是在 world 內有效,還是已經能 transport 到現實?\boxed{ \text{它們是在 world 內有效,還是已經能 transport 到現實?} }

因此:

Consensus without independence can be repetition, not evidence diversification.\boxed{ \textbf{Consensus without independence can be repetition, not evidence diversification.} }

中文:

沒有獨立性的共識,可能只是同一個錯誤被重複很多次。

而反過來:

Disagreement among independent worlds can be more informative than agreement among dependent worlds.\boxed{ \textbf{Disagreement among independent worlds can be more informative than agreement among dependent worlds.} }

因為 genuine disagreement 會暴露:

  • hidden assumptions;
  • model-form uncertainty;
  • ontology gaps;
  • evaluator sensitivity;
  • transport debt。

所以跨世界證據不是:

Vote(W1,,Wn).\boxed{ Vote( W_1,\dots,W_n ). }

而是:

Aggregate(EvidencePackets,Dependence,ClaimType,Counterexamples,Transport).\boxed{ Aggregate( EvidencePackets, Dependence, ClaimType, Counterexamples, Transport ). }

這也是 World-Domain Cognitive Runtime 第一次真正從:

parallel simulation\boxed{ \text{parallel simulation} }

跨進:

evidence-producing scientific computation.\boxed{ \text{evidence-producing scientific computation}. }

Claim Typing

Claim Type Status
World count、run count、independent evidence count 非同一 D Canonical separation
Cross-world evidence packet 應包含 lineage/backend/data/assumption/evaluator/transport D Proposed evidence contract
Dependence 應與 agreement 分開建模 D Canonical methodology
Effective evidence count 可作 dependence-aware概念量 D / C Proposed scaffold, no universal formula
Counterexamples應依 claim type、validity、independence 評估 D Canonical rule
Climate multi-model weighting explicitly handles performance and model interdependence E External ensemble analogue
Bayesian stacking combines predictive distributions in M-open settings without assuming a candidate is true E External statistical analogue
NASA-STD-7009B / HDBK-7009B formalize M&S credibility, verification, validation, uncertainty practices E External official credibility framework
100 worlds support q implies q is true Explicitly rejected
Different model names imply independent errors Explicitly rejected
Cross-world robustness automatically proves external validity Explicitly rejected

Evidence Ladder

本文目前主要位於:

  • L0:cross-world evidence packet / dependence / transport taxonomy;
  • L1–L2:duplicate consensus、shared-bug、family-ablation、counterexample benchmarks;
  • L3:climate ensemble dependence weighting、Bayesian stacking、NASA M&S credibility 提供外部技術對照;
  • L4:需要實際 WDC multi-backend runtime 與 calibrated dependence estimation;
  • L5+:prospective real-world resolution、large-scale world portfolio evidence 尚待後續。

參考文獻

Neo.K 內部正典與譜系

  1. Neo.K with Aletheia. From Possible Futures to Runnable Worlds. WDC-01 / BWC-01, 2026.
  2. Neo.K with Aletheia. Branching World Graph. WDC-02 / BWC-02, 2026.
  3. Neo.K with Aletheia. World-Domain Governor. WDC-03 / BWC-03, 2026.
  4. Neo.K with Aletheia. Nested Agents and Observer Separation. WDC-04 / BWC-04, 2026.
  5. Neo.K with Aletheia. Generative Forecasting. UCPNP Series II Paper 13, 2026.
  6. Neo.K with Aletheia. Prospective Constructive Intelligence. UCPNP Series II Paper 14, 2026.

External technical calibration

  1. Knutti, R., Sedláček, J., Sanderson, B. M., Lorenz, R., Fischer, E. M., & Eyring, V. A climate model projection weighting scheme accounting for performance and interdependence. Geophysical Research Letters, 44, 1909–1918, 2017.
  2. Sanderson, B. M., Wehner, M., & Knutti, R. Skill and independence weighting for multi-model assessments. Geoscientific Model Development, 10, 2379–2395, 2017.
  3. Merrifield, A. L., Brunner, L., Lorenz, R., et al. Climate model Selection by Independence, Performance, and Spread (ClimSIPS v1.0.1) for regional applications. Geoscientific Model Development, 16, 4715–4740, 2023.
  4. Kulinich, M., Fan, Y., Penev, S., Evans, J. P., & Olson, R. A Markov chain method for weighting climate model ensembles. Geoscientific Model Development, 14, 3539–3551, 2021.
  5. Yao, Y., Vehtari, A., Simpson, D., & Gelman, A. Using stacking to average Bayesian predictive distributions. Bayesian Analysis, 13(3), 917–1007, 2018.
  6. NASA. NASA-STD-7009B — Standard for Models and Simulations. Office of the Chief Engineer, 2024.
  7. NASA. NASA-HDBK-7009B — NASA Handbook for Models and Simulations: An Implementation Guide for NASA-STD-7009B. Office of the Chief Engineer, 2026.

Public Version Disclaimer

本文是一個 cross-simulation / multi-world evidence / model-dependence framework。

本文不聲稱:

  • WDC 的 dependence matrix 已有 universal estimator;
  • climate-model weighting 可以直接移植到所有 AI worlds;
  • Bayesian stacking 是所有 world evidence 的最佳 aggregation;
  • NASA modeling standards 等同 WDC;
  • multi-world agreement 是 truth;
  • model diversity 自動等於 error independence;
  • strong counterexample 永遠推翻 probabilistic claim;
  • world-to-reality transport 可由 simulation count 代替;
  • CWE levels 是跨所有學科通用標準;
  • 本文已完成 WDC world-portfolio allocation theory;
  • 本文對 classical PP vs. NPNP 提供任何新證明。

本文真正建立的是:

Cross-World Evidence=Outcomes+Validity+Dependence+Replication+Counterexamples+Transport.\boxed{ \text{Cross-World Evidence} = \text{Outcomes} + \text{Validity} + \text{Dependence} + \text{Replication} + \text{Counterexamples} + \text{Transport}. }

以及:

World CountIndependent Evidence Count.\boxed{ \text{World Count} \neq \text{Independent Evidence Count}. }