← Archive
lm-004652 · 2026-10

SRBC Theory Validation Program — Round 12:Conflict Resolution _ Clarification Optimization

下載 MD 檔 ⬇
📎 附件 · Companion files — 隨文交付的程式 / 證明 / 資料,可獨立下載重驗
title
SRBC Theory Validation Program — Round 12:Conflict Resolution / Clarification Optimization
author
Neo.K with Aletheia (GPT-5.6 Sol)
institution
EveMissLab / 一言諾科技有限公司
version
v0.1
date
2026-09-23
language
zh-TW
document_type
Clarification Planning / Conflict Resolution Validation Report
status
Round 12 completed

SRBC Theory Validation Program — Round 12

Conflict Resolution / Clarification Optimization

作者: Neo.K with Aletheia (GPT-5.6 Sol)
機構: EveMissLab / 一言諾科技有限公司
日期: 2026-09-23
版本: v0.1


Canonical Source Note

本報告以 UTF-8 Markdown 保存;數學 source 僅使用 $...$ 與 $$...$$。

輸入是 Round 11 的 8000 個 three-way merge / conflict cases。

本輪不是再判「有沒有 conflict」,而是問:

為了當前 query,最少還需要知道什麼?\boxed{ \text{為了當前 query,最少還需要知道什麼?} }

並同時區分:

  • query resolution;
  • canonical semantic resolution。

0. 第一個修正:不能強迫所有 Clarification 都以 QRC 為終點

Round 11 dependency conflicts 中有:

764/1000\boxed{ 764/1000 }

案例的合法 resolution outcome 包含:

Undef⁡.\operatorname{Undef}.

如果 optimizer 的成功條件固定是:

QRC⁡,\operatorname{QRC},

它會傾向要求額外假設,直到硬擠出數值。

這違反 SRBC 自己的原則。

因此 Round 12 改成:

Resolution Closure.\boxed{ \text{Resolution Closure}. }

也就是:

能確定 query 的合法 resolution outcome 即可。


1. Undef Planning Tag

為了做 decision tree,可以定義:

YQ(M)=Ans⁡(a)Y_Q(M) = \operatorname{Ans}(a)

或:

YQ(M)=Undef⁡.Y_Q(M) = \operatorname{Undef}.

但:

Undef⁡\boxed{ \operatorname{Undef} }

只是 resolution-planning label。

它不是把 undefined 偷塞回:

AQ.\mathcal A_Q.

因此:

undefined≠answer.\boxed{ \text{undefined} \neq \text{answer}. }

仍然成立。


2. 8000 Cases

各類成本:

class n expected query-resolution cost expected semantic-resolution cost minimax query cost resolve-all cost
auto_disjoint 1000 0.000 0.000 0.000 0.000
same_intent 1000 0.000 0.000 0.000 0.000
query_equivalent_conflict 1000 0.000 1.000 0.000 1.000
query_sensitive_conflict 1000 1.000 1.000 1.000 1.000
dependency_interaction 1000 1.157 1.667 1.236 2.000
evidence_conflict 1000 0.000 1.000 0.000 1.000
nuisance_conflict 1000 0.000 1.000 0.000 1.000
root_conflict 1000 0.000 1.000 0.000 1.000

最明顯的是:

Query-equivalent semantic conflict

κres,Q=0\kappa_{\mathrm{res},Q}=0

但:

κsem=1.\kappa_{\mathrm{sem}}=1.

Evidence / metadata / root conflict

同樣:

κres,Q=0.\kappa_{\mathrm{res},Q}=0.

query 已經閉合,不必為了回答當前問題追問。


3. Query-Sensitive Conflict

兩 branch:

Q(BA)≠Q(BB).Q(B_A)\neq Q(B_B).

只要問一個 distinguishing semantic question:

κres,Q=1.\boxed{ \kappa_{\mathrm{res},Q}=1. }

semantic identity 也同樣需要一次:

κsem=1.\kappa_{\mathrm{sem}}=1.

4. Dependency Interaction

candidate worlds:

A,B,AB.A,\quad B,\quad AB.

兩個 atomic questions:

  • A change 是否 intended?
  • B change 是否 intended?

Round 11 分兩種 pattern。

QDU Pattern

764 cases:

YQ(A)=Undef⁡,Y_Q(A)=\operatorname{Undef}, YQ(B)=a,Y_Q(B)=a, YQ(AB)=Undef⁡.Y_Q(AB)=\operatorname{Undef}.

只要問:

A / parent change 是否 intended?

即可直接知道:

  • yes → undefined;
  • no → branch B answer。

因此:

κres,Q=1.\boxed{ \kappa_{\mathrm{res},Q}=1. }

但如果目標是找唯一 semantic world,平均仍需:

53.\frac53.

Three-Unique Pattern

236 cases:

三個 worlds 都有不同 defined outcome。

因此 uniform prior / unit-cost 下:

E[κres,Q]=53.\boxed{ \mathbb E[ \kappa_{\mathrm{res},Q} ] = \frac53. }

5. Dependency Aggregate

1000 dependency cases:

expected query-resolution cost:

1.157\boxed{ 1.157 }

minimax worst-case average:

1.236.\boxed{ 1.236. }

semantic resolution:

1.667\boxed{ 1.667 }

expected,worst-case:

2.2.

所以 query-directed clarification 確實能利用:

A 與 AB 對當前 query 等價\text{A 與 AB 對當前 query 等價}

而不必硬分開它們。


6. Conflict Set 的總成本

只看真正 unresolved / preserved-conflict 六類,共:

6000\boxed{ 6000 }

cases。

問完所有 conflict dimensions

平均:

1.167\boxed{ 1.167 }

要唯一 semantic branch

adaptive expected cost:

1.111\boxed{ 1.111 }

只為當前 query 取得 resolution closure

adaptive expected cost:

0.360\boxed{ 0.360 }

相對 resolve-all 減少:

69.18%\boxed{ 69.18\% }

相對 semantic-resolution 減少:

67.64%.\boxed{ 67.64\%. }

7. Branch EIG 會讓 Query-Neutral Conflict 看起來值得問

本輪有:

4000\boxed{ 4000 }

個 query-neutral preserved conflicts:

  • query-equivalent semantic;
  • evidence;
  • metadata;
  • root identity。

branch identity 尚未確定,因此:

H(B)>0.H(B)>0.

distinguishing question 也有:

EIG⁡branch>0.\operatorname{EIG}_{\mathrm{branch}}>0.

但當前 query:

H(YQ)=0,H(Y_Q)=0,

因此:

EIG⁡Q=0.\boxed{ \operatorname{EIG}_Q=0. }

所以:

branch entropy reduction≠query utility.\boxed{ \text{branch entropy reduction} \neq \text{query utility}. }

8. Dependency QDU Pattern 的 EIG

uniform three-world prior:

μ(A)=μ(B)=μ(AB)=13.\mu(A)=\mu(B)=\mu(AB)=\frac13.

兩個 questions 的 branch EIG 都約:

0.9183 bits.0.9183\text{ bits}.

所以 branch-EIG 無法區分誰比較適合當前 query。

但 resolution EIG:

對 parent question:

0.9183\boxed{ 0.9183 }

bits。

對 child question只有:

0.2516\boxed{ 0.2516 }

bits。

因此 query-relative information gain 會明確優先問 parent。


9. Uniform Prior / Unit Cost

在本 Round 11 benchmark 的 candidate questions 下,真正需要 clarification 的:

2000\boxed{ 2000 }

cases 中,

greedy:

EIG⁡Qc\frac{\operatorname{EIG}_Q}{c}

和 exact dynamic programming:

2000/2000\boxed{ 2000/2000 }

相同。

但這只是這個特殊小型結構。


10. EIG / Cost 不是 Exact Optimizer

本輪進一步隨機化:

  • branch prior;
  • question cost。

prior 用 synthetic random distributions;

question cost 在約:

[0.1,10][0.1,10]

上 log-uniform。

這些都是 benchmark input,不是自然 prior,也不是人類實測成本。

結果:

pattern trials greedy suboptimal avg regret when wrong max regret
QDU_PATTERN 100,000 8.20% 14.53% 131.10%
THREE_UNIQUE 100,000 11.68% 9.24% 80.39%

所以:

arg⁡max⁡EIG⁡c\boxed{ \arg\max \frac{\operatorname{EIG}}{c} }

不能被定義成:

κres,Q.\kappa_{\mathrm{res},Q}.

它只是 heuristic。


11. Exact DP

小 branch family 應直接解:

V(S)=min⁡λ[c(λ)+∑rP(r)V(Sλ=r)].V(S) = \min_{\lambda} \left[ c(\lambda) + \sum_r P(r) V(S_{\lambda=r}) \right].

stop condition:

對 query resolution:

∣{YQ(M):M∈S}∣=1.|\{Y_Q(M):M\in S\}|=1.

對 semantic resolution:

∣S∣=1.|S|=1.

這才是真正的 minimum expected clarification cost。


12. 無 Prior 時

如果沒有:

μ,\mu,

不應偷偷假設 uniform。

本輪定義:

Vmax⁡(S)=min⁡λ[c(λ)+max⁡rVmax⁡(Sλ=r)].\boxed{ V_{\max}(S) = \min_\lambda \left[ c(\lambda) + \max_r V_{\max}(S_{\lambda=r}) \right]. }

這給 minimax clarification cost。

dependency family 的平均 minimax query cost:

1.236.\boxed{ 1.236. }

13. Semantic Conflict 不一定要問

Round 11 的:

QUERY_EQUIVALENT_CONFLICT
EVIDENCE_CONFLICT
METADATA_CONFLICT
ROOT_IDENTITY_CONFLICT

對目前 query 全部:

κres,Q=0.\boxed{ \kappa_{\mathrm{res},Q}=0. }

runtime 可以:

  1. 保留 conflict artifact;
  2. 直接回答 query;
  3. future query 需要時再 reopen。

這正是 delayed commitment 在 merge runtime 的版本。


14. Canonical Merge 仍然可以需要 Clarification

如果使用者的真正 task 是:

我要一份唯一 canonical source / evidence / semantic state。

那目標不是 query resolution,而是:

κsem.\kappa_{\mathrm{sem}}.

所以同一 conflict:

κres,Q≠κsem.\boxed{ \kappa_{\mathrm{res},Q} \neq \kappa_{\mathrm{sem}}. }

這兩個 cost 不能混報。


15. 本輪對 SRBC-05 的修正

SRBC-05 原來的 clarification priority:

Pclarify=SQcask+ϵP_{\mathrm{clarify}} = \frac{S_Q}{c_{\mathrm{ask}}+\epsilon}

仍可做 heuristic。

EIG 版本:

EIG⁡Qc\frac{\operatorname{EIG}_Q}{c}

也可做 heuristic。

但 exact κ\kappa 應由:

  • DP;
  • search;
  • optimization;
  • formal branch elimination;

計算。

不能用局部 score 直接冒充 global minimum。


16. Prior Laundering 再次被阻止

如果沒有 branch prior:

μ,\mu,

就不能因方便寫:

μi=1n\mu_i=\frac1n

然後把 EIG ranking 說成「最理性」。

Round 12 所有 uniform-prior 數字都明確標:

synthetic benchmark prior.\boxed{ \text{synthetic benchmark prior}. }

17. Clarification 本質是 Partition Operator

一個問題不是只有字串:

Which rule did you mean?

正式 artifact 是:

λ:I→Rλ.\lambda: \mathfrak I \rightarrow \mathcal R_\lambda.

每個 response:

rr

對應:

Iλ=r⊆I.\mathfrak I_{\lambda=r} \subseteq \mathfrak I.

所以 clarification planner 可以直接跑:

  • expected elimination;
  • resolution EIG;
  • minimax cost;
  • exact DP。

18. 本輪沒有證明什麼

沒有證明:

  1. 人類 clarification cost 真的是 1;
  2. 人類 branch prior 是 uniform;
  3. EIG 在自然語言對話的心理成本準確;
  4. synthetic atomic question 能直接變成最佳人類問句;
  5. 大型 branch space 的 exact DP 可實務擴展。

所以這輪是:

clarification planning semantics + algorithmic benchmark.\boxed{ \text{clarification planning semantics + algorithmic benchmark}. }

19. 下一輪

Round 13 最自然就是:

Adaptive Clarification Dialogue / Noisy Human Answers.\boxed{ \text{Adaptive Clarification Dialogue / Noisy Human Answers}. }

因為現在的 clarification response 都是假設 deterministic / truthful。

下一輪可以主動加入:

  • 使用者答錯;
  • 模糊回答;
  • 部分回答;
  • 改口;
  • refusal;
  • confidence;
  • contradictory clarification;
  • question fatigue cost;
  • repeated-question consistency。

然後測:

Bayesian / non-Bayesian belief update\text{Bayesian / non-Bayesian belief update}

和:

hard commitment update\text{hard commitment update}

何時應該分開。


附錄 A:Class Costs

class n expected query cost expected semantic cost minimax query cost resolve-all
auto_disjoint 1000 0.000 0.000 0.000 0.000
same_intent 1000 0.000 0.000 0.000 0.000
query_equivalent_conflict 1000 0.000 1.000 0.000 1.000
query_sensitive_conflict 1000 1.000 1.000 1.000 1.000
dependency_interaction 1000 1.157 1.667 1.236 2.000
evidence_conflict 1000 0.000 1.000 0.000 1.000
nuisance_conflict 1000 0.000 1.000 0.000 1.000
root_conflict 1000 0.000 1.000 0.000 1.000

附錄 B:Random Prior / Cost Stress

pattern trials greedy suboptimal avg regret when wrong max regret
QDU_PATTERN 100,000 8.20% 14.53% 131.10%
THREE_UNIQUE 100,000 11.68% 9.24% 80.39%

附錄 C:Machine Results

{
  "round": "SRBC Theory Validation Round 12",
  "engine_revision": "v0.12-resolution-closure",
  "cases": 8000,
  "class_summary": {
    "auto_disjoint": {
      "n": 1000,
      "avg_query_expected_cost": 0.0,
      "avg_semantic_expected_cost": 0.0,
      "avg_query_worst_cost": 0.0,
      "avg_semantic_worst_cost": 0.0,
      "avg_resolve_all_cost": 0.0,
      "avg_greedy_query_eig_cost": 0.0
    },
    "same_intent": {
      "n": 1000,
      "avg_query_expected_cost": 0.0,
      "avg_semantic_expected_cost": 0.0,
      "avg_query_worst_cost": 0.0,
      "avg_semantic_worst_cost": 0.0,
      "avg_resolve_all_cost": 0.0,
      "avg_greedy_query_eig_cost": 0.0
    },
    "query_equivalent_conflict": {
      "n": 1000,
      "avg_query_expected_cost": 0.0,
      "avg_semantic_expected_cost": 1.0,
      "avg_query_worst_cost": 0.0,
      "avg_semantic_worst_cost": 1.0,
      "avg_resolve_all_cost": 1.0,
      "avg_greedy_query_eig_cost": 0.0
    },
    "query_sensitive_conflict": {
      "n": 1000,
      "avg_query_expected_cost": 1.0,
      "avg_semantic_expected_cost": 1.0,
      "avg_query_worst_cost": 1.0,
      "avg_semantic_worst_cost": 1.0,
      "avg_resolve_all_cost": 1.0,
      "avg_greedy_query_eig_cost": 1.0
    },
    "dependency_interaction": {
      "n": 1000,
      "avg_query_expected_cost": 1.1573333333333324,
      "avg_semantic_expected_cost": 1.6666666666666865,
      "avg_query_worst_cost": 1.236,
      "avg_semantic_worst_cost": 2.0,
      "avg_resolve_all_cost": 2.0,
      "avg_greedy_query_eig_cost": 1.1573333333333324
    },
    "evidence_conflict": {
      "n": 1000,
      "avg_query_expected_cost": 0.0,
      "avg_semantic_expected_cost": 1.0,
      "avg_query_worst_cost": 0.0,
      "avg_semantic_worst_cost": 1.0,
      "avg_resolve_all_cost": 1.0,
      "avg_greedy_query_eig_cost": 0.0
    },
    "nuisance_conflict": {
      "n": 1000,
      "avg_query_expected_cost": 0.0,
      "avg_semantic_expected_cost": 1.0,
      "avg_query_worst_cost": 0.0,
      "avg_semantic_worst_cost": 1.0,
      "avg_resolve_all_cost": 1.0,
      "avg_greedy_query_eig_cost": 0.0
    },
    "root_conflict": {
      "n": 1000,
      "avg_query_expected_cost": 0.0,
      "avg_semantic_expected_cost": 1.0,
      "avg_query_worst_cost": 0.0,
      "avg_semantic_worst_cost": 1.0,
      "avg_resolve_all_cost": 1.0,
      "avg_greedy_query_eig_cost": 0.0
    }
  },
  "conflict_set": {
    "cases": 6000,
    "avg_query_resolution_expected_cost": 0.3595555555555556,
    "avg_semantic_resolution_expected_cost": 1.1111111111111112,
    "avg_resolve_all_cost": 1.1666666666666667,
    "query_vs_resolve_all_reduction": 0.6918095238095239,
    "query_vs_semantic_resolution_reduction": 0.6764
  },
  "greedy_query_eig": {
    "active_query_clarification_cases": 2000,
    "matches_exact_dp_under_uniform_unit_cost": 2000,
    "failures": 0,
    "query_neutral_cases_where_branch_eig_positive_but_resolution_eig_zero": 4000
  },
  "dependency_interaction": {
    "cases": 1000,
    "avg_uniform_expected_resolution_cost": 1.1573333333333333,
    "avg_minimax_worst_case_cost": 1.236,
    "eig_patterns": [
      {
        "pattern": "QDU_PATTERN",
        "question_scores": [
          [
            "ask_A_change",
            0.918296,
            0.918296
          ],
          [
            "ask_B_change",
            0.251629,
            0.918296
          ]
        ],
        "count": 764
      },
      {
        "pattern": "THREE_UNIQUE",
        "question_scores": [
          [
            "ask_A_change",
            0.918296,
            0.918296
          ],
          [
            "ask_B_change",
            0.918296,
            0.918296
          ]
        ],
        "count": 236
      }
    ]
  },
  "prior_cost_stress": {
    "explicit_note": "Priors and question costs are synthetic benchmark inputs, not natural priors or measured human costs.",
    "QDU_PATTERN": {
      "n": 100000,
      "greedy_suboptimal": 8196,
      "max_relative_regret": 1.311015936140403,
      "suboptimal_rate": 0.08196,
      "avg_relative_regret_when_suboptimal": 0.14529424714734357
    },
    "THREE_UNIQUE": {
      "n": 100000,
      "greedy_suboptimal": 11685,
      "max_relative_regret": 0.8038866550313561,
      "suboptimal_rate": 0.11685,
      "avg_relative_regret_when_suboptimal": 0.09239126242904895
    }
  },
  "core_findings": {
    "qrc_is_not_only_legal_closure_target": true,
    "resolution_outcome_variable_includes_undefined_tag_for_planning_only": true,
    "undefined_tag_is_not_promoted_into_answer_domain": true,
    "semantic_merge_cost_separate_from_query_resolution_cost": true,
    "eig_requires_declared_prior": true,
    "eig_per_cost_not_globally_optimal": true,
    "without_prior_use_minimax_or_bounds": true
  }
}

附錄 D:Artifacts

  • SRBC_Theory_Validation_Round12_Report_v0.1_2026-09-23.md
  • SRBC_Core_Patch_v0.12_2026-09-23.md
  • clarification_optimizer_protocol_v012.yaml
  • round12_clarification_optimizer.py
  • round12_eig_stress.py
  • round12_replay.py
  • results_round12.json
  • clarification_cases_round12.jsonl
  • merge_cases_round11.jsonl
  • validation.json