← Archive
lm-004653 · 2026-10

SRBC Theory Validation Program — Round 13:Adaptive Clarification Dialogue _ Noisy Human Answers

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 13:Adaptive Clarification Dialogue / Noisy Human Answers
author
Neo.K with Aletheia (GPT-5.6 Sol)
institution
EveMissLab / 一言諾科技有限公司
version
v0.1
date
2026-09-24
language
zh-TW
document_type
Noisy Clarification / Evidence-Ledger Validation Report
status
Round 13 corrected r2 completed

SRBC Theory Validation Program — Round 13

Adaptive Clarification Dialogue / Noisy Human Answers

作者: Neo.K with Aletheia (GPT-5.6 Sol)
機構: EveMissLab / 一言諾科技有限公司
日期: 2026-09-24
版本: v0.1


Canonical Source Note

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

本輪使用 synthetic answer channels。所有錯誤率、refusal rate、correlation rate 都是 benchmark parameters,不是人類行為統計。

正式結果使用 common-response schedule:同一 synthetic session、同一 question、第 nn 次詢問,所有 policy 看到同一個 synthetic response,避免 policy 比較被不同 RNG draw 污染。

第一版 correction harness 曾誤共用「已被另一 policy 消耗的 remaining-question list」,導致 soft-retraction 95.28% 的中間結果;該結果已作廢。修正後 5000-case correction benchmark 已全部重跑。


0. 問題

Round 12 假設:

clarification response is truthful and deterministic.\boxed{ \text{clarification response is truthful and deterministic}. }

Round 13 把這個假設拿掉。

現在 response 可能:

  • 正確;
  • 答錯;
  • refusal;
  • ambiguous;
  • 前後矛盾;
  • explicit correction;
  • 同一來源重複偏誤。

1. Clarification Speech Act 不是同一 epistemic type

Round 13 最重要的理論結果之一:

clarification answer≠report of hidden fact\boxed{ \text{clarification answer} \neq \text{report of hidden fact} }

至少要分:

Constitutive

使用者定義自己的 intended meaning / preference。

例如:

我說的「有利」就是 posterior 比較高。

這通常可以成為 hard commitment,因為 speaker 正在設定 target specification。

Reportive

我記得主持人是隨機開門。

這是在報告世界資訊或記憶,應先進 evidence ledger。

Inferential

我覺得答案應該是 2/32/3。

這是推論 evidence。


2. Noise Benchmark

active query cases:

2000.\boxed{ 2000. }

每 scenario 每 case:

2525

repetitions。

因此每 policy / scenario:

50000\boxed{ 50000 }

sessions。

scenario policy resolution rate accuracy when resolved wrong resolution / all avg turns
iid05 hard_first 100.00% 94.54% 5.460% 1.077
iid05 confirm2 100.00% 99.21% 0.794% 2.259
iid05 soft_iid 99.98% 96.67% 3.328% 1.435
iid05 soft_corr_aware 99.98% 96.67% 3.328% 1.435
iid20 hard_first 100.00% 79.09% 20.912% 1.072
iid20 confirm2 100.00% 88.94% 11.062% 2.497
iid20 soft_iid 93.23% 97.80% 2.052% 4.421
iid20 soft_corr_aware 93.23% 97.80% 2.052% 4.421
mixed10 hard_first 100.00% 86.75% 13.250% 1.343
mixed10 confirm2 99.90% 95.53% 4.462% 2.987
mixed10 soft_iid 98.27% 97.97% 1.998% 3.362
mixed10 soft_corr_aware 98.27% 97.97% 1.998% 3.362
correlated15 hard_first 100.00% 84.03% 15.966% 1.073
correlated15 confirm2 100.00% 84.03% 15.966% 2.146
correlated15 soft_iid 100.00% 84.03% 15.966% 2.343
correlated15 soft_corr_aware 0.00% — 0.000% 9.000

3. IID 5% Error

single hard prune:

5.460%\boxed{ 5.460\% }

wrong resolution。

要求兩次一致:

0.794%.\boxed{ 0.794\%. }

所以 IID 模型下 repeated confirmation 確實有價值。

代價是平均 turns:

1.077→2.259.1.077 \rightarrow 2.259.

4. IID 20% Error

hard-first:

20.912%\boxed{ 20.912\% }

wrong resolution。

confirm-two:

11.062%.\boxed{ 11.062\%. }

soft IID confidence policy:

  • resolution rate:93.23%93.23\%
  • wrong resolution / all:2.052%\boxed{ 2.052\% }
  • resolved-case accuracy:97.80%.97.80\%.

這不是「soft 一定更好」。

它是在用更多 turns 與:

允許 unresolved\boxed{ \text{允許 unresolved} }

換取較低錯閉合。


5. Mixed Noise

synthetic channel:

  • 10% refusal;
  • 10% ambiguous;
  • 10% wrong;
  • 70% correct。

hard-first wrong resolution:

13.250%.13.250\%.

confirm-two:

4.462%.4.462\%.

soft evidence policy:

1.998%1.998\%

wrong resolution / all,

但仍保留約:

1.73%1.73\%

sessions unresolved。

因此:

abstention is part of safety policy.\boxed{ \text{abstention is part of safety policy}. }

6. Correlated 15% Session Bias

這是本輪最重要的反例。

同一 question 在一個 session 中要嘛一直正確,要嘛一直反向。

hard-first:

15.966%15.966\%

wrong。

confirm-two:

15.966%15.966\%

wrong。

IID Bayes:

15.966%15.966\%

wrong。

也就是:

再問一次完全沒有改善.\boxed{ \text{再問一次完全沒有改善}. }

因為不是兩個獨立 errors。


7. Correlation-Aware Policy

correlation-aware model 知道:

同一 source 對同一 question 的 repetition 共享 latent bias。

因此重問九次也不把它當九份證據。

在:

τ=0.95\tau=0.95

confidence threshold 下:

  • false high-confidence:0\boxed{ 0 }
  • resolution:0.\boxed{ 0. }

它選擇「不知道」。

這是刻意保守,不是免費午餐。


8. Exact Repetition Control

100,000 個二世界 correlated sessions。

confirm-two wrong commit:

15.286%.\boxed{ 15.286\%. }

IID Bayes 把三個相同回答當 independent:

P(reported branch∣3 same)≈99.45%.P(\text{reported branch}\mid 3\text{ same}) \approx 99.45\%.

但因真實錯誤是 session-correlated,false high-confidence rate 仍:

15.286%.\boxed{ 15.286\%. }

correlation-aware:

  • false high-confidence:00
  • unresolved:100%.100\%.

9. Correlated Repetition Lemma

perfect shared-bias model 下:

P(R1:n=r∣T=r)=1−ϵ,P(R_{1:n}=r\mid T=r)=1-\epsilon, P(R1:n=r∣T≠r)=ϵ.P(R_{1:n}=r\mid T\neq r)=\epsilon.

所以 likelihood ratio 只是:

1−ϵϵ.\boxed{ \frac{1-\epsilon}{\epsilon}. }

IID 誤模型卻用:

(1−ϵϵ)n.\boxed{ \left( \frac{1-\epsilon}{\epsilon} \right)^n. }

這就是虛假 certainty 的來源。


10. Correction Benchmark

第一個 response 強制答錯,

接著 explicit correction。

共:

50005000

cases。

policy n true branch retained final outcome accuracy
hard irreversible 5000 0.00% —
reversible ledger 5000 100.00% 100.00%
soft retraction 5000 100.00% 100.00%

hard irreversible:

0%\boxed{ 0\% }

true branch recovery。

reversible ledger:

100%.\boxed{ 100\%. }

soft evidence retraction + truthful follow-up:

100%.\boxed{ 100\%. }

11. Irreversible Pruning Lemma

若:

St+1=St∩Ct,S_{t+1}=S_t\cap C_t,

且一次錯誤 response:

M∗∉Ck,M^*\notin C_k,

則:

M∗∉Sk+1.M^*\notin S_{k+1}.

而:

St+1⊆St.S_{t+1}\subseteq S_t.

所以:

∀t>k:M∗∉St.\boxed{ \forall t>k: M^*\notin S_t. }

後面再說:

抱歉,我剛才說錯了。

如果 runtime 沒有 retract / reopen,

數學上就恢復不了真 branch。


12. Contradiction Benchmark

5000 cases:

先正確回答,再給相反但沒有 correction marker的回答。

Hard append

100%\boxed{ 100\% }

得到 empty / inconsistent active family。

Last-write-wins

true branch retention:

0%.\boxed{ 0\%. }

也就是直接把前一次 evidence 靜默反轉。

Evidence ledger

  • conflict detection:100%\boxed{ 100\% }
  • true branch retained:100%.\boxed{ 100\%. }

Soft IID

correct + opposite evidence 抵消後:

100%\boxed{ 100\% }

避免達到 95% confidence closure。


13. Correction 與 Contradiction 不能共用同一 operator

明示:

supersedes: event-17

是 correction。

沒有 supersession relation 的相反 response 是:

EvidenceConflict

不能自動 last-write-wins。


14. Query-Neutral Conflict

Round 12 已經知道 query-neutral conflict 不需 clarification。

Round 13 再驗:

4000/4000\boxed{ 4000/4000 }

query-neutral preserved conflicts 問題數為零。

所以 noisy-dialogue support 沒有讓系統退化成:

有 conflict 就一直問。


15. Probabilistic Confidence 不是 QRC

soft policy 的:

P(YQ=y∣E)≥0.95P(Y_Q=y\mid\mathcal E)\ge0.95

最多叫:

PCR⁡0.95.\boxed{ \operatorname{PCR}_{0.95}. }

不能寫:

QRC⁡.\operatorname{QRC}.

因為 admissible semantic branches 並沒有被形式證明成同一答案。


16. Hard Promotion 的正確邊界

本輪 synthetic noise 結果主要針對:

reportive clarification.\boxed{ \text{reportive clarification}. }

不能錯誤解讀成:

使用者自己定義自己的 intended meaning 也不能直接相信。

如果是 constitutive speech act:

我這裡就是定義成 X。

那 X 本身可成為新 source commitment。

問題不是「人會不會答錯」,

而是:

這句話是在設定世界,還是在報告世界?\boxed{ \text{這句話是在設定世界,還是在報告世界?} }

17. Authority Scope

即使是 constitutive,也有 scope。

使用者可以定義:

  • 自己的術語;
  • 自己的 preference;
  • 自己要問的 query;
  • 自己建立的 fictional / formal world。

但不能靠一句:

我定義黎曼猜想已證明。

就 constitutively 改變數學真理。

所以:

Authority⁡(speaker,domain)\boxed{ \operatorname{Authority} ( \text{speaker},\text{domain} ) }

必須進 act typing。


18. Round 13 對 Runtime 的真正修改

clarification runtime 從:

question
→ answer
→ commitment

改成:

question
→ response event
→ speech-act typing
→ authority check
→ evidence / commitment routing
→ correlation model
→ optional belief update
→ promotion / retract / reopen

19. Evidence Ledger

建議 event:

event_id:
question_id:
speaker_id:
act_type:
response:
authority_scope:
channel_model:
correlation_group:
timestamp:
supersedes:
retracts:

所以 future correction 不需要重建整個對話語意。


20. 本輪沒有證明什麼

所有 noise rates 都是 synthetic。

沒有證明:

  1. 人類 clarification 真實 error rate 是 5%、15% 或 20%;
  2. 同一人的重複回答一定 perfect-correlated;
  3. Bayesian likelihood 是人類心理真模型;
  4. 0.95 是最佳 confidence threshold;
  5. constitutive / reportive speech act 能被 AI 100% 正確分類。

本輪證明的是:

這些 error models 一旦存在,更新規則會產生哪些結構性後果。\boxed{ \text{這些 error models 一旦存在,更新規則會產生哪些結構性後果。} }

21. 下一輪

Round 14 最自然的方向是:

Speech-Act Classification / Authority Stress Test.\boxed{ \text{Speech-Act Classification / Authority Stress Test}. }

現在真正最薄弱的環節變成:

系統怎麼知道一句 clarification 是 constitutive、reportive、inferential,還是 correction?

可以故意生成:

  • 「我是說 X」;
  • 「我記得是 X」;
  • 「假設 X」;
  • 「應該是 X」;
  • 「不,我剛才說錯,是 Y」;
  • 「在這個 fictional world 裡 X」;
  • 「現實世界就是 X」。

測錯分型會如何把 evidence 變成 hard commitment,或把 authoritative intent 錯當 probabilistic evidence。


附錄 A:Noise Benchmark

scenario policy resolution rate accuracy when resolved wrong resolution / all avg turns
iid05 hard_first 100.00% 94.54% 5.460% 1.077
iid05 confirm2 100.00% 99.21% 0.794% 2.259
iid05 soft_iid 99.98% 96.67% 3.328% 1.435
iid05 soft_corr_aware 99.98% 96.67% 3.328% 1.435
iid20 hard_first 100.00% 79.09% 20.912% 1.072
iid20 confirm2 100.00% 88.94% 11.062% 2.497
iid20 soft_iid 93.23% 97.80% 2.052% 4.421
iid20 soft_corr_aware 93.23% 97.80% 2.052% 4.421
mixed10 hard_first 100.00% 86.75% 13.250% 1.343
mixed10 confirm2 99.90% 95.53% 4.462% 2.987
mixed10 soft_iid 98.27% 97.97% 1.998% 3.362
mixed10 soft_corr_aware 98.27% 97.97% 1.998% 3.362
correlated15 hard_first 100.00% 84.03% 15.966% 1.073
correlated15 confirm2 100.00% 84.03% 15.966% 2.146
correlated15 soft_iid 100.00% 84.03% 15.966% 2.343
correlated15 soft_corr_aware 0.00% — 0.000% 9.000

附錄 B:Correction Benchmark

policy n true branch retained final outcome accuracy
hard irreversible 5000 0.00% —
reversible ledger 5000 100.00% 100.00%
soft retraction 5000 100.00% 100.00%

附錄 C:Machine Results

{
  "round": "SRBC Theory Validation Round 13",
  "engine_revision": "v0.13-r2-common-response-schedule",
  "active_cases": 2000,
  "repetitions_per_case_scenario": 25,
  "query_neutral_cases": 4000,
  "query_neutral_zero_question_rate": 1.0,
  "noise_benchmark": {
    "iid05": {
      "hard_first": {
        "n": 50000,
        "resolution_rate": 1.0,
        "accuracy_when_resolved": 0.9454,
        "wrong_resolution_rate_all": 0.0546,
        "avg_turns": 1.07654
      },
      "confirm2": {
        "n": 50000,
        "resolution_rate": 1.0,
        "accuracy_when_resolved": 0.99206,
        "wrong_resolution_rate_all": 0.00794,
        "avg_turns": 2.25926
      },
      "soft_iid": {
        "n": 50000,
        "resolution_rate": 0.99978,
        "accuracy_when_resolved": 0.9667126767888936,
        "wrong_resolution_rate_all": 0.03328,
        "avg_turns": 1.4354
      },
      "soft_corr_aware": {
        "n": 50000,
        "resolution_rate": 0.99978,
        "accuracy_when_resolved": 0.9667126767888936,
        "wrong_resolution_rate_all": 0.03328,
        "avg_turns": 1.4354
      }
    },
    "iid20": {
      "hard_first": {
        "n": 50000,
        "resolution_rate": 1.0,
        "accuracy_when_resolved": 0.79088,
        "wrong_resolution_rate_all": 0.20912,
        "avg_turns": 1.07152
      },
      "confirm2": {
        "n": 50000,
        "resolution_rate": 1.0,
        "accuracy_when_resolved": 0.88938,
        "wrong_resolution_rate_all": 0.11062,
        "avg_turns": 2.49676
      },
      "soft_iid": {
        "n": 50000,
        "resolution_rate": 0.9323,
        "accuracy_when_resolved": 0.9779899174085595,
        "wrong_resolution_rate_all": 0.02052,
        "avg_turns": 4.42124
      },
      "soft_corr_aware": {
        "n": 50000,
        "resolution_rate": 0.9323,
        "accuracy_when_resolved": 0.9779899174085595,
        "wrong_resolution_rate_all": 0.02052,
        "avg_turns": 4.42124
      }
    },
    "mixed10": {
      "hard_first": {
        "n": 50000,
        "resolution_rate": 0.99998,
        "accuracy_when_resolved": 0.8674973499469989,
        "wrong_resolution_rate_all": 0.1325,
        "avg_turns": 1.34264
      },
      "confirm2": {
        "n": 50000,
        "resolution_rate": 0.99904,
        "accuracy_when_resolved": 0.9553371236386932,
        "wrong_resolution_rate_all": 0.04462,
        "avg_turns": 2.98686
      },
      "soft_iid": {
        "n": 50000,
        "resolution_rate": 0.9827,
        "accuracy_when_resolved": 0.9796682609138089,
        "wrong_resolution_rate_all": 0.01998,
        "avg_turns": 3.36202
      },
      "soft_corr_aware": {
        "n": 50000,
        "resolution_rate": 0.9827,
        "accuracy_when_resolved": 0.9796682609138089,
        "wrong_resolution_rate_all": 0.01998,
        "avg_turns": 3.36202
      }
    },
    "correlated15": {
      "hard_first": {
        "n": 50000,
        "resolution_rate": 1.0,
        "accuracy_when_resolved": 0.84034,
        "wrong_resolution_rate_all": 0.15966,
        "avg_turns": 1.07314
      },
      "confirm2": {
        "n": 50000,
        "resolution_rate": 1.0,
        "accuracy_when_resolved": 0.84034,
        "wrong_resolution_rate_all": 0.15966,
        "avg_turns": 2.14628
      },
      "soft_iid": {
        "n": 50000,
        "resolution_rate": 1.0,
        "accuracy_when_resolved": 0.84034,
        "wrong_resolution_rate_all": 0.15966,
        "avg_turns": 2.34268
      },
      "soft_corr_aware": {
        "n": 50000,
        "resolution_rate": 0.0,
        "accuracy_when_resolved": null,
        "wrong_resolution_rate_all": 0.0,
        "avg_turns": 9.0
      }
    }
  },
  "correction_benchmark": {
    "hard_irreversible": {
      "n": 5000,
      "true_branch_retention_rate": 0.0
    },
    "reversible_ledger": {
      "n": 5000,
      "true_branch_retention_rate": 1.0,
      "final_outcome_accuracy": 1.0
    },
    "soft_retraction": {
      "n": 5000,
      "true_branch_nonzero_rate": 1.0,
      "final_MAP_outcome_accuracy": 1.0
    }
  },
  "contradiction_benchmark": {
    "n": 5000,
    "hard_append_empty_rate": 1.0,
    "last_write_wins_true_branch_retention_rate": 0.0,
    "evidence_ledger_conflict_detection_rate": 1.0,
    "evidence_ledger_true_branch_retention_rate": 1.0,
    "soft_iid_avoids_95_confidence_rate": 1.0
  },
  "correlated_repeat_control": {
    "n": 100000,
    "confirm2_wrong_commit_rate": 0.15286,
    "iid_bayes_false_high_confidence_rate": 0.15286,
    "correlated_aware_false_high_confidence_rate": 0.0,
    "correlated_aware_unresolved_rate": 1.0
  },
  "speech_act_typing": {
    "constitutive": "authorized declaration that sets intended meaning/preference; eligible for hard commitment within authority scope",
    "reportive": "report of external fact/memory/belief; evidence by default",
    "inferential": "derived judgment/answer; evidence for a claim, not source fact",
    "refusal_or_ambiguous": "no branch elimination by default"
  },
  "core_findings": {
    "hard_promotion_from_single_reportive_answer_is_noise_sensitive": true,
    "correction_requires_retraction_or_reopen": true,
    "confidence_closure_is_not_QRC": true,
    "same_source_repetition_requires_correlation_model": true,
    "speech_act_type_controls_commitment_promotion": true
  },
  "discarded_intermediate_result": {
    "bug": "The first correction harness reused a remaining-question list already consumed by the reversible-ledger branch.",
    "discarded_value": "soft_retraction final_MAP_outcome_accuracy = 0.9528",
    "fix": "Use independent follow-up state for each policy, then rerun all 5000 correction cases."
  }
}

附錄 D:Artifacts

  • SRBC_Theory_Validation_Round13_Report_v0.1_2026-09-24.md
  • SRBC_Core_Patch_v0.13_2026-09-24.md
  • adaptive_clarification_protocol_v013.yaml
  • round13_noisy_clarification.py
  • results_round13.json
  • noise_results_round13.jsonl
  • merge_cases_round11.jsonl
  • validation.json