- title
- SRBC Theory Validation Program — Round 05:Formalizing Real Cases
- 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
- Machine-Replayable Real-Case Formalization Report
- status
- Round 05 completed
SRBC Theory Validation Program — Round 05
Formalizing Real Cases:從真實文本 transfer 到可重放機器證書
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 $$...$$。Round 05 不再新增 taxonomy,而是把四類 real-text case 轉成 machine-replayable model / symbolic certificate:
- Monty Hall;
- Boy/Girl;
- Bertrand-style measure family;
- Two Envelopes。
每一項 certificate 都區分 source / completion assumptions、query、proof scope、proof type,以及 exact theorem 與 stress evidence。
0. 本輪目的
Round 04 已證明 SRBC 的 closure machinery 可以 transfer 到真實經典問題,但 underdetermination origin 需要分層。
Round 05 進一步問:
答案是:
1. Machine certificates
本輪產生:
個 machine-readable certificates。
| claim | case | proof type | status |
|---|---|---|---|
| MH-C1 | MH | symbolic_identity | PASS |
| MH-C2 | MH | symbolic_substitution | PASS |
| MH-C3 | MH | symbolic_substitution | PASS |
| MH-C4 | MH | symbolic_family | PASS |
| MH-C5 | MH | finite_enumeration | PASS |
| BG-C1 | BG | finite_enumeration | PASS |
| BG-C2 | BG | finite_observation_protocol | PASS |
| BG-C3 | BG | finite_enumeration | PASS |
| BG-C4 | BG | symbolic_likelihood_family | PASS |
| BER-C1 | BER | symbolic_change_of_parameter | PASS |
| BER-C2 | BER | symbolic_monotonicity | PASS |
| TE-C1 | TE | symbolic_bayes_expectation | PASS |
| TE-C2 | TE | exact_countermodel_pair | PASS |
| TE-C3 | TE | symbolic_symmetry_plus_random_stress | PASS |
使用兩種 backend:
- Python
Fraction:exact finite enumeration; - SymPy:symbolic identity / parameter-family algebra。
本環境沒有 Z3,因此本輪沒有假裝產生 SMT-solver proof。這不影響有限枚舉與 symbolic identity 的可重放性,但也意味著「SMT certificate」仍是未來工作。
2. Monty Hall:finite kernel + symbolic posterior
一般 observed-door posterior:
SymPy 代入 standard symmetric:
得到:
random-unchosen:
得到:
standard goat family:
得到:
而:
所以:
3. Monty Hall:ex-ante query 的有限證書
標準 goat-opening protocol 下,車只可能在:
player 先選:
always-switch:
- :輸;
- :贏;
- :贏。
等先驗:
此 finite certificate 完全沒有用:
所以再次確認:
4. Boy/Girl:observation protocol 被真正機器化
完整 ordered family:
各先驗:
4.1 Condition on family event
「至少一個 boy」的 likelihood:
因此:
4.2 Random-child reporting protocol
隨機挑一個孩子並回報該孩子為 boy。
report likelihood:
因此:
4.3 Older-child query
older child is girl:
因此:
5. Boy/Girl 的一般 observation formula
令:
且:
等先驗下:
因此:
這和 Monty Hall 的 host action likelihood 具有同一高階結構。
6. Bertrand-style cube:measure family 的 symbolic certificate
令:
為 cube side length。
不直接假設 uniform,而令:
在:
上 uniform,其中:
則:
所以:
具體:
是 uniform length;
是 uniform face area;
是 uniform volume。
而:
所以 時,這個 family 的 answer image 是:
7. Round 05 第一個理論修正:measure-open 不等於 QDU
這個 formalization 把一件事逼得很清楚。
source:
randomly produce a cube
沒有固定:
但每一個 completion:
都已經是一個完整 probability model,而且:
完全有定義。
所以:
可以造成:
但不必有:
因此:
這就是 v0.5 的核心修正。
8. Two Envelopes:general prior formula
令:
為較小 envelope amount,具有 prior:
隨機拿一封 envelope 後觀察:
只有兩個來源:
- ,目前拿到較小 envelope,另一封是 ;
- ,目前拿到較大 envelope,另一封是 。
因此:
減掉留下的:
得到:
所以 switch beneficial 當且僅當:
9. Two Envelopes:exact countermodel pair
固定:
Prior A
其餘 finite support 合法補滿。
得到:
所以 switch。
Prior B
得到:
所以 stay。
因此:
這是一個 exact finite countermodel certificate。
10. Two Envelopes:ex-ante query 反而閉合
尚未打開 envelope 前,若:
隨機拿到的 envelope 期望:
另一封 envelope 的 ex-ante 期望因交換對稱性也為:
所以:
本輪另外生成:
組 proper finite-support rational priors 做 exact Fraction implementation stress,全數通過。
這再次得到:
而:
11. Stress evidence 不得冒充 theorem
10,000 組 prior 全通過,不是 theorem 的來源。
真正 theorem 來自 envelope symmetry:
random stress 只是:
因此 v0.5 certificate schema 強制區分:
proof_type: symbolic_identity
theorem: true
和:
proof_type: randomized_stress
theorem: false
12. QDU 必須 completion-relative
v0.5 將 QDU 的型別邊界固定為:
其中:
必須已是 admissible completion。
所以:
source 沒寫 prior
不是 QDU。
正確可能是:
source_specification: open
open_slots:
- prior
qru: true
qdu: false
若未來 completion family 納入:
- infinite expectation;
- non-integrable utility;
- genuinely undefined conditional;
才會真正進 QDU。
13. Source Specification Status
新增:
因此目前完整 state 可寫:
這比最初的:
QRC | QRU | QDU
已經嚴格很多。
14. 四案例 machine status
Monty Hall
- open slots:protocol / prior;
- observed-door query:QRU across protocol completions;
- standard ex-ante query:QRC;
- proof:symbolic + finite enumeration。
Boy/Girl
- open slot:information-generation protocol;
- report query:QRU;
- protocol fixed 後:QRC;
- proof:finite enumeration + symbolic likelihood family。
Bertrand
- open slot:probability measure;
- query 在每個 completion 中有定義;
- across measures:QRU;
- QDU:false in this completion family;
- proof:symbolic parameter family。
Two Envelopes
- open slots:prior / expectation regime;
- proper finite-support prior family:observed query QRU;
- finite-expectation ex-ante query:QRC;
- proof:symbolic identity + exact countermodel pair。
15. 本輪對 SRBC 的結果
得到更強支持
以下結構已在 real-case machine formalization 中真正可執行:
- completion family;
- query-relative status;
- countermodel certificate;
- parameterized answer family;
- coarse / fine query split;
- underdetermination origin typing;
- proof-scope provenance。
必須再修正
- source-open slot 與 QDU 分開;
- QDU 僅對 completed admissible models 定義;
- source specification state 獨立成 ;
- certificate 必須明示 exact theorem / enumeration / countermodel / stress;
- countermodel pair 只證非唯一與 GAP lower bound,不能自動宣稱完整 answer range。
16. Round 05 最重要的結構
現在 SRBC 的流程可以排成:
其中:
以及:
除非答案真的分岔。
最後:
除非 origin analysis 確實指向 source / protocol / measure 等 specification layer。
到 Round 05,最早較直覺式的 SRBC 已經被四輪驗證逐步修成 typed framework。
17. Round 06
下一輪最自然的不是再加 puzzle,而是:
對本輪 14 個 certificates 主動做:
- 刪 assumption;
- 加錯 assumption;
- 改 query;
- 改 support;
- 改 prior;
- 改 observation likelihood;
- 把 stress evidence 冒充 theorem。
直接測 certificate system 是否能:
- 正確 invalidate;
- 找到最小破壞條件;
- 保留仍有效的子證書;
- 給出 reopen reason。
這才真正開始測 QRSRC / SART 本身。
參考錨點
- Adam Elga, Self-locating belief and the Sleeping Beauty problem, Analysis 60(2), 143–147, 2000.
- Stanford Encyclopedia of Philosophy, Infinity: Tacitly Infinite Decision Problems — Two Envelopes.
- Stanford Encyclopedia of Philosophy, Causal Decision Theory / Newcomb's Problem.
- Scientific American / Martin Gardner two-child problem historical formulation.
- SRBC-06 Monty Hall Semantic Surgery.
附錄 A:Machine results
{
"monty_hall": {
"general_posterior": "l2*p2/(l1*p1 + l2*p2)",
"standard_symmetric": "2/3",
"random_unchosen": "1/2",
"standard_goat_family": "1/(alpha + 1)",
"family_derivative": "-1/(alpha + 1)**2",
"family_range_alpha_0_1": [
"1/2",
"1"
],
"ex_ante_always_switch": "2/3",
"ex_ante_tie_break_dependency": false
},
"boy_girl": {
"at_least_one_boy": "1/3",
"random_child_report_boy": "1/2",
"older_child_girl": "1/2",
"general_report_formula": "a/(a + b + c)"
},
"bertrand_cube": {
"family": "U=L^r Uniform(0,1), r>0",
"P_L_lt_half": "2**(-r)",
"derivative": "-log(2)/2**r",
"r_1_2_3": {
"1": "1/2",
"2": "1/4",
"3": "1/8"
},
"range_r_positive": "(0,1)"
},
"two_envelopes": {
"conditional_expected_other": "y*(p_half + 4*p_y)/(2*(p_half + p_y))",
"conditional_gain": "-y*(p_half - 2*p_y)/(2*(p_half + p_y))",
"switch_condition": "2*p_y > p_half",
"prior_A_E_other_given_Y4": "50/7",
"prior_B_E_other_given_Y4": "20/7",
"prior_A_switch": true,
"prior_B_switch": false,
"ex_ante_expected_gain": "0",
"random_rational_priors_checked": 10000,
"stress_pass": true
},
"formal_status": {
"MH": {
"open_slots": [
"host policy",
"tie-breaking",
"prior"
],
"status": "QRU for observed-door query across protocol completions; QRC ex-ante under standard protocol."
},
"BG": {
"open_slots": [
"information-generation protocol"
],
"status": "QRU across report protocols; QRC after protocol is fixed."
},
"BER": {
"open_slots": [
"probability measure"
],
"status": "QRU across completed measures; source openness alone is not QDU."
},
"TE": {
"open_slots": [
"prior distribution",
"expectation regime"
],
"status": "QRU across proper finite priors for observed-value query; ex-ante query QRC under finite expectation."
}
}
}
附錄 B:Artifacts
SRBC_Theory_Validation_Round05_Report_v0.1_2026-09-23.mdSRBC_Core_Patch_v0.5_2026-09-23.mdresults_round05.jsoncertificate_index.jsoncertificates/*.jsonround05_replay.pymanifest.yamlvalidation.json