← Archive
lm-004254 · 2026-10

ESC-EXP-27:Shared Semantic Residual Factorization

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

ESC-EXP-27:Shared Semantic Residual Factorization

系列: Extensional Structural Convergence — Experimental Phase
文件編號: ESC-EXP-27
版本: v0.1
日期: 2026-09-23
前置: ESC-00 ~ ESC-06、ESC-EXP-00 ~ ESC-EXP-26
狀態: Shared Latent Factor / Overlapping Task Provenance Experiment

作者: Neo.K
機構: EveMissLab/一言諾科技有限公司


摘要

ESC-EXP-26 建立了:

RJ(H)=H(J∣H)\boxed{ R_J(H) = H(J\mid H) }

並證明:

future task若只需要 underlying state 的一部分 distinction,就不必每次解碼完整 universal provenance。

但 EXP-26 deliberately 使用兩個完全互補、互不重疊的 selective task layers:

MID,OUTER.MID, \quad OUTER.

它們的 provenance bits沒有重複,因此:

RMID+ROUTER=RFULL.R_{\mathrm{MID}} + R_{\mathrm{OUTER}} = R_{\mathrm{FULL}}.

這是一個乾淨 baseline,但真實 task families通常不是這樣。

多個 tasks 很可能:

  • 各自需要一些 private distinctions;
  • 同時重複依賴一批 shared latent distinctions。

如果我們直接:

每個 task各存一份 task-specific provenance code,

那 shared latent bits 就會被重複保存。

本輪因此故意建立兩個 overlap tasks:

JAJ_A

與:

JB.J_B.

Hot representation仍是 H4。

Hot 已知:

(T,D),(T,D),

隱藏:

G.G.

Task A 需要 GG 的 Hot regions:

{00,01,10}\boxed{ \{00,01,10\} }

Task B 需要:

{01,10,11}.\boxed{ \{01,10,11\}. }

所以:

A-only region

{00}.\{00\}.

Shared region

{01,10}.\boxed{ \{01,10\}. }

B-only region

{11}.\{11\}.

兩個 tasks 的 union:

A∪B={00,01,10,11}\boxed{ A\cup B = \{00,01,10,11\} }

剛好覆蓋完整 hidden residual。

因此本輪可以比較兩種 coding architecture。

Naive Task-Wise Coding

分別保存:

ZAZ_A

與:

ZB.Z_B.

storage:

H(JA∣H)+H(JB∣H).\boxed{ H(J_A\mid H) + H(J_B\mid H). }

Shared distinctions被存兩次。

Shared-Factor Coding

改存:

ZA-only,Zshared,ZB-only.\boxed{ Z_{A\text{-only}}, \quad Z_{\mathrm{shared}}, \quad Z_{B\text{-only}}. }

Task A讀:

ZA-only+Zshared.Z_{A\text{-only}} + Z_{\mathrm{shared}}.

Task B讀:

Zshared+ZB-only.Z_{\mathrm{shared}} + Z_{B\text{-only}}.

Joint A+B讀三層。

因此理想 storage:

H(ZA-only,Zshared,ZB-only∣H).\boxed{ H( Z_{A\text{-only}}, Z_{\mathrm{shared}}, Z_{B\text{-only}} \mid H ). }

本輪主要結果:

Uniform Prior

Task A:

0.75 bit.0.75 \text{ bit}.

Task B:

0.75 bit.0.75 \text{ bit}.

Joint:

1 bit.1 \text{ bit}.

所以 naive task-wise storage:

1.5.1.5.

factorized storage:

1.1.

省:

0.5 bit/state\boxed{ 0.5\text{ bit/state} }

即:

33.33%.\boxed{ 33.33\%. }

這 0.50.5 bit正好等於 shared factor本身:

H(Zshared∣H)=0.5.\boxed{ H(Z_{\mathrm{shared}}\mid H) = 0.5. }

Skewed Prior

Task A:

0.648390484031.0.648390484031.

Task B:

0.441549035869.0.441549035869.

Joint:

0.811278124459.0.811278124459.

所以 naive storage:

1.089939519900.1.089939519900.

factorized storage:

0.811278124459.0.811278124459.

省:

0.278661395441 bits/state.\boxed{ 0.278661395441 \text{ bits/state}. }

即:

25.57%.\boxed{ 25.57\%. }

這個 savings 又精確等於 shared factor:

0.278661395441.\boxed{ 0.278661395441. }

因此:

conditional task redundancy=shared latent factor information\boxed{ \text{conditional task redundancy} = \text{shared latent factor information} }

在本 finite overlap geometry 中完全成立。


1. Runtime

EXP-27 regression:

6 passed

測試內容:

  • uniform overlap geometry 精確為:RA=RB=0.75,RS=0.5,RAB=1;R_A=R_B=0.75, \quad R_S=0.5, \quad R_{AB}=1;
  • shared factor等於 conditional task redundancy;
  • A+B joint tasks精確恢復 full residual;
  • factorized storage小於 naive per-task storage;
  • factorized retrieval同時優於 monolithic與 naive task-wise read;
  • uniform / skewed shared factor都非零。

2. Hot Geometry

Hot H4:

{{∅,G},{D,DG},{T,TG},{TD,TDG}}.\{ \{\varnothing,G\}, \{D,DG\}, \{T,TG\}, \{TD,TDG\} \}.

四個 Hot regions可用:

(T,D)(T,D)

標記:

00,01,10,11.00,\quad01,\quad10,\quad11.

每個 region內只剩:

GG

尚未決定。


3. Overlap Tasks

Task A:

JA=Gon {00,01,10}.\boxed{ J_A = G \quad \text{on } \{00,01,10\}. }

Task B:

JB=Gon {01,10,11}.\boxed{ J_B = G \quad \text{on } \{01,10,11\}. }

所以:

A∩B={01,10}.A\cap B = \boxed{ \{01,10\} }.

4. Factor Regions

定義三個 semantic factors:

ZAZ_A

只服務:

00.00. ZSZ_S

服務 shared:

01,10.01,10. ZBZ_B

只服務:

11.11.

所以:

JA⟺(ZA,ZS)\boxed{ J_A \Longleftrightarrow (Z_A,Z_S) }

在給定 Hot 後成立。

同理:

JB⟺(ZS,ZB).\boxed{ J_B \Longleftrightarrow (Z_S,Z_B). }

而:

(JA,JB)⟺(ZA,ZS,ZB).\boxed{ (J_A,J_B) \Longleftrightarrow (Z_A,Z_S,Z_B). }

5. Joint Tasks Recover Full Hidden State

因:

A∪B={00,01,10,11},A\cup B = \{00,01,10,11\},

joint task family知道每個 Hot region中的:

G.G.

所以:

H(JA,JB∣H)=H(G∣H).\boxed{ H( J_A,J_B \mid H ) = H( G \mid H ). }

也就是:

RAB=RFULL.\boxed{ R_{AB} = R_{\mathrm{FULL}}. }

Uniform / skewed都精確成立。


6. Conditional Redundancy

對兩 tasks:

JA,JB,J_A,J_B,

定義:

RAB∣H=H(JA∣H)+H(JB∣H)−H(JA,JB∣H).\boxed{ \mathcal R_{AB\mid H} = H(J_A\mid H) + H(J_B\mid H) - H(J_A,J_B\mid H). }

如果:

R>0,\mathcal R>0,

代表兩個 task-specific code中有 shared information。


7. Uniform Conditional Redundancy

Uniform:

H(JA∣H)=0.75,H(J_A\mid H) = 0.75, H(JB∣H)=0.75,H(J_B\mid H) = 0.75, H(JA,JB∣H)=1.H(J_A,J_B\mid H) = 1.

所以:

RAB∣H=0.5.\boxed{ \mathcal R_{AB\mid H} = 0.5. }

8. Shared Factor Information

Shared regions:

01,10.01, 10.

每個 Hot region probability:

0.25,0.25,

而其 hidden GG entropy為:

1.1.

所以 shared residual:

RS=0.25+0.25=0.5.\boxed{ R_S = 0.25+0.25 = 0.5. }

因此:

RAB∣H=RS.\boxed{ \mathcal R_{AB\mid H} = R_S. }

9. Skewed Conditional Redundancy

Skewed:

RA=0.648390484031,R_A = 0.648390484031, RB=0.441549035869,R_B = 0.441549035869, RAB=0.811278124459.R_{AB} = 0.811278124459.

所以:

RAB∣H=0.278661395441.\boxed{ \mathcal R_{AB\mid H} = 0.278661395441. }

10. Skewed Shared Factor

直接計算 shared regions:

01,1001, 10

所對應的 hidden conditional entropy contribution:

RS=0.278661395441.\boxed{ R_S = 0.278661395441. }

所以再次:

RAB∣H=RS.\boxed{ \mathcal R_{AB\mid H} = R_S. }

11. Task Boundary vs Latent-Factor Boundary

Naive task-wise coding以:

task\boxed{ \text{task} }

當 coding boundary。

所以:

ZAZ_A

與:

ZBZ_B

各自包含 shared residual。

Factorized coding改以:

latent distinction region\boxed{ \text{latent distinction region} }

當 boundary。

因此 shared部分只存一次。

這是本輪最核心的 coding architecture差異。


12. Uniform Storage Comparison

Naive:

RA+RB=0.75+0.75=1.5.R_A+R_B = 0.75+0.75 = \boxed{ 1.5. }

Factorized:

RA-only+RS+RB-onlyR_{A\text{-only}} + R_S + R_{B\text{-only}} =0.25+0.5+0.25=1.= 0.25+0.5+0.25 = \boxed{ 1. }

所以:

33.33%\boxed{ 33.33\% }

storage reduction。


13. Skewed Storage Comparison

A-only:

0.369729088590.\boxed{ 0.369729088590. }

Shared:

0.278661395441.\boxed{ 0.278661395441. }

B-only:

0.162887640428.\boxed{ 0.162887640428. }

總:

0.811278124459.\boxed{ 0.811278124459. }

naive:

0.648390484031+0.441549035869=1.089939519900.0.648390484031 + 0.441549035869 = 1.089939519900.

省:

0.278661395441.\boxed{ 0.278661395441. }

14. Storage Savings 正好是 Shared Factor

一般在本 separable overlap construction:

RA=RA-only+RS,R_A = R_{A\text{-only}} + R_S, RB=RS+RB-only,R_B = R_S + R_{B\text{-only}},

而:

RAB=RA-only+RS+RB-only.R_{AB} = R_{A\text{-only}} + R_S + R_{B\text{-only}}.

所以:

RA+RB−RAB=RS.R_A+R_B-R_{AB} = \boxed{ R_S. }

即:

duplicated storage=shared latent information.\boxed{ \text{duplicated storage} = \text{shared latent information}. }

15. Huffman Storage

Uniform one-shot Huffman:

naive:

1.5.1.5.

factorized:

1.1.

仍省:

0.5.0.5.

Skewed:

naive:

1.343484419263.\boxed{ 1.343484419263. }

factorized:

1.\boxed{ 1. }

省:

0.343484419263.\boxed{ 0.343484419263. }

16. 為什麼 Huffman Savings 比 Shannon Shared Bit 更大?

Skewed Shannon shared factor:

0.278661395441.0.278661395441.

Huffman storage savings:

0.343484419263.0.343484419263.

因 task-wise one-shot coding不只重複 shared information,

還各自支付 prefix quantization overhead。

factorization同時消掉:

  • semantic duplication;
  • 部分 coding overhead duplication。

17. Retrieval Workload

本輪 workload:

P(A)=0.45,P(A)=0.45, P(B)=0.45,P(B)=0.45, P(AB)=0.10.P(AB)=0.10.

所有 queries都需要某種 hidden residual。


18. Universal Monolithic Read

如果每次都讀完整 universal residual:

Uniform:

1 bit/query.\boxed{ 1\text{ bit/query}. }

Skewed:

0.811278124459.\boxed{ 0.811278124459. }

19. Naive Task-Wise Read

A query讀:

RA.R_A.

B query讀:

RB.R_B.

AB query因兩份 task code彼此獨立,要讀:

RA+RB.R_A+R_B.

Uniform:

0.825.\boxed{ 0.825. }

Skewed:

0.599466735945.\boxed{ 0.599466735945. }

20. Factorized Read

A query讀:

ZA+ZS.Z_A+Z_S.

B query讀:

ZS+ZB.Z_S+Z_B.

AB query讀:

ZA+ZS+ZB.Z_A+Z_S+Z_B.

Uniform:

0.775.\boxed{ 0.775. }

Skewed:

0.571600596401.\boxed{ 0.571600596401. }

21. Factorized vs Monolithic

Uniform:

1→0.775.1 \rightarrow 0.775.

下降:

22.5%.\boxed{ 22.5\%. }

Skewed:

0.811278→0.571601.0.811278 \rightarrow 0.571601.

下降:

29.54%.\boxed{ 29.54\%. }

22. Factorized vs Naive Task-Wise

Uniform:

0.825→0.775.0.825 \rightarrow 0.775.

下降:

6.06%.\boxed{ 6.06\%. }

Skewed:

0.599467→0.571601.0.599467 \rightarrow 0.571601.

下降:

4.65%.\boxed{ 4.65\%. }

這個差距主要來自:

P(AB)=0.10P(AB)=0.10

的 joint queries不再重複讀 shared factor。

若 joint-task probability更高,這個收益會更大。


23. Storage Savings 與 Retrieval Savings 是不同來源

Storage savings:

Shared factor不用存兩次。

Retrieval savings:

Joint query不用讀 Shared factor兩次。

所以:

shared factorization improves both write/storage and read/retrieval sides.\boxed{ \text{shared factorization improves both write/storage and read/retrieval sides}. }

24. Shared Factor Fraction

Uniform full residual:

1.1.

shared:

0.5.0.5.

所以:

50%\boxed{ 50\% }

的 universal residual同時被 A 與 B 使用。

Skewed:

0.2786613954410.811278124459=34.35%.\frac{ 0.278661395441 }{ 0.811278124459 } = \boxed{ 34.35\%. }

所以相同 geometric overlap,在不同 prior下有不同 information overlap。


25. Geometry Overlap ≠ Information Overlap

A/B overlap regions占:

2/4=50%.2/4 = 50\%.

但 skewed information overlap只有:

34.35%.34.35\%.

因此:

overlap in state-space geometry≠overlap in information mass.\boxed{ \text{overlap in state-space geometry} \neq \text{overlap in information mass}. }

這和 EXP-25 的 block-count vs bits差異完全一致。


26. Shared-Latent Factor Value

可以定義:

VS=H(JA∣H)+H(JB∣H)−H(JA,JB∣H).\boxed{ V_S = H(J_A\mid H) + H(J_B\mid H) - H(J_A,J_B\mid H). }

它就是:

task-wise provenance storage中最少可被共享/去重的 conditional information。

本輪:

Uniform:

VS=0.5.V_S=0.5.

Skewed:

VS=0.278661395441.V_S=0.278661395441.

27. Shared Semantic Factor

因此可以把:

ZSZ_S

定義為:

能同時服務多個 future tasks 的最小共同 residual component.\boxed{ \text{能同時服務多個 future tasks 的最小共同 residual component}. }

本 finite construction中它具有非常直接的 latent-region realization。


28. Task Factorization vs Common Information

一般情況下,shared factor未必能像本輪這麼乾淨地對應 disjoint state regions。

更一般可能需要尋找:

ZSZ_S

使:

H(JA∣H,ZA,ZS)=0,H(J_A\mid H,Z_A,Z_S)=0, H(JB∣H,ZB,ZS)=0,H(J_B\mid H,Z_B,Z_S)=0,

並最小化:

H(ZA,ZS,ZB∣H).H(Z_A,Z_S,Z_B\mid H).

這開始接近 conditional common-information / multiterminal coding 類型的問題。

本輪只是最簡單 finite baseline。


29. Shared Factor 不能只靠 Task 名稱決定

如果兩個 tasks:

  • 名稱不同;
  • output格式不同;

但 underlying residual需求高度重疊,

仍應共用同一 latent layer。

反過來,同一 task family中的不同 contexts也可能需要完全不同 residual factors。

所以:

code factorization should follow latent dependency, not API/task labels.\boxed{ \text{code factorization should follow latent dependency, not API/task labels}. }

30. Semantic Residual Factor Graph

可以把 future tasks與 latent factors建成 bipartite graph:

Gtask−factor=(J,Z,E).\boxed{ G_{\mathrm{task-factor}} = ( \mathcal J, \mathcal Z, E ). }

本輪:

JA↔{ZA,ZS},J_A \leftrightarrow \{ Z_A,Z_S \}, JB↔{ZS,ZB}.J_B \leftrightarrow \{ Z_S,Z_B \}.

這已經是一個最小 shared-factor graph。


31. Factor Degree

Shared factor:

ZSZ_S

degree:

2.2.

private factors:

ZA,ZBZ_A,Z_B

degree:

1.1.

未來如果有大量 tasks,

高-degree latent factors可能值得:

  • Hot-cache;
  • 優先 replicate;
  • 強保護;
  • 高可用性 storage。

因此 factor degree會成為新的 tier-placement signal。


32. Frequency × Sharedness

一個 residual factor的 operational value不只取決於:

H(Z∣H).H(Z\mid H).

還取決於:

∑J:Z∈I(J)P(J).\boxed{ \sum_{J: Z\in I(J)} P(J). }

也就是它被多少/多常 tasks使用。

所以 Hot promotion policy可以由:

information bits×task reuse frequency\boxed{ \text{information bits} \times \text{task reuse frequency} }

推動。


33. Shared Factor 優先 Hot 的可能性

如果:

ZSZ_S

被 A / B / C / D 多個 tasks共同依賴,

即使它本身 information size不大,

把它放 Hot可能非常划算。

private factors則留 Cold。

所以 multi-tier coding的 natural architecture可能是:

Hot shared core+Cold private residuals.\boxed{ \text{Hot shared core} + \text{Cold private residuals}. }

34. 這與 Representation Core 很接近

如果跨大量 tasks都反覆使用同一 latent residual,

那這個 residual開始具有:

representation core\boxed{ \text{representation core} }

性質。

所以 ESC可以從:

shared provenance factor

進一步發展成:

minimal common latent substrate across task families。


35. Factorized Task Lattice

本輪 factors:

ZA,ZS,ZB.Z_A, Z_S, Z_B.

可形成 layer subsets。

重要 nodes:

E0,E_0, EA,E_A, ES,E_S, EB,E_B, EAS,E_{AS}, ESB,E_{SB}, EASB.E_{ASB}.

其中:

EASE_{AS}

足以回答 Task A,

ESBE_{SB}

足以回答 Task B,

EASBE_{ASB}

足以回答 joint/full task。


36. Task Lattice 不再只是 B2B_2

EXP-26 是兩個完全互補 task layers,

所以 recoverability requirement lattice是:

B2.B_2.

EXP-27 加入 shared factor後,

底層 factor subset lattice更接近:

B3\boxed{ B_3 }

但只有部分 nodes對應完整 task contracts。

所以:

factor lattice≠task lattice.\boxed{ \text{factor lattice} \neq \text{task lattice}. }

這是很重要的新區分。


37. Task Quotient of Factor Lattice

多個 factor subsets可能對目前 task set具有相同 operational utility。

所以可以再做 quotient:

P(Z)→Task Capability Classes.\boxed{ \mathcal P(\mathcal Z) \rightarrow \text{Task Capability Classes}. }

這其實又回到 ESC effect quotient / contextual equivalence。


38. Representation Factorization 又回到 ESC 主題

我們現在有:

  • underlying state;
  • Hot projection;
  • task outputs;
  • residual factors;
  • factor subsets;
  • task capability equivalence。

不同 representations再次需要比較:

what distinctions they preserve, what tasks they support, what future composition they enable.\boxed{ \text{what distinctions they preserve, what tasks they support, what future composition they enable}. }

這幾乎正是 ESC 最早的共同底空間命題。


39. Shared Factor Storage Lower Bound

本輪 separable overlap case:

RS=RAB∣H.\boxed{ R_S = \mathcal R_{AB\mid H}. }

因此任何不重複保存 task-specific codes的 factorization,至少需要保存這一份 shared information。


40. Naive Duplication Tax

定義:

Tdup=RA+RB−RAB.\boxed{ T_{\mathrm{dup}} = R_A+R_B-R_{AB}. }

本輪:

Tdup=RS.\boxed{ T_{\mathrm{dup}} = R_S. }

它可以被解讀成:

使用 task boundaries 而不是 latent-factor boundaries 所支付的 storage tax。


41. Uniform Duplication Tax

Tdup=0.5 bit/state.\boxed{ T_{\mathrm{dup}} = 0.5\text{ bit/state}. }

相對 naive:

1.5,1.5,

比例:

33.33%.\boxed{ 33.33\%. }

42. Skewed Duplication Tax

0.278661395441 bits/state.\boxed{ 0.278661395441 \text{ bits/state}. }

相對:

1.089939519900,1.089939519900,

比例:

25.57%.\boxed{ 25.57\%. }

43. One-Shot Coding Duplication Tax

Skewed Huffman:

naive:

1.343484419263.1.343484419263.

factorized:

1.1.

所以:

Tdup,Huff=0.343484419263.\boxed{ T_{\mathrm{dup,Huff}} = 0.343484419263. }

這比 Shannon semantic redundancy更高,因還包含 duplicated coding overhead。


44. Shared Factorization Principle

因此可以提出:

Shared Semantic Residual Factorization Principle.\boxed{ \text{Shared Semantic Residual Factorization Principle}. }

若多個 future tasks對 Hot representation存在重疊 residual dependency,

則 provenance code不應簡單按 task各自獨立儲存。

應優先尋找:

private factors+shared factors\boxed{ \text{private factors} + \text{shared factors} }

使 joint task code避免重複保存 shared information。


45. 本輪錨點

ESC-EXP-27.ARAB∣H=H(JA∣H)+H(JB∣H)−H(JA,JB∣H).\boxed{ \textbf{ESC-EXP-27.A} \quad \mathcal R_{AB\mid H} = H(J_A\mid H) + H(J_B\mid H) - H(J_A,J_B\mid H). } ESC-EXP-27.BRAB∣H=RS\boxed{ \textbf{ESC-EXP-27.B} \quad \mathcal R_{AB\mid H} = R_S }

在本 finite overlap construction精確成立。

ESC-EXP-27.Cfactorized storage相對 naive task-wise storage降低 33.33%(uniform)與 25.57%(skewed)。\boxed{ \textbf{ESC-EXP-27.C} \quad \text{factorized storage相對 naive task-wise storage降低 }33.33\%\text{(uniform)與 }25.57\%\text{(skewed)。} } ESC-EXP-27.Dfactorized expected retrieval同時低於 universal monolithic 與 naive task-wise retrieval。\boxed{ \textbf{ESC-EXP-27.D} \quad \text{factorized expected retrieval同時低於 universal monolithic 與 naive task-wise retrieval。} } ESC-EXP-27.Ecode boundary應追隨 shared latent dependency,而不應只追隨 task/API boundary。\boxed{ \textbf{ESC-EXP-27.E} \quad \text{code boundary應追隨 shared latent dependency,而不應只追隨 task/API boundary。} }

46. 從 EXP-26 到 EXP-27

EXP-26:

Which provenance layer does each task need?\boxed{ \text{Which provenance layer does each task need?} }

EXP-27:

Which provenance layers are actually shared by multiple tasks?\boxed{ \text{Which provenance layers are actually shared by multiple tasks?} }

所以:

Task-Adaptive Coding→Shared Latent Residual Factorization.\boxed{ \text{Task-Adaptive Coding} \rightarrow \text{Shared Latent Residual Factorization}. }

47. 下一輪:ESC-EXP-28

EXP-27 的 shared factor還是人工設計:

A-only,Shared,B-only.A\text{-only}, \quad \text{Shared}, \quad B\text{-only}.

真正下一步應該讓 runtime自己找 factorization。

也就是:

Automatic Semantic Residual Factor Discovery.\boxed{ \text{Automatic Semantic Residual Factor Discovery}. }

給定:

  • Hot representation HH ;
  • 多個 future tasks:J1,…,Jm;J_1,\dots,J_m;
  • task demand distribution;
  • storage / read cost;

自動搜尋 latent factors:

Z1,…,ZkZ_1,\dots,Z_k

以及每個 task要讀的 factor subset:

I(Ji).I(J_i).

目標:

min⁡[total stored provenance+λexpected task read cost]\boxed{ \min \left[ \text{total stored provenance} + \lambda \text{expected task read cost} \right] }

subject to:

H(Ji∣H,ZI(Ji))=0∀i.\boxed{ H( J_i \mid H, Z_{I(J_i)} ) = 0 \quad \forall i. }

真正問題變成:

AI 能不能自己從 task family 中發現「哪些 hidden distinctions 應該共用一層 residual factor」?

這會把人工 semantic decomposition推進到真正可自動編譯的:

provenance factor compiler.\boxed{ \text{provenance factor compiler}. }