數學問題不是只有可解與不可解:多域問題診斷與受約束數學域補全
Mathematical Problems Are Not Merely Solved or Unsolved: Multi-Domain Diagnosis and Constrained Mathematical Domain Completion
系列:Autonomous Mathematical Research / Paper 02 of 04 版本:v0.1 日期:2026-08-23 作者:Neo.K
摘要
在傳統數學解題與自動定理證明框架中,一個問題往往被簡化成「已解/未解」、「證成/未證成」或「成功/失敗」。然而,真正研究級數學中的失敗通常不是單一狀態。問題可能尚未被正確定義,必要假設可能缺失,成功條件可能模糊,當前表示可能遮蔽關鍵結構,方法域可能不充分,證明依賴可能缺少中間引理,計算資源可能不足,形式化可能偏離原意,甚至所謂「新結果」可能只是既有數學的重新命名。若 AI 將所有這些情況都壓縮為「解不出來」,則增加推理時間、模型大小或 proof search budget 並不一定能改善研究品質。
本文提出 Multi-Domain Mathematical Diagnosis(多域數學問題診斷) 與 Constrained Mathematical Domain Completion, CMDC(受約束數學域補全) 。本文首先將研究狀態拆分為一組可擴張的數學域:問題/規格域、定義域、假設/邊界域、判定域、表示域、解決域、方法域、引理/依賴域、計算/搜尋域、例證/反例域、驗證/忠實性域、理論橋接域與新穎性/有趣性域。對任一研究狀態,診斷器產生 Mathematical Domain Gap Map ,其中每個 gap 具有類型、目標、證據、信心、阻塞程度、依賴與候選修復方式。
本文接著提出 CMDC:當 AI 辨認某一缺口後,可以提出對應的數學補全,例如新增定義、澄清術語、補必要假設、改變表示、生成中間引理、建立新方法、產生反例、加入 verifier、或建立與既有理論的橋接。但任何 repair 都只能先成為候選變更,而不能直接覆寫原問題或升格為正式數學。本文因此定義 Typed Repair , Repair Contract , Gap Closure Condition , Repair-Induced Obligation , Repair Dependency Graph 與 Local Domain Completion ,並提出一組禁止「捷徑式修復」的守恆條件:問題身份必須保留、非保守改變必須明示、新定義不得以重新命名冒充新理論、證明成功不得取代語義忠實性、以及局部補全不得被誇張成全域數學完成。
本文最後將該方法論映射至 Autonomous Mathematical Research Runtime:以 missing / failures / uncertainty / risks 等公開狀態為輸入,讓 AI 自主決定是否 DIAGNOSE / REFRAME / DEFINE / ADD_ASSUMPTION / GENERATE_LEMMA / SEARCH_COUNTEREXAMPLE / CHANGE_METHOD / INCREASE_COMPUTE / BRIDGE / VERIFY / STOP / DEFER / ESCALATE,並建立可測量的 diagnosis accuracy、repair validity、gap-closure precision、obligation coverage、problem-identity preservation 與 downstream research gain。本文的核心命題是:數學自治的前提不是 AI 永遠能解題,而是 AI 能辨認自己究竟在哪一個域失敗,以及在不偷換問題的條件下,合法地生成使研究重新可進行的缺失結構。
關鍵詞: 數學問題診斷、受約束數學域補全、CMDC、問題重構、問題表示、數學缺口、證明修復、自動猜想、理論探索、自主數學研究
1. 引言:Unsolved 是一個過度壓縮的狀態
給定數學問題:
Q . Q. Q .
最簡單的分類是:
S t a t u s ( Q ) ∈ { S o l v e d , U n s o l v e d } . Status(Q)
\in
\{
Solved,
Unsolved
\}. S t a t u s ( Q ) ∈ { S o l v e d , U n so l v e d } .
對形式證明系統,則可能是:
S t a t u s ( Q ) ∈ { P r o v e d , U n p r o v e d } . Status(Q)
\in
\{
Proved,
Unproved
\}. S t a t u s ( Q ) ∈ { P r o v e d , U n p r o v e d } .
但這些標籤對真正研究過程而言資訊量過低。一個未解問題可能至少包含:
definition missing , \text{definition missing}, definition missing ,
assumption missing , \text{assumption missing}, assumption missing ,
representation mismatch , \text{representation mismatch}, representation mismatch ,
method failure , \text{method failure}, method failure ,
missing lemma , \text{missing lemma}, missing lemma ,
counterexample not searched , \text{counterexample not searched}, counterexample not searched ,
compute budget insufficient , \text{compute budget insufficient}, compute budget insufficient ,
formalization unfaithful , \text{formalization unfaithful}, formalization unfaithful ,
或:
problem itself malformed . \text{problem itself malformed}. problem itself malformed .
因此:
U n s o l v e d = Compress ( F a i l u r e S t r u c t u r e ) . \boxed{
Unsolved
=
\operatorname{Compress}
(
FailureStructure
).
} U n so l v e d = Compress ( F ai l u r e S t r u c t u r e ) .
本文的第一個主張是:
Before solving a problem, an autonomous mathematical AI should be able to diagnose the structure of its failure. \boxed{
\text{Before solving a problem, an autonomous mathematical AI should be able to diagnose the structure of its failure.}
} Before solving a problem, an autonomous mathematical AI should be able to diagnose the structure of its failure.
也就是,真正需要的函數不只是:
S o l v e ( Q ) , Solve(Q), S o l v e ( Q ) ,
而是:
D i a g n o s e ( Q , S t , K t ) . Diagnose(Q,S_t,K_t). D ia g n ose ( Q , S t , K t ) .
2. 為什麼增加計算能力不能解決所有數學失敗
假設某系統目前對問題 Q Q Q 的計算能力為:
C t . C_t. C t .
若主要瓶頸屬於 computation/search domain,增加:
C t → C t + 1 C_t
\rightarrow
C_{t+1} C t → C t + 1
可能有效。
但若問題真正是:
D e f i n i t i o n ( Q ) = A m b i g u o u s , Definition(Q)=Ambiguous, D e f ini t i o n ( Q ) = A mbi g u o u s ,
則:
C t + 1 ≫ C t C_{t+1}\gg C_t C t + 1 ≫ C t
不保證:
A m b i g u i t y → R e s o l v e d . Ambiguity
\rightarrow
Resolved. A mbi g u i t y → R eso l v e d .
同樣地,若形式化命題:
F ( Q ) F(Q) F ( Q )
根本不忠於原問題 Q Q Q ,則對:
F ( Q ) F(Q) F ( Q )
投入更多 proof search,只會更有效率地證明錯誤目標。
因此:
M o r e C o m p u t e ⇏ B e t t e r P r o b l e m S t a t e . \boxed{
MoreCompute
\not\Rightarrow
BetterProblemState.
} M or e C o m p u t e ⇒ B e tt er P r o b l e m S t a t e .
更完整地說:
R e s e a r c h P r o g r e s s = f ( P r o b l e m Q u a l i t y , D e f i n i t i o n s , A s s u m p t i o n s , R e p r e s e n t a t i o n , M e t h o d s , D e p e n d e n c i e s , C o m p u t e , V e r i f i c a t i o n , B r i d g e s ) . ResearchProgress
=
f(
ProblemQuality,
Definitions,
Assumptions,
Representation,
Methods,
Dependencies,
Compute,
Verification,
Bridges
). R ese a r c h P r o g r ess = f ( P r o b l e m Q u a l i t y , D e f ini t i o n s , A ss u m pt i o n s , R e p r ese n t a t i o n , M e t h o d s , D e p e n d e n c i es , C o m p u t e , V er i f i c a t i o n , B r i d g es ) .
3. 從 Failure Code 到 Mathematical Domain State
傳統軟體系統常將失敗分成:
syntax_error
type_error
runtime_error
timeout
這些分類讓系統能夠採取不同修復策略。數學研究也需要類似但更高階的診斷空間。
令:
S t M S_t^M S t M
表示第 t t t 時點的數學研究狀態。
我們不只問:
S u c c e s s ( S t M ) ? Success(S_t^M)? S u ccess ( S t M )?
而要抽取:
G a p S t a t e ( S t M ) . GapState(S_t^M). G a pS t a t e ( S t M ) .
因此:
S t M → D i a g n o s e Δ t . S_t^M
\xrightarrow{Diagnose}
\Delta_t. S t M D ia g n ose Δ t .
其中:
Δ t = { δ 1 , δ 2 , … , δ n } \Delta_t
=
\{
\delta_1,\delta_2,\ldots,\delta_n
\} Δ t = { δ 1 , δ 2 , … , δ n }
稱為:
Mathematical Domain Gap Map . \boxed{
\text{Mathematical Domain Gap Map}.
} Mathematical Domain Gap Map .
4. 第一版 Mathematical Research Domain Set
本文採用可擴張域集合:
D ( 1 ) = { D P , D D , D A , D J , D R , D S , D M , D L , D C , D E , D V , D B , D N } . \mathfrak D^{(1)}
=
\{
D_P,
D_D,
D_A,
D_J,
D_R,
D_S,
D_M,
D_L,
D_C,
D_E,
D_V,
D_B,
D_N
\}. D ( 1 ) = { D P , D D , D A , D J , D R , D S , D M , D L , D C , D E , D V , D B , D N } .
這裡上標 ( 1 ) (1) ( 1 ) 表示:這是第一版 active taxonomy,而不是不可修改的終極本體論。未來可以:
D ( 1 ) → D ( 2 ) . \mathfrak D^{(1)}
\rightarrow
\mathfrak D^{(2)}. D ( 1 ) → D ( 2 ) .
但變更必須版本化。
5. D P D_P D P :Problem / Specification Domain
D P D_P D P 處理:問題本身是否被正確指定?
令原始問題:
Q 0 . Q_0. Q 0 .
若存在多種合理語義:
I n t e r p r e t ( Q 0 ) = { Q 0 ( 1 ) , Q 0 ( 2 ) , … } , Interpret(Q_0)
=
\{
Q_0^{(1)},
Q_0^{(2)},
\ldots
\}, I n t er p r e t ( Q 0 ) = { Q 0 ( 1 ) , Q 0 ( 2 ) , … } ,
則問題首先不是 proof search,而是:
D i s a m b i g u a t i o n . Disambiguation. D i s ambi g u a t i o n .
典型 gap:
δ P a m b i g u i t y , \delta_P^{ambiguity}, δ P ambi g u i t y ,
δ P u n d e r s p e c i f i c a t i o n , \delta_P^{underspecification}, δ P u n d er s p ec i f i c a t i o n ,
δ P g o a l m i s m a t c h , \delta_P^{goal\ mismatch}, δ P g o a l mi s ma t c h ,
δ P i n c o n s i s t e n t s t a t e m e n t . \delta_P^{inconsistent\ statement}. δ P in co n s i s t e n t s t a t e m e n t .
此域的 repair 不一定產生 theorem。它可能只產生:
P r o b l e m C l a r i f i c a t i o n C a n d i d a t e . ProblemClarificationCandidate. P r o b l e m C l a r i f i c a t i o n C an d i d a t e .
6. D D D_D D D :Definition Domain
令問題依賴定義集合:
D e f s ( Q ) = { d 1 , … , d k } . Defs(Q)
=
\{
d_1,\ldots,d_k
\}. D e f s ( Q ) = { d 1 , … , d k } .
Definition Domain 診斷至少檢查:
D e f i n e d ( d i ) ? Defined(d_i)? D e f in e d ( d i )?
W e l l F o r m e d ( d i ) ? WellFormed(d_i)? W e l l F or m e d ( d i )?
N o n C i r c u l a r ( d i ) ? NonCircular(d_i)? N o n C i r c u l a r ( d i )?
B o u n d a r y C o v e r e d ( d i ) ? BoundaryCovered(d_i)? B o u n d a r y C o v er e d ( d i )?
P r i o r A r t M a p p e d ( d i ) ? PriorArtMapped(d_i)? P r i or A r tM a pp e d ( d i )?
O p e r a t i o n a l l y U s a b l e ( d i ) ? OperationallyUsable(d_i)? O p er a t i o na l l y U s ab l e ( d i )?
因此:
D e f i n i t i o n F a i l u r e ≠ P r o o f F a i l u r e . DefinitionFailure
\neq
ProofFailure. D e f ini t i o n F ai l u r e = P r oo f F ai l u r e .
一個 theorem 可以無法被合理證明,只因為它的核心對象從未被足夠精確地定義。
7. D A D_A D A :Assumption / Boundary Domain
數學命題通常是:
A ⇒ Q . A
\Rightarrow
Q. A ⇒ Q .
但自然語言問題經常省略:
A . A. A .
若 AI 發現:
Q Q Q
在一般情況下存在反例,而在:
A ∗ A^* A ∗
下可能成立,則它不能直接把:
Q Q Q
改成:
A ∗ ⇒ Q A^*\Rightarrow Q A ∗ ⇒ Q
再宣稱原問題已解。正確輸出應是:
A s s u m p t i o n G a p D e t e c t e d . AssumptionGapDetected. A ss u m pt i o n G a p D e t ec t e d .
並產生:
R e p a i r C a n d i d a t e = A d d A s s u m p t i o n ( A ∗ ) . RepairCandidate
=
AddAssumption(A^*). R e p ai r C an d i d a t e = A dd A ss u m pt i o n ( A ∗ ) .
此 repair 同時產生:
P r o b l e m M u t a t i o n O b l i g a t i o n . ProblemMutationObligation. P r o b l e m M u t a t i o n O b l i g a t i o n .
8. D J D_J D J :Judgment / Success Criterion Domain
同一句「解這個問題」可能隱含完全不同的 success criteria:
E x i s t e n c e , U n i q u e n e s s , C o n s t r u c t i b i l i t y , C o m p u t a b i l i t y , E x a c t V a l u e , B o u n d , C l a s s i f i c a t i o n , P r o o f , D i s p r o o f . Existence,
Uniqueness,
Constructibility,
Computability,
ExactValue,
Bound,
Classification,
Proof,
Disproof. E x i s t e n ce , U ni q u e n ess , C o n s t r u c t ibi l i t y , C o m p u t abi l i t y , E x a c t V a l u e , B o u n d , C l a ss i f i c a t i o n , P r oo f , D i s p r oo f .
因此成功判準應成為:
J ( Q ) . J(Q). J ( Q ) .
若:
J ( Q ) = U n d e f i n e d , J(Q)=Undefined, J ( Q ) = U n d e f in e d ,
則系統不應自行假裝:
J ( Q ) = F o r m a l P r o o f . J(Q)=FormalProof. J ( Q ) = F or ma l P r oo f .
而應先形成:
J u d g m e n t G a p . JudgmentGap. J u d g m e n tG a p .
9. D R D_R D R :Representation Domain
數學中的一個問題可以有多個表示:
R 1 ( Q ) , R 2 ( Q ) , … , R m ( Q ) . R_1(Q),R_2(Q),\ldots,R_m(Q). R 1 ( Q ) , R 2 ( Q ) , … , R m ( Q ) .
例如:
N a t u r a l L a n g u a g e ↔ S e t T h e o r y ↔ G r a p h ↔ M a t r i x ↔ G e o m e t r y ↔ P r o g r a m . NaturalLanguage
\leftrightarrow
SetTheory
\leftrightarrow
Graph
\leftrightarrow
Matrix
\leftrightarrow
Geometry
\leftrightarrow
Program. N a t u r a l L an g u a g e ↔ S e tT h eor y ↔ G r a p h ↔ M a t r i x ↔ G eo m e t r y ↔ P r o g r am .
在不同表示下:
S e a r c h S p a c e ( R i ( Q ) ) SearchSpace(R_i(Q)) S e a r c h S p a ce ( R i ( Q ))
可能有巨大差異。因此:
P r o b l e m D i f f i c u l t y ≠ R e p r e s e n t a t i o n I n v a r i a n t . \boxed{
ProblemDifficulty
\neq
RepresentationInvariant.
} P r o b l e m D i f f i c u l t y = R e p r ese n t a t i o n I n v a r ian t .
至少在實際研究成本意義上如此。
Representation repair 可以是:
R i ( Q ) → R j ( Q ) . R_i(Q)
\rightarrow
R_j(Q). R i ( Q ) → R j ( Q ) .
但它需要:
S e m a n t i c B r i d g e ( R i , R j ) . SemanticBridge(R_i,R_j). S e man t i c B r i d g e ( R i , R j ) .
否則換表示可能悄悄改變問題。
10. D S D_S D S :Solution Domain
即使問題本身清楚,AI 仍可能搜尋錯誤答案型態。
令:
S ( Q ) \mathcal S(Q) S ( Q )
為合法 solution space。
若系統一直搜尋:
C l o s e d F o r m ClosedForm C l ose d F or m
而自然結果其實是:
N o C l o s e d F o r m B u t B o u n d , NoClosedFormButBound, N o C l ose d F or m B u tB o u n d ,
則搜尋可能長期無效。因此需要:
I n f e r S o l u t i o n D o m a i n ( Q ) . InferSolutionDomain(Q). I n f er S o l u t i o n D o main ( Q ) .
Solution-domain gap 的典型 repair 是:
R e f i n e S u c c e s s O b j e c t . RefineSuccessObject. R e f in e S u ccess O bj ec t .
例如從:
E x a c t S o l u t i o n ExactSolution E x a c tS o l u t i o n
改為候選:
U p p e r B o u n d , L o w e r B o u n d , C l a s s i f i c a t i o n , C o u n t e r e x a m p l e , A l g o r i t h m . UpperBound,
LowerBound,
Classification,
Counterexample,
Algorithm. U pp er B o u n d , L o w er B o u n d , C l a ss i f i c a t i o n , C o u n t er e x am pl e , A l g or i t hm .
這類改變仍需由 Problem Identity / Research Contract 判斷是否屬於原任務。
11. D M D_M D M :Method Domain
對問題 Q Q Q ,已嘗試方法集合:
M t = { M 1 , … , M k } . \mathcal M_t
=
\{
M_1,\ldots,M_k
\}. M t = { M 1 , … , M k } .
若:
∀ M i ∈ M t , F a i l ( M i , Q ) , \forall M_i\in\mathcal M_t,
\quad
Fail(M_i,Q), ∀ M i ∈ M t , F ai l ( M i , Q ) ,
不能直接推出:
Q is unsolvable . Q
\text{ is unsolvable}. Q is unsolvable .
可能只是:
M e t h o d C o v e r a g e ( M t , Q ) MethodCoverage(\mathcal M_t,Q) M e t h o d C o v er a g e ( M t , Q )
不足。
Method gap 可導向:
R e t r i e v e M e t h o d , C o m p o s e M e t h o d s , M o d i f y M e t h o d , G e n e r a t e M e t h o d C a n d i d a t e . RetrieveMethod,
ComposeMethods,
ModifyMethod,
GenerateMethodCandidate. R e t r i e v e M e t h o d , C o m p ose M e t h o d s , M o d i f y M e t h o d , G e n er a t e M e t h o d C an d i d a t e .
但最後一項必須有更強 obligation。
12. D L D_L D L :Lemma / Dependency / Theory Domain
很多 proof failure 的真正瓶頸不是 final theorem,而是缺少:
L ∗ . L^*. L ∗ .
若:
T + L ∗ ⊢ Q , T
+
L^*
\vdash
Q, T + L ∗ ⊢ Q ,
但:
T ⊬ L ∗ T
\nvdash
L^* T ⊬ L ∗
尚未被研究,則:
L ∗ L^* L ∗
成為 research agenda。
因此:
P r o o f S e a r c h ( Q ) ProofSearch(Q) P r oo f S e a r c h ( Q )
應能轉成:
D e p e n d e n c y D i s c o v e r y ( Q ) . DependencyDiscovery(Q). D e p e n d e n cy D i sco v er y ( Q ) .
現有 automated conjecturing 與 theory exploration 已把 lemma discovery 視為機械化數學的重要方向;本文將它納入更一般的 domain diagnosis,而不是把所有 lemma generation 都視為無條件有益。
13. D C D_C D C :Computation / Search Domain
某些問題的主要缺口確實是:
B u d g e t . Budget. B u d g e t .
此時可以檢查:
T o k e n B u d g e t , S e a r c h D e p t h , M e m o r y , P a r a l l e l i s m , S o l v e r T i m e , N u m e r i c a l P r e c i s i o n . TokenBudget,
SearchDepth,
Memory,
Parallelism,
SolverTime,
NumericalPrecision. T o k e n B u d g e t , S e a r c h D e pt h , M e m or y , P a r a l l e l i s m , S o l v er T im e , N u m er i c a l P r ec i s i o n .
並建立:
C o m p u t e G a p . ComputeGap. C o m p u t e G a p .
但增加計算前仍需回答:
E x p e c t e d V a l u e ( M o r e C o m p u t e ∣ C u r r e n t D i a g n o s i s ) . ExpectedValue(
MoreCompute
\mid
CurrentDiagnosis
). E x p ec t e d V a l u e ( M or e C o m p u t e ∣ C u r r e n t D ia g n os i s ) .
若主要 gap 不在 D C D_C D C ,則應避免將算力當成預設修復。
14. D E D_E D E :Examples / Counterexamples / Evidence Domain
對 conjecture:
C . C. C .
系統應問:
E x a m p l e s ( C ) ? Examples(C)? E x am pl es ( C )?
N o n E x a m p l e s ( C ) ? NonExamples(C)? N o n E x am pl es ( C )?
B o u n d a r y C a s e s ( C ) ? BoundaryCases(C)? B o u n d a r y C a ses ( C )?
C o u n t e r e x a m p l e S e a r c h P e r f o r m e d ( C ) ? CounterexampleSearchPerformed(C)? C o u n t er e x am pl e S e a r c h P er f or m e d ( C )?
沒有反例不代表:
C = True . C=\text{True}. C = True .
因此:
N o C o u n t e r e x a m p l e F o u n d ≠ P r o o f . NoCounterexampleFound
\neq
Proof. N o C o u n t er e x am pl e F o u n d = P r oo f .
同時,反例的發現可能反過來產生:
P r o b l e m R e p a i r , A s s u m p t i o n R e p a i r , D e f i n i t i o n R e p a i r . ProblemRepair,
AssumptionRepair,
DefinitionRepair. P r o b l e m R e p ai r , A ss u m pt i o n R e p ai r , D e f ini t i o n R e p ai r .
這使 gap diagnosis 具有循環性。
15. D V D_V D V :Verification / Semantic Faithfulness Domain
Proof assistant 可以驗證:
F o r m a l S t a t e m e n t . FormalStatement. F or ma l S t a t e m e n t .
但研究系統還必須驗證:
F o r m a l S t a t e m e n t ≈ I n t e n d e d S t a t e m e n t . FormalStatement
\approx
IntendedStatement. F or ma l S t a t e m e n t ≈ I n t e n d e d S t a t e m e n t .
因此:
V e r i f i c a t i o n = F o r m a l V a l i d i t y + S e m a n t i c F a i t h f u l n e s s + S c o p e F a i t h f u l n e s s . Verification
=
FormalValidity
+
SemanticFaithfulness
+
ScopeFaithfulness. V er i f i c a t i o n = F or ma l V a l i d i t y + S e man t i c F ai t h f u l n ess + S co p e F ai t h f u l n ess .
其中任何一項可能失敗。
特別是:
L e a n A c c e p t s ( π , Q f ) LeanAccepts(\pi,Q_f) L e an A cce pt s ( π , Q f )
只能推出 Q f Q_f Q f 在相應 formal environment 下具有 kernel-accepted proof。它不能單獨推出:
Q f = Q i n t e n d e d . Q_f
=
Q_{intended}. Q f = Q in t e n d e d .
因此 D V D_V D V 是 CMDC 不可省略的一域。
16. D B D_B D B :Bridge / Existing Mathematics Domain
新生成概念 X X X 若沒有 bridge,至少存在四種風險: X X X 其實是舊概念換名字、是已知概念的特殊情形、是已知理論的簡單重新編碼、或與既有理論矛盾但系統沒有察覺。
因此:
B r i d g e D i a g n o s i s ( X ) BridgeDiagnosis(X) B r i d g eD ia g n os i s ( X )
至少搜尋:
E q u i v a l e n t T o , S p e c i a l C a s e O f , G e n e r a l i z e s , R e d u c e s T o , I s o m o r p h i c T o , I n d e p e n d e n t F r o m , C o n t r a d i c t s . EquivalentTo,
SpecialCaseOf,
Generalizes,
ReducesTo,
IsomorphicTo,
IndependentFrom,
Contradicts. E q u i v a l e n tT o , S p ec ia l C a se O f , G e n er a l i z es , R e d u ces T o , I so m or p hi c T o , I n d e p e n d e n tF r o m , C o n t r a d i c t s .
Bridge gap 不是附加文獻工作,而是數學合法性的一部分。
17. D N D_N D N :Novelty / Interestingness Domain
自動猜想系統可以產生大量真命題。但:
T r u t h ⇏ I n t e r e s t i n g n e s s . Truth
\not\Rightarrow
Interestingness. T r u t h ⇒ I n t er es t in g n ess .
同樣:
N o v e l t y ⇏ I m p o r t a n c e . Novelty
\not\Rightarrow
Importance. N o v e l t y ⇒ I m p or t an ce .
因此系統至少應分離:
C o r r e c t n e s s , N o v e l t y , I n t e r e s t i n g n e s s , U t i l i t y . Correctness,
Novelty,
Interestingness,
Utility. C or r ec t n ess , N o v e l t y , I n t er es t in g n ess , U t i l i t y .
Interestingness 可以由多種信號近似,例如:
C o m p r e s s i o n G a i n , U n i f i c a t i o n P o w e r , S u r p r i s e , D e p e n d e n c y C e n t r a l i t y , P r o o f U t i l i t y , H u m a n I n t e r e s t . CompressionGain,
UnificationPower,
Surprise,
DependencyCentrality,
ProofUtility,
HumanInterest. C o m p r ess i o n G ain , U ni f i c a t i o n P o w er , S u r p r i se , D e p e n d e n cy C e n t r a l i t y , P r oo f U t i l i t y , H u man I n t er es t .
但任何自動 interestingness score 都只能是研究排序器,而不是共同體價值的最終裁決。
18. 一個 gap 的 canonical 結構
本文定義:
δ i = ( I i , D i , T i , E i , S i , C i , B i , R i , D e p i ) . \delta_i
=
(
I_i,
D_i,
T_i,
E_i,
S_i,
C_i,
B_i,
R_i,
Dep_i
). δ i = ( I i , D i , T i , E i , S i , C i , B i , R i , D e p i ) .
其中 I i I_i I i 是 gap identity, D i D_i D i 是 domain, T i T_i T i 是 target, E i E_i E i 是 evidence, S i S_i S i 是 severity, C i C_i C i 是 confidence, B i B_i B i 是 blocking status, R i R_i R i 是 candidate repair classes, D e p i Dep_i D e p i 是 dependency gaps。
例如:
{
"gap_id": "gap:def:017",
"domain": "definition",
"target_ref": "term:density",
"evidence_refs": ["obs:221", "formalizer:error:18"],
"severity": "high",
"confidence": 0.91,
"blocking": true,
"candidate_repairs": [
"clarify_definition",
"map_to_existing_definition",
"split_overloaded_term"
],
"depends_on": []
}
19. Gap 不是永遠互相獨立
實際上:
δ i → δ j \delta_i
\rightarrow
\delta_j δ i → δ j
可能表示修復 δ i \delta_i δ i 後才能判斷 δ j \delta_j δ j 。
例如:
D e f i n i t i o n G a p → F o r m a l i z a t i o n G a p → P r o o f G a p . DefinitionGap
\rightarrow
FormalizationGap
\rightarrow
ProofGap. D e f ini t i o n G a p → F or ma l i z a t i o n G a p → P r oo f G a p .
因此 Gap Map 更合理地表示為:
G Δ = ( V Δ , E Δ ) . G_\Delta
=
(
V_\Delta,
E_\Delta
). G Δ = ( V Δ , E Δ ) .
其中:
V Δ = { δ 1 , … , δ n } . V_\Delta
=
\{
\delta_1,\ldots,\delta_n
\}. V Δ = { δ 1 , … , δ n } .
而:
δ i → δ j \delta_i
\rightarrow
\delta_j δ i → δ j
代表 δ j \delta_j δ j 在診斷或修復上依賴 δ i \delta_i δ i 。
這稱為:
Repair Dependency Graph . \boxed{
\text{Repair Dependency Graph}.
} Repair Dependency Graph .
20. Dominant Gap 與 Secondary Gap
不是每個 gap 都值得立即處理。
定義:
P r i o r i t y ( δ i ) = f ( B l o c k i n g , S e v e r i t y , C o n f i d e n c e , E x p e c t e d G a i n , R e p a i r C o s t , D e p e n d e n c y C e n t r a l i t y ) . Priority(\delta_i)
=
f(
Blocking,
Severity,
Confidence,
ExpectedGain,
RepairCost,
DependencyCentrality
). P r i or i t y ( δ i ) = f ( B l oc k in g , S e v er i t y , C o n f i d e n ce , E x p ec t e d G ain , R e p ai r C os t , D e p e n d e n cy C e n t r a l i t y ) .
因此:
δ ∗ = arg max δ i ∈ Δ t P r i o r i t y ( δ i ) . \delta^*
=
\arg\max_{\delta_i\in\Delta_t}
Priority(\delta_i). δ ∗ = arg δ i ∈ Δ t max P r i or i t y ( δ i ) .
但若:
C o n f i d e n c e ( δ ∗ ) < τ , Confidence(\delta^*)<\tau, C o n f i d e n ce ( δ ∗ ) < τ ,
合理操作可能是:
G a t h e r E v i d e n c e GatherEvidence G a t h er E v i d e n ce
而不是立即 repair。
21. Gap Diagnosis 本身也可以錯
本文不將 diagnosis 視為 oracle。
令真實缺口集合:
Δ ∗ . \Delta^*. Δ ∗ .
AI 估計:
Δ ^ . \hat\Delta. Δ ^ .
則可測:
P r e c i s i o n g a p , R e c a l l g a p , F 1 g a p , Precision_{gap},
Recall_{gap},
F1_{gap}, P r ec i s i o n g a p , R ec a l l g a p , F 1 g a p ,
以及 typed confusion matrix。
例如 DefinitionGap 可能被錯判為 MethodGap。這種誤診很重要,因為它會使 AI 投入錯誤研究方向。
因此:
D i a g n o s i s ≠ T r u t h . \boxed{
Diagnosis
\neq
Truth.
} D ia g n os i s = T r u t h .
22. 從 Gap 到 Repair:CMDC 的核心映射
對 gap:
δ i , \delta_i, δ i ,
CMDC 產生:
R ( δ i ) = { r 1 , … , r k } . \mathcal R(\delta_i)
=
\{
r_1,\ldots,r_k
\}. R ( δ i ) = { r 1 , … , r k } .
其中每個 r j r_j r j 都是 typed repair candidate。
形式:
r j = ( R e p a i r T y p e , T a r g e t , P r e c o n d i t i o n s , M u t a t i o n , E x p e c t e d E f f e c t , R i s k s , O b l i g a t i o n s ) . r_j
=
(
RepairType,
Target,
Preconditions,
Mutation,
ExpectedEffect,
Risks,
Obligations
). r j = ( R e p ai r T y p e , T a r g e t , P r eco n d i t i o n s , M u t a t i o n , E x p ec t e d E f f ec t , R i s k s , O b l i g a t i o n s ) .
因此:
G a p → R e p a i r C a n d i d a t e \boxed{
Gap
\rightarrow
RepairCandidate
} G a p → R e p ai r C an d i d a t e
不是:
G a p → D i r e c t M u t a t i o n . Gap
\rightarrow
DirectMutation. G a p → D i r ec tM u t a t i o n .
23. Typed Repair Taxonomy
第一版 repair class 可以包括:
R P = ProblemClarification , R_P=\text{ProblemClarification}, R P = ProblemClarification ,
R D = DefinitionRepair , R_D=\text{DefinitionRepair}, R D = DefinitionRepair ,
R A = AssumptionRepair , R_A=\text{AssumptionRepair}, R A = AssumptionRepair ,
R J = CriterionRepair , R_J=\text{CriterionRepair}, R J = CriterionRepair ,
R R = RepresentationRepair , R_R=\text{RepresentationRepair}, R R = RepresentationRepair ,
R S = SolutionSpaceRepair , R_S=\text{SolutionSpaceRepair}, R S = SolutionSpaceRepair ,
R M = MethodRepair , R_M=\text{MethodRepair}, R M = MethodRepair ,
R L = Lemma/TheoryRepair , R_L=\text{Lemma/TheoryRepair}, R L = Lemma/TheoryRepair ,
R C = ComputeRepair , R_C=\text{ComputeRepair}, R C = ComputeRepair ,
R E = EvidenceRepair , R_E=\text{EvidenceRepair}, R E = EvidenceRepair ,
R V = VerificationRepair , R_V=\text{VerificationRepair}, R V = VerificationRepair ,
R B = BridgeRepair , R_B=\text{BridgeRepair}, R B = BridgeRepair ,
R N = NoveltyAssessmentRepair . R_N=\text{NoveltyAssessmentRepair}. R N = NoveltyAssessmentRepair .
這裡「repair」不是說原問題一定有錯,而是對當前 research state 的一個結構性修復或補全。
24. Repair Contract
每個 repair 都應具有:
R C ( r ) = ( A l l o w e d M u t a t i o n , F o r b i d d e n M u t a t i o n , R e q u i r e d E v i d e n c e , R e q u i r e d O b l i g a t i o n s , R o l l b a c k P o l i c y , P r o m o t i o n P o l i c y ) . RC(r)
=
(
AllowedMutation,
ForbiddenMutation,
RequiredEvidence,
RequiredObligations,
RollbackPolicy,
PromotionPolicy
). R C ( r ) = ( A l l o w e d M u t a t i o n , F or bi dd e n M u t a t i o n , R e q u i r e d E v i d e n ce , R e q u i r e d O b l i g a t i o n s , R o l l ba c k P o l i cy , P r o m o t i o n P o l i cy ) .
例如 DefinitionRepair 可允許:
- split an overloaded term
- introduce an explicit parameter
- map to a standard mathematical definition
但禁止:
- silently alter theorem scope
- silently strengthen assumptions
- erase original wording
因此 CMDC 是:
constrained generation , \boxed{
\text{constrained generation},
} constrained generation ,
不是自由改寫。
25. Gap Closure Condition
Repair 完成不能只靠 AI 說「我覺得問題修好了」。
對 gap δ i \delta_i δ i 定義:
C l o s e ( δ i ) ⟺ C 1 ∧ C 2 ∧ ⋯ ∧ C m . Close(\delta_i)
\iff
C_1
\land
C_2
\land
\cdots
\land
C_m. C l ose ( δ i ) ⟺ C 1 ∧ C 2 ∧ ⋯ ∧ C m .
例如 DefinitionGap 的 closure conditions 可以是:
D e f i n i t i o n E x p l i c i t ∧ B o u n d a r y H a n d l e d ∧ F o r m a l i z a b l e ∧ P r i o r A r t R e l a t i o n C h e c k e d . DefinitionExplicit
\land
BoundaryHandled
\land
Formalizable
\land
PriorArtRelationChecked. D e f ini t i o n E x pl i c i t ∧ B o u n d a r y H an d l e d ∧ F or ma l i z ab l e ∧ P r i or A r tR e l a t i o n C h ec k e d .
MethodGap 的 closure conditions 可以是:
M e t h o d S p e c i f i e d ∧ A p p l i c a b i l i t y E s t a b l i s h e d ∧ F a i l u r e M o d e K n o w n . MethodSpecified
\land
ApplicabilityEstablished
\land
FailureModeKnown. M e t h o d S p ec i f i e d ∧ A ppl i c abi l i t y E s t ab l i s h e d ∧ F ai l u r e M o d eK n o w n .
因此:
R e p a i r P r o p o s e d ≠ G a p C l o s e d . RepairProposed
\neq
GapClosed. R e p ai r P r o p ose d = G a pC l ose d .
26. Repair-Induced Obligation
任何 repair 都可能製造新的義務。
令:
r : S t M → S t + 1 M . r:
S_t^M
\rightarrow
S_{t+1}^M. r : S t M → S t + 1 M .
則:
O ( r ) = O b l i g a t i o n G e n e r a t o r ( r ) . \mathcal O(r)
=
ObligationGenerator(r). O ( r ) = O b l i g a t i o n G e n er a t or ( r ) .
例如:
A d d A s s u m p t i o n ( A ) AddAssumption(A) A dd A ss u m pt i o n ( A )
產生:
O 1 = declare problem mutation , O_1=\text{declare problem mutation}, O 1 = declare problem mutation ,
O 2 = justify necessity or relevance , O_2=\text{justify necessity or relevance}, O 2 = justify necessity or relevance ,
O 3 = track old and new problem separately . O_3=\text{track old and new problem separately}. O 3 = track old and new problem separately .
RepresentationRepair 產生:
O = semantic equivalence or relationship proof . O=\text{semantic equivalence or relationship proof}. O = semantic equivalence or relationship proof .
MethodRepair 產生:
O = correctness/applicability obligations . O=\text{correctness/applicability obligations}. O = correctness/applicability obligations .
因此:
R e p a i r → N e w O b l i g a t i o n s . \boxed{
Repair
\rightarrow
NewObligations.
} R e p ai r → N e w O b l i g a t i o n s .
27. Repair Cascade
修復一個 gap 可能產生另一個 gap:
δ 1 → r 1 δ 2 . \delta_1
\xrightarrow{r_1}
\delta_2. δ 1 r 1 δ 2 .
例如 DefinitionRepair 引入新 parameter,接著產生 AssumptionGap;或 RepresentationRepair 導致 VerificationGap。
因此 CMDC 不是:
D e t e c t → F i x → D o n e . Detect
\rightarrow
Fix
\rightarrow
Done. D e t ec t → F i x → D o n e .
而是:
D i a g n o s e → R e p a i r → R e D i a g n o s e . Diagnose
\rightarrow
Repair
\rightarrow
ReDiagnose. D ia g n ose → R e p ai r → R eD ia g n ose .
即:
S t M → Δ t → r t → S t + 1 M → Δ t + 1 . \boxed{
S_t^M
\rightarrow
\Delta_t
\rightarrow
r_t
\rightarrow
S_{t+1}^M
\rightarrow
\Delta_{t+1}.
} S t M → Δ t → r t → S t + 1 M → Δ t + 1 .
28. 最小變更原則
在多個 repair 都可解除 gap 時,初步可以採:
r ∗ = arg min r M u t a t i o n C o s t ( r ) r^*
=
\arg\min_r
MutationCost(r) r ∗ = arg r min M u t a t i o n C os t ( r )
subject to:
G a p C l o s e d ( r ) = 1. GapClosed(r)=1. G a pC l ose d ( r ) = 1.
Mutation cost 可包含:
P r o b l e m I d e n t i t y D i s t a n c e , T h e o r y C h a n g e M a g n i t u d e , N e w A s s u m p t i o n s , V e r i f i c a t i o n C o s t . ProblemIdentityDistance,
TheoryChangeMagnitude,
NewAssumptions,
VerificationCost. P r o b l e m I d e n t i t y D i s t an ce , T h eor y C han g e M a g ni t u d e , N e w A ss u m pt i o n s , V er i f i c a t i o n C os t .
這是一種:
Minimal Mathematical Repair Principle . \boxed{
\text{Minimal Mathematical Repair Principle}.
} Minimal Mathematical Repair Principle .
但它不是絕對規則。有時較大的理論重構能帶來更自然、更一般的結果,因此系統應允許 human/contract 指定:
P r e f e r M i n i m a l R e p a i r PreferMinimalRepair P r e f er M inima l R e p ai r
或:
P e r m i t T h e o r y R e f a c t o r . PermitTheoryRefactor. P er mi tT h eor y R e f a c t or .
29. 禁止「捷徑式修復」
CMDC 至少禁止以下模式。
29.1 Silent Assumption Injection
原問題:
Q . Q. Q .
AI 改成:
A ⇒ Q A\Rightarrow Q A ⇒ Q
但仍回報「證明 Q Q Q 」。禁止。
29.2 Silent Restriction
原域:
X . X. X .
AI 只證:
X ′ ⊊ X X'\subsetneq X X ′ ⊊ X
但回報全域成立。禁止。
29.3 Definition Drift
研究中途改變:
D e f i n i t i o n t ( x ) → D e f i n i t i o n t + 1 ( x ) Definition_t(x)
\rightarrow
Definition_{t+1}(x) D e f ini t i o n t ( x ) → D e f ini t i o n t + 1 ( x )
卻不版本化。禁止。
29.4 Formalization Substitution
證成形式命題:
F ( Q ) F(Q) F ( Q )
後直接宣稱自然語言 Q Q Q 已證,而未做 faithfulness check。禁止。
29.5 Rename-as-Novelty
把既有概念換名字後宣稱新理論。禁止。
30. 何時 Repair 應該變成 New Problem
不是所有修改都應被視為原問題修復。
定義 problem distance:
d P ( Q i , Q j ) . d_P(Q_i,Q_j). d P ( Q i , Q j ) .
若:
d P ( Q i , Q j ) > τ P , d_P(Q_i,Q_j)>\tau_P, d P ( Q i , Q j ) > τ P ,
或變換類型為:
N e w P r o b l e m , NewProblem, N e w P r o b l e m ,
則 Q j Q_j Q j 應成為新的 research branch。
因此:
Q 0 → { Q 1 , Q 2 , Q 3 } Q_0
\rightarrow
\{
Q_1,
Q_2,
Q_3
\} Q 0 → { Q 1 , Q 2 , Q 3 }
可以形成 Problem DAG,而不是單一路徑。這使 AI 可以探索新問題,但不能把新問題冒充舊問題。
31. Local Domain Completion
對研究 scope:
Ω , \Omega, Ω ,
令所需構件集合:
R e q ( Ω ) . Req(\Omega). R e q ( Ω ) .
已具備構件:
H a v e t ( Ω ) . Have_t(\Omega). H a v e t ( Ω ) .
則缺失:
M i s s i n g t ( Ω ) = R e q ( Ω ) − H a v e t ( Ω ) . Missing_t(\Omega)
=
Req(\Omega)
-
Have_t(\Omega). M i ss in g t ( Ω ) = R e q ( Ω ) − H a v e t ( Ω ) .
若:
C r i t i c a l M i s s i n g t ( Ω ) = ∅ , CriticalMissing_t(\Omega)
=
\varnothing, C r i t i c a l M i ss in g t ( Ω ) = ∅ ,
且必要 obligations 已解除,則可以說:
L o c a l D o m a i n C o m p l e t e ( Ω , t ) = 1. LocalDomainComplete(\Omega,t)=1. L oc a l D o main C o m pl e t e ( Ω , t ) = 1.
這裡的 complete 表示:在目前研究 scope 與契約下,已具備足夠結構繼續研究或形成 theory package。
它不表示:
all mathematics in this domain is complete . \text{all mathematics in this domain is complete}. all mathematics in this domain is complete .
32. Completion 不是 Closure of Mathematics
本文特別區分:
L o c a l D o m a i n C o m p l e t i o n LocalDomainCompletion L oc a l D o main C o m pl e t i o n
與:
G l o b a l M a t h e m a t i c a l C l o s u r e . GlobalMathematicalClosure. Gl o ba l M a t h e ma t i c a l C l os u r e .
後者不是本文主張。
因此:
L o c a l D o m a i n C o m p l e t i o n ( Ω ) ⇏ N o F u r t h e r M a t h e m a t i c s ( Ω ) . \boxed{
LocalDomainCompletion(\Omega)
\not\Rightarrow
NoFurtherMathematics(\Omega).
} L oc a l D o main C o m pl e t i o n ( Ω ) ⇒ N o F u r t h er M a t h e ma t i cs ( Ω ) .
即使局部 scope 已足夠:
N e w Q u e s t i o n s NewQuestions N e w Q u es t i o n s
仍可無界產生。
33. 生成新詞彙的合法流程
若 gap 為:
T e r m i n o l o g y G a p , TerminologyGap, T er min o l o g y G a p ,
AI 可以生成:
τ n e w . \tau_{new}. τ n e w .
但 promotion 前至少要:
D e f i n e ( τ n e w ) , E x a m p l e s ( τ n e w ) , N o n E x a m p l e s ( τ n e w ) , B o u n d a r y C a s e s ( τ n e w ) , P r i o r A r t S e a r c h ( τ n e w ) , B r i d g e K n o w n C o n c e p t s ( τ n e w ) . Define(\tau_{new}),
Examples(\tau_{new}),
NonExamples(\tau_{new}),
BoundaryCases(\tau_{new}),
PriorArtSearch(\tau_{new}),
BridgeKnownConcepts(\tau_{new}). D e f in e ( τ n e w ) , E x am pl es ( τ n e w ) , N o n E x am pl es ( τ n e w ) , B o u n d a r y C a ses ( τ n e w ) , P r i or A r tS e a r c h ( τ n e w ) , B r i d g eK n o w n C o n ce pt s ( τ n e w ) .
因此:
N a m e C r e a t e d ≠ C o n c e p t C r e a t e d . NameCreated
\neq
ConceptCreated. N am e C r e a t e d = C o n ce ptC r e a t e d .
而:
C o n c e p t C r e a t e d ≠ M a t h e m a t i c a l C o n t r i b u t i o n A c c e p t e d . ConceptCreated
\neq
MathematicalContributionAccepted. C o n ce ptC r e a t e d = M a t h e ma t i c a l C o n t r ib u t i o n A cce pt e d .
34. 生成新方法論的合法流程
若:
M e t h o d G a p MethodGap M e t h o d G a p
成立,AI 可以提出:
M n e w . M_{new}. M n e w .
但應形成:
M e t h o d C a n d i d a t e = ( D o m a i n , I n p u t T y p e , P r e c o n d i t i o n s , T r a n s f o r m a t i o n , O u t p u t T y p e , I n v a r i a n t s , T e r m i n a t i o n , C o r r e c t n e s s C l a i m , C o m p l e x i t y C l a i m , F a i l u r e M o d e s ) . MethodCandidate
=
(
Domain,
InputType,
Preconditions,
Transformation,
OutputType,
Invariants,
Termination,
CorrectnessClaim,
ComplexityClaim,
FailureModes
). M e t h o d C an d i d a t e = ( D o main , I n p u tT y p e , P r eco n d i t i o n s , T r an s f or ma t i o n , O u tp u tT y p e , I n v a r ian t s , T er mina t i o n , C or r ec t n ess C l aim , C o m pl e x i t y C l aim , F ai l u r e M o d es ) .
之後:
T e s t → C o m p a r e → V e r i f y → R e g i s t e r . Test
\rightarrow
Compare
\rightarrow
Verify
\rightarrow
Register. T es t → C o m p a r e → V er i f y → R e g i s t er .
若 correctness 尚未證成,可以保存:
M e t h o d S t a t u s = E x p e r i m e n t a l . MethodStatus=Experimental. M e t h o d S t a t u s = E x p er im e n t a l .
而不是:
M e t h o d S t a t u s = E s t a b l i s h e d . MethodStatus=Established. M e t h o d S t a t u s = E s t ab l i s h e d .
35. 生成新引理與 theory fragment
對 dependency gap:
δ L , \delta_L, δ L ,
系統可以生成候選:
L 1 , … , L k . L_1,\ldots,L_k. L 1 , … , L k .
但候選引理除了可證性外,還應評估:
U t i l i t y ( L i , Q ) , Utility(L_i,Q), U t i l i t y ( L i , Q ) ,
因為:
P r o v a b l e ( L i ) Provable(L_i) P r o v ab l e ( L i )
不代表:
U s e f u l ( L i , Q ) . Useful(L_i,Q). U se f u l ( L i , Q ) .
可使用:
D e p e n d e n c y R e d u c t i o n , P r o o f S e a r c h G a i n , R e u s e P o t e n t i a l DependencyReduction,
ProofSearchGain,
ReusePotential D e p e n d e n cy R e d u c t i o n , P r oo f S e a r c h G ain , R e u se P o t e n t ia l
等信號排序。
36. 反例不是失敗,而是 Repair Trigger
若:
C o u n t e r e x a m p l e ( C ) = x , Counterexample(C)=x, C o u n t er e x am pl e ( C ) = x ,
則:
C = False . C=\text{False}. C = False .
但 autonomous researcher 不應只輸出:
FAIL . \text{FAIL}. FAIL .
它應重新診斷:
W h y F a l s e ( C , x ) ? WhyFalse(C,x)? W h y F a l se ( C , x )?
並可能產生:
R e s t r i c t i o n C a n d i d a t e , A s s u m p t i o n C a n d i d a t e , D e f i n i t i o n R e v i s i o n C a n d i d a t e , N e w C o n j e c t u r e C a n d i d a t e . RestrictionCandidate,
AssumptionCandidate,
DefinitionRevisionCandidate,
NewConjectureCandidate. R es t r i c t i o n C an d i d a t e , A ss u m pt i o n C an d i d a t e , D e f ini t i o n R e v i s i o n C an d i d a t e , N e w C o nj ec t u r e C an d i d a t e .
因此:
C o u n t e r e x a m p l e = R e s e a r c h S t a t e T r a n s i t i o n , \boxed{
Counterexample
=
ResearchStateTransition,
} C o u n t er e x am pl e = R ese a r c h S t a t e T r an s i t i o n ,
而不是終止碼。
37. Proof Repair 與 Domain Repair 的差別
現有 proof-repair 系統通常處理:
E r r o n e o u s P r o o f + C o m p i l e r F e e d b a c k → R e p a i r e d P r o o f . ErroneousProof
+
CompilerFeedback
\rightarrow
RepairedProof. E r r o n eo u s P r oo f + C o m p i l er F ee d ba c k → R e p ai r e d P r oo f .
本文提出的 Domain Repair 更一般:
R e s e a r c h S t a t e + M u l t i D o m a i n E v i d e n c e → T y p e d M a t h e m a t i c a l R e p a i r . ResearchState
+
MultiDomainEvidence
\rightarrow
TypedMathematicalRepair. R ese a r c h S t a t e + M u l t i D o main E v i d e n ce → T y p e d M a t h e ma t i c a l R e p ai r .
Proof repair 可以被視為:
P r o o f R e p a i r ⊂ V e r i f i c a t i o n / M e t h o d R e p a i r . ProofRepair
\subset
Verification/MethodRepair. P r oo f R e p ai r ⊂ V er i f i c a t i o n / M e t h o d R e p ai r .
因此 CMDC 不是取代 proof repair,而是把它放進更大的 research-diagnosis ontology。
38. Autoformalization 與 Definition Gap
真實數學定義的 autoformalization 已被證明比受控 benchmark 更困難。這支持一個重要設計:
F o r m a l i z a t i o n F a i l u r e FormalizationFailure F or ma l i z a t i o n F ai l u r e
不應永遠被判成:
M o d e l F a i l u r e . ModelFailure. M o d e l F ai l u r e .
它也可能表示:
D e f i n i t i o n G a p , L i b r a r y G r o u n d i n g G a p , R e p r e s e n t a t i o n G a p , C o n t e x t G a p . DefinitionGap,
LibraryGroundingGap,
RepresentationGap,
ContextGap. D e f ini t i o n G a p , L ib r a r y G r o u n d in g G a p , R e p r ese n t a t i o n G a p , C o n t e x tG a p .
因此 formalizer error 應成為:
E v i d e n c e ( δ i ) , Evidence(\delta_i), E v i d e n ce ( δ i ) ,
而不是只是一個 retry signal。
39. Theory-Level Autoformalization 與 CMDC
Theory-level autoformalization 強調:
A x i o m s + D e f i n i t i o n s + L e m m a s + D e p e n d e n c i e s Axioms
+
Definitions
+
Lemmas
+
Dependencies A x i o m s + D e f ini t i o n s + L e mma s + D e p e n d e n c i es
必須作為 coherent library 處理。
CMDC 往前多走一步:若這個 theory context 本身有缺口,AI 能否先診斷缺少哪一種構件,再提出受約束的候選補全?
因此:
T h e o r y L e v e l A u t o f o r m a l i z a t i o n TheoryLevelAutoformalization T h eor y L e v e l A u t o f or ma l i z a t i o n
主要處理:
I n f o r m a l T h e o r y → F o r m a l T h e o r y . InformalTheory
\rightarrow
FormalTheory. I n f or ma l T h eor y → F or ma l T h eor y .
而 CMDC 允許:
I n c o m p l e t e R e s e a r c h S t a t e → D i a g n o s e d G a p s → C a n d i d a t e T h e o r y E x t e n s i o n s . IncompleteResearchState
\rightarrow
DiagnosedGaps
\rightarrow
CandidateTheoryExtensions. I n co m pl e t e R ese a r c h S t a t e → D ia g n ose d G a p s → C an d i d a t e T h eor y E x t e n s i o n s .
兩者可以互補。
40. Automated Conjecturing 與 CMDC
Automated conjecturing 一般研究:
D e f i n i t i o n s + A x i o m s → C o n j e c t u r e s . Definitions+Axioms
\rightarrow
Conjectures. D e f ini t i o n s + A x i o m s → C o nj ec t u r es .
CMDC 則先問:
S h o u l d C o n j e c t u r e G e n e r a t i o n B e I n v o k e d ? ShouldConjectureGenerationBeInvoked? S h o u l d C o nj ec t u r e G e n er a t i o n B e I n v o k e d ?
若當前主要缺口是:
D e f i n i t i o n G a p , DefinitionGap, D e f ini t i o n G a p ,
大量 conjecture generation 可能是錯誤行動。
因此:
C o n j e c t u r i n g = O n e R e p a i r C l a s s , \boxed{
Conjecturing
=
OneRepairClass,
} C o nj ec t u r in g = O n e R e p ai r C l a ss ,
而不是所有數學自治的共同預設。
41. ACR 中的數學診斷狀態
一般 Semantic State 可已有:
progress
uncertainty
failures
missing
risks
budget
authority
數學 extension 可以加:
problem_ref
problem_version
theory_refs
definition_refs
assumption_refs
representation_ref
gap_refs
obligation_refs
verification_refs
bridge_refs
novelty_state
因此:
S e m a n t i c S t a t e → M a t h e m a t i c a l S e m a n t i c S t a t e SemanticState
\rightarrow
MathematicalSemanticState S e man t i c S t a t e → M a t h e ma t i c a l S e man t i c S t a t e
可採 extension 而不是重寫。
42. 數學 Affordance Retrieval
給定:
S t M S_t^M S t M
與 Gap Map:
Δ t , \Delta_t, Δ t ,
系統取得:
A M ( S t M , Δ t ) = { Ω 1 , … , Ω k } . \mathcal A_M(S_t^M,\Delta_t)
=
\{
\Omega_1,\ldots,\Omega_k
\}. A M ( S t M , Δ t ) = { Ω 1 , … , Ω k } .
例如:
cog://math/clarify-problem
cog://math/define
cog://math/add-assumption-candidate
cog://math/change-representation
cog://math/generate-lemma
cog://math/search-counterexample
cog://math/retrieve-method
cog://math/generate-method
cog://math/increase-compute
cog://math/verify-faithfulness
cog://math/bridge-theory
因此:
G a p D i a g n o s i s → C o g n i t i v e A f f o r d a n c e . GapDiagnosis
\rightarrow
CognitiveAffordance. G a p D ia g n os i s → C o g ni t i v e A f f or d an ce .
43. Diagnosis-Aware Cognitive Routing
令候選 cognition:
Ω i . \Omega_i. Ω i .
評分可以是:
S c o r e ( Ω i ) = w g G a p F i t + w b B l o c k i n g R e d u c t i o n + w e E x p e c t e d R e s e a r c h G a i n + w h H i s t o r i c a l U t i l i t y − w c C o s t − w r R i s k . Score(\Omega_i)
=
w_g GapFit
+
w_b BlockingReduction
+
w_e ExpectedResearchGain
+
w_h HistoricalUtility
-
w_c Cost
-
w_r Risk. S cor e ( Ω i ) = w g G a pF i t + w b B l oc k in g R e d u c t i o n + w e E x p ec t e d R ese a r c h G ain + w h H i s t or i c a l U t i l i t y − w c C os t − w r R i s k .
而不只是:
S e m a n t i c S i m i l a r i t y . SemanticSimilarity. S e man t i c S imi l a r i t y .
例如:
G a p = D e f i n i t i o n G a p Gap=DefinitionGap G a p = D e f ini t i o n G a p
時:
S c o r e ( D E F I N E ) > S c o r e ( M O R E _ P R O O F _ S E A R C H ) Score(DEFINE)
>
Score(MORE\_PROOF\_SEARCH) S cor e ( D E F I N E ) > S cor e ( M O R E _ P R O O F _ S E A R C H )
應可被 benchmark 驗證。
44. 自主研究中的 DEFER / ESCALATE
若:
C o n f i d e n c e ( D i a g n o s i s ) < τ , Confidence(Diagnosis)<\tau, C o n f i d e n ce ( D ia g n os i s ) < τ ,
AI 可以:
D E F E R DEFER D E F E R
並等待:
M o r e E v i d e n c e . MoreEvidence. M or e E v i d e n ce .
若 repair 需要:
N e w A x i o m NewAxiom N e w A x i o m
或重大 problem mutation,而 contract 不允許自主 promotion,則:
E S C A L A T E . ESCALATE. E S C A L A T E .
這避免:
U n c e r t a i n t y → F a b r i c a t e d C e r t a i n t y . Uncertainty
\rightarrow
FabricatedCertainty. U n cer t ain t y → F ab r i c a t e d C er t ain t y .
45. 自主研究中的 IDLE / STOP
若沒有 cognition 具有正 expected gain,AI 可以:
I D L E . IDLE. I D L E .
若目前 scope 已完成:
L o c a l D o m a i n C o m p l e t e ( Ω ) = 1 , LocalDomainComplete(\Omega)=1, L oc a l D o main C o m pl e t e ( Ω ) = 1 ,
則可:
S T O P . STOP. S T O P .
因此自治不等於:
I n f i n i t e T h e o r y G e n e r a t i o n . InfiniteTheoryGeneration. I n f ini t e T h eor y G e n er a t i o n .
46. Diagnosis Receipt
每次重大 diagnosis 應形成:
D i a g n o s i s R e c e i p t t . DiagnosisReceipt_t. D ia g n os i s R ece i p t t .
至少包含:
( S t a t e R e f , G a p R e f s , E v i d e n c e R e f s , C o n f i d e n c e , A l t e r n a t i v e s , S e l e c t e d G a p , K n o w l e d g e B o u n d a r y R e f ) . (
StateRef,
GapRefs,
EvidenceRefs,
Confidence,
Alternatives,
SelectedGap,
KnowledgeBoundaryRef
). ( S t a t e R e f , G a pR e f s , E v i d e n ce R e f s , C o n f i d e n ce , A l t er na t i v es , S e l ec t e d G a p , K n o w l e d g e B o u n d a r y R e f ) .
它不是 private chain-of-thought,而是當時系統公開採用的 gap classification 與證據。
這讓未來可以審計:AI 為什麼認為問題缺的是定義,而不是方法?
47. Repair Receipt
Repair 決策應保存:
R e p a i r R e c e i p t = ( G a p R e f , R e p a i r C a n d i d a t e R e f , P r o b l e m B e f o r e R e f , P r o b l e m A f t e r R e f , M u t a t i o n T y p e , O b l i g a t i o n R e f s , V a l i d a t i o n R e f s , D e c i s i o n ) . RepairReceipt
=
(
GapRef,
RepairCandidateRef,
ProblemBeforeRef,
ProblemAfterRef,
MutationType,
ObligationRefs,
ValidationRefs,
Decision
). R e p ai r R ece i pt = ( G a pR e f , R e p ai r C an d i d a t e R e f , P r o b l e m B e f or e R e f , P r o b l e m A f t er R e f , M u t a t i o n T y p e , O b l i g a t i o n R e f s , V a l i d a t i o n R e f s , D ec i s i o n ) .
因此:
D i a g n o s i s R e c e i p t ≠ R e p a i r R e c e i p t . DiagnosisReceipt
\neq
RepairReceipt. D ia g n os i s R ece i pt = R e p ai r R ece i pt .
前者回答「缺什麼?」,後者回答「決定怎麼補?」
48. Gap Closure Receipt
當系統宣稱 gap 已關閉:
G a p S t a t u s : O p e n → C l o s e d , GapStatus:
Open
\rightarrow
Closed, G a pS t a t u s : O p e n → C l ose d ,
需要:
G a p C l o s u r e R e c e i p t . GapClosureReceipt. G a pC l os u r e R ece i pt .
至少保存:
C l o s u r e C o n d i t i o n s , E v i d e n c e R e f s , V e r i f i e r R e f s , R e m a i n i n g R i s k s . ClosureConditions,
EvidenceRefs,
VerifierRefs,
RemainingRisks. C l os u r e C o n d i t i o n s , E v i d e n ce R e f s , V er i f i er R e f s , R e mainin g R i s k s .
因此:
C l o s e d Closed C l ose d
成為可審計狀態,而不是 prose assertion。
49. 評估:Diagnosis Accuracy
建立 benchmark:
B d i a g = B P ∪ B D ∪ B A ∪ ⋯ ∪ B N . B_{diag}
=
B_P
\cup
B_D
\cup
B_A
\cup
\cdots
\cup
B_N. B d ia g = B P ∪ B D ∪ B A ∪ ⋯ ∪ B N .
其中每個 case 故意放入已知 gap。
測:
A c c u r a c y d o m a i n , P r e c i s i o n d o m a i n , R e c a l l d o m a i n , C a l i b r a t i o n . Accuracy_{domain},
Precision_{domain},
Recall_{domain},
Calibration. A cc u r a c y d o main , P r ec i s i o n d o main , R ec a l l d o main , C a l ib r a t i o n .
另外需要 multi-label evaluation,因為一個問題可同時具有多個 gap。
50. 評估:Repair Validity
對 repair:
r , r, r ,
定義:
R e p a i r V a l i d i t y ( r ) = G a p R e d u c t i o n − N e w U n d e c l a r e d D a m a g e . RepairValidity(r)
=
GapReduction
-
NewUndeclaredDamage. R e p ai r V a l i d i t y ( r ) = G a pR e d u c t i o n − N e w U n d ec l a r e d D ama g e .
更具體可拆:
R V = ( C l o s u r e S u c c e s s , I d e n t i t y P r e s e r v a t i o n , O b l i g a t i o n C o v e r a g e , N o S i l e n t M u t a t i o n , D o w n s t r e a m G a i n ) . RV
=
(
ClosureSuccess,
IdentityPreservation,
ObligationCoverage,
NoSilentMutation,
DownstreamGain
). R V = ( C l os u r e S u ccess , I d e n t i t y P r eser v a t i o n , O b l i g a t i o n C o v er a g e , N o S i l e n tM u t a t i o n , D o w n s t r e am G ain ) .
因此「修完能證」只是其中一軸。
51. 評估:Downstream Research Gain
最重要的實驗之一是比較:
D i r e c t R e s e a r c h DirectResearch D i r ec tR ese a r c h
與:
D i a g n o s e T h e n R e s e a r c h . DiagnoseThenResearch. D ia g n ose T h e n R ese a r c h .
設:
P e r f o r m a n c e d i r e c t Performance_{direct} P er f or man c e d i r ec t
與:
P e r f o r m a n c e d i a g . Performance_{diag}. P er f or man c e d ia g .
檢驗:
P e r f o r m a n c e d i a g > P e r f o r m a n c e d i r e c t ? Performance_{diag}
>
Performance_{direct}? P er f or man c e d ia g > P er f or man c e d i r ec t ?
但 performance 不只看 final solved rate,也可以看:
T i m e T o C o r r e c t D i a g n o s i s , I n v a l i d M u t a t i o n R a t e , P r o o f S e a r c h W a s t e , F o r m a l i z a t i o n F a i l u r e R a t e . TimeToCorrectDiagnosis,
InvalidMutationRate,
ProofSearchWaste,
FormalizationFailureRate. T im e T o C or r ec t D ia g n os i s , I n v a l i d M u t a t i o n R a t e , P r oo f S e a r c hW a s t e , F or ma l i z a t i o n F ai l u r e R a t e .
52. 評估:Matched Repair vs Random Repair
對同一 gap:
δ , \delta, δ ,
比較:
M a t c h e d R e p a i r ( δ ) MatchedRepair(\delta) M a t c h e d R e p ai r ( δ )
與:
R a n d o m R e p a i r . RandomRepair. R an d o m R e p ai r .
若 CMDC 有意義,應期待:
E x p e c t e d G a i n ( M a t c h e d R e p a i r ) > E x p e c t e d G a i n ( R a n d o m R e p a i r ) . ExpectedGain(
MatchedRepair
)
>
ExpectedGain(
RandomRepair
). E x p ec t e d G ain ( M a t c h e d R e p ai r ) > E x p ec t e d G ain ( R an d o m R e p ai r ) .
53. 評估:Wrong-Domain Intervention
也應故意測:
G a p = D e f i n i t i o n Gap=Definition G a p = D e f ini t i o n
但強制:
A c t i o n = I n c r e a s e C o m p u t e . Action=IncreaseCompute. A c t i o n = I n cr e a se C o m p u t e .
與:
A c t i o n = P r o o f S e a r c h . Action=ProofSearch. A c t i o n = P r oo f S e a r c h .
比較 matched action。這可以回答:域診斷是否真的改變研究效率,而不只是產生漂亮標籤?
54. 評估:Repair Minimality
測:
M u t a t i o n D i s t a n c e ( Q 0 , Q 1 ) MutationDistance(Q_0,Q_1) M u t a t i o n D i s t an ce ( Q 0 , Q 1 )
與 gap closure success。
希望系統避免:
S m a l l G a p → H u g e T h e o r y R e w r i t e . SmallGap
\rightarrow
HugeTheoryRewrite. S ma l l G a p → H ug e T h eor y R e w r i t e .
但不能把 minimality 設成唯一 objective。
因此評估應為 Pareto:
( G a p C l o s u r e , M u t a t i o n C o s t , R e s e a r c h G a i n ) . (
GapClosure,
MutationCost,
ResearchGain
). ( G a pC l os u r e , M u t a t i o n C os t , R ese a r c h G ain ) .
55. 評估:Problem Identity Preservation
對刻意設計的陷阱題,測 AI 是否靜默加假設、靜默限制域、靜默改定義、靜默換形式命題、或將新問題結果回報為原問題結果。
定義:
I d e n t i t y V i o l a t i o n R a t e . IdentityViolationRate. I d e n t i t y V i o l a t i o n R a t e .
理想:
I d e n t i t y V i o l a t i o n R a t e → 0. IdentityViolationRate
\rightarrow
0. I d e n t i t y V i o l a t i o n R a t e → 0.
56. 評估:Obligation Coverage
對每種 repair class,人工建立:
R e q u i r e d O b l i g a t i o n s ( r ) . RequiredObligations(r). R e q u i r e d O b l i g a t i o n s ( r ) .
AI 產生:
G e n e r a t e d O b l i g a t i o n s ( r ) . GeneratedObligations(r). G e n er a t e d O b l i g a t i o n s ( r ) .
測:
C o v e r a g e = ∣ G e n e r a t e d ∩ R e q u i r e d ∣ ∣ R e q u i r e d ∣ . Coverage
=
\frac{
|Generated\cap Required|
}{
|Required|
}. C o v er a g e = ∣ R e q u i r e d ∣ ∣ G e n er a t e d ∩ R e q u i r e d ∣ .
同時測 false obligations,以避免 GenerateEverything 造成表面高 recall。
57. 評估:Long-Horizon Repair Stability
某 repair 在第 t t t 輪看似成功:
G a p C l o s e d t . GapClosed_t. G a pC l ose d t .
但在:
t + k t+k t + k
可能暴露:
H i d d e n C o n t r a d i c t i o n . HiddenContradiction. H i dd e n C o n t r a d i c t i o n .
因此 closure 應允許:
C l o s e d → R e o p e n e d . Closed
\rightarrow
Reopened. C l ose d → R eo p e n e d .
並測:
R e o p e n R a t e , D e l a y e d F a i l u r e D e t e c t i o n . ReopenRate,
DelayedFailureDetection. R eo p e n R a t e , D e l a y e d F ai l u r eD e t ec t i o n .
這比一次性 success 更接近真實研究。
58. Multi-Agent Diagnosis
不同角色可以平行產生:
D i a g n o s i s A , D i a g n o s i s B , D i a g n o s i s C . Diagnosis_A,
Diagnosis_B,
Diagnosis_C. D ia g n os i s A , D ia g n os i s B , D ia g n os i s C .
例如:
F o r m a l i z e r , C o u n t e r e x a m p l e A g e n t , L i t e r a t u r e A g e n t , M e t h o d A g e n t . Formalizer,
CounterexampleAgent,
LiteratureAgent,
MethodAgent. F or ma l i z er , C o u n t er e x am pl e A g e n t , L i t er a t u r e A g e n t , M e t h o d A g e n t .
最後:
J o i n → G a p M a p . Join
\rightarrow
GapMap. J o in → G a pM a p .
但多 Agent 不應被誤認為天然更正確。仍需要:
E v i d e n c e M e r g e , C o n f l i c t R e s o l u t i o n , C a l i b r a t i o n . EvidenceMerge,
ConflictResolution,
Calibration. E v i d e n ce M er g e , C o n f l i c tR eso l u t i o n , C a l ib r a t i o n .
59. 人類的角色
CMDC 不要求人類逐輪提供下一步。但某些 repair 類別可以被 contract 設為:
A p p r o v a l R e q u i r e d . ApprovalRequired. A pp r o v a l R e q u i r e d .
例如:
N e w A x i o m , C a n o n i c a l P r o b l e m M u t a t i o n , N o v e l t y C l a i m , S u b m i s s i o n R e a d y P r o m o t i o n . NewAxiom,
CanonicalProblemMutation,
NoveltyClaim,
SubmissionReadyPromotion. N e w A x i o m , C an o ni c a l P r o b l e m M u t a t i o n , N o v e l t y C l aim , S u bmi ss i o n R e a d y P r o m o t i o n .
因此人類從:
S t e p A u t h o r StepAuthor S t e p A u t h or
逐步轉為:
C o n t r a c t A u t h o r i t y + E x p e r t R e v i e w e r . ContractAuthority
+
ExpertReviewer. C o n t r a c t A u t h or i t y + E x p er tR e v i e w er .
60. 與數學共同體的接口
數學共同體接受的不是「AI 說這是一個完整的新理論」,而是更可審查的:
T h e o r y P a c k a g e = ( O r i g i n a l P r o b l e m , P r o b l e m V e r s i o n s , G a p M a p , D e f i n i t i o n s , A s s u m p t i o n s , R e s u l t s , P r o o f s , C o u n t e r e x a m p l e s , M e t h o d s , B r i d g e s , O p e n O b l i g a t i o n s , P r o v e n a n c e ) . TheoryPackage
=
(
OriginalProblem,
ProblemVersions,
GapMap,
Definitions,
Assumptions,
Results,
Proofs,
Counterexamples,
Methods,
Bridges,
OpenObligations,
Provenance
). T h eor y P a c k a g e = ( O r i g ina l P r o b l e m , P r o b l e mV er s i o n s , G a pM a p , D e f ini t i o n s , A ss u m pt i o n s , R es u l t s , P r oo f s , C o u n t er e x am pl es , M e t h o d s , B r i d g es , O p e n O b l i g a t i o n s , P r o v e nan ce ) .
這使人類可以:
I n s p e c t , R e p r o d u c e , R e j e c t , R e v i s e , E x t e n d . Inspect,
Reproduce,
Reject,
Revise,
Extend. I n s p ec t , R e p r o d u ce , R e j ec t , R e v i se , E x t e n d .
因此:
C o m m u n i t y C o m p a t i b i l i t y = A u d i t a b i l i t y + R e p r o d u c i b i l i t y + E x p l i c i t T h e o r y C h a n g e . \boxed{
CommunityCompatibility
=
Auditability
+
Reproducibility
+
ExplicitTheoryChange.
} C o mm u ni t y C o m p a t ibi l i t y = A u d i t abi l i t y + R e p r o d u c ibi l i t y + E x pl i c i tT h eor y C han g e .
61. 與 Paper 01 的關係
Paper 01 建立總體:
A M R R = A C R + D o m a i n D i a g n o s i s + C M D C + V e r i f i c a t i o n + T h e o r y B r i d g e . AMRR
=
ACR
+
DomainDiagnosis
+
CMDC
+
Verification
+
TheoryBridge. A M R R = A C R + D o main D ia g n os i s + C M D C + V er i f i c a t i o n + T h eor y B r i d g e .
本文只深入:
D o m a i n D i a g n o s i s + C M D C . DomainDiagnosis
+
CMDC. D o main D ia g n os i s + C M D C .
其目標不是重新介紹 AMRR,而是回答兩個問題:
What exactly is missing? \boxed{
\text{What exactly is missing?}
} What exactly is missing?
以及:
What counts as a legitimate repair? \boxed{
\text{What counts as a legitimate repair?}
} What counts as a legitimate repair?
62. 與 Paper 03 的接口
本文故意不把所有 obligation、problem identity 與 theory extension 問題完全展開。下一篇將專門形式化:
P r o b l e m I d e n t i t y , T h e o r y E x t e n s i o n , C o n s e r v a t i v i t y , M a t h e m a t i c a l O b l i g a t i o n s . ProblemIdentity,
TheoryExtension,
Conservativity,
MathematicalObligations. P r o b l e m I d e n t i t y , T h eor y E x t e n s i o n , C o n ser v a t i v i t y , M a t h e ma t i c a l O b l i g a t i o n s .
本文只建立必要接口:
R e p a i r → M a t h C h a n g e S e t → O b l i g a t i o n s . Repair
\rightarrow
MathChangeSet
\rightarrow
Obligations. R e p ai r → M a t h C han g e S e t → O b l i g a t i o n s .
63. 與 Paper 04 的接口
Paper 04 將把本文的 Gap、Repair、Closure、Receipt 變成真正 runtime schema 與 executable loop。
即:
S t M → D i a g n o s e → G a p M a p → R o u t e → R e p a i r → V e r i f y → R e D i a g n o s e . S_t^M
\rightarrow
Diagnose
\rightarrow
GapMap
\rightarrow
Route
\rightarrow
Repair
\rightarrow
Verify
\rightarrow
ReDiagnose. S t M → D ia g n ose → G a pM a p → R o u t e → R e p ai r → V er i f y → R eD ia g n ose .
64. 研究限制
第一,十三域 taxonomy 不是終極分類。
第二,domain diagnosis 的 ground truth 在高階研究問題上可能本身具有爭議。
第三,不同數學領域的 gap structure 可能非常不同。
第四,某些 breakthrough 可能來自看似「錯域」的跨域操作,因此 matched repair 不應變成硬禁止。
第五,interestingness 與 novelty 仍高度依賴人類與歷史脈絡。
第六,局部 repair 可能累積成長期理論漂移,因此必須有 CTCL / ledger 與 Problem Identity 層。
第七,本文仍屬方法論與架構論文,實證 benchmark 留待 AMRR MVP 完成後進行。
65. 可否證命題
本文提出至少六個可直接否證的命題。
H1 — Diagnosis Utility
D i a g n o s e T h e n A c t > D i r e c t A c t DiagnoseThenAct
>
DirectAct D ia g n ose T h e n A c t > D i r ec t A c t
在預先植入 domain gaps 的 benchmark 上成立。
H2 — Matched Repair
M a t c h e d R e p a i r > R a n d o m R e p a i r . MatchedRepair
>
RandomRepair. M a t c h e d R e p ai r > R an d o m R e p ai r .
H3 — Wrong-Domain Cost
C o s t ( W r o n g D o m a i n A c t i o n ) > C o s t ( M a t c h e d D o m a i n A c t i o n ) . Cost(WrongDomainAction)
>
Cost(MatchedDomainAction). C os t ( W r o n g D o main A c t i o n ) > C os t ( M a t c h e d D o main A c t i o n ) .
H4 — Identity Preservation
加入 CMDC / Problem Identity gate 後:
S i l e n t M u t a t i o n R a t e ↓ . SilentMutationRate
\downarrow. S i l e n tM u t a t i o n R a t e ↓ .
H5 — Obligation Coverage
typed repair 能提高:
R e q u i r e d O b l i g a t i o n R e c a l l . RequiredObligationRecall. R e q u i r e d O b l i g a t i o n R ec a l l .
H6 — Long-Horizon Stability
Re-diagnosis loop 相較一次性 repair:
D e l a y e d F a i l u r e D e t e c t i o n ↑ . DelayedFailureDetection
\uparrow. D e l a y e d F ai l u r eD e t ec t i o n ↑ .
若這些命題大多不成立,CMDC 作為獨立方法論的價值就必須重新評估。
66. 核心不變量
本文最後固定十個不變量。
U n s o l v e d ≠ S i n g l e F a i l u r e M o d e . \boxed{
Unsolved
\neq
SingleFailureMode.
} U n so l v e d = S in g l e F ai l u r e M o d e .
D i a g n o s i s ≠ T r u t h . \boxed{
Diagnosis
\neq
Truth.
} D ia g n os i s = T r u t h .
G a p ≠ R e p a i r . \boxed{
Gap
\neq
Repair.
} G a p = R e p ai r .
R e p a i r C a n d i d a t e ≠ A c c e p t e d M u t a t i o n . \boxed{
RepairCandidate
\neq
AcceptedMutation.
} R e p ai r C an d i d a t e = A cce pt e d M u t a t i o n .
R e p a i r P r o p o s e d ≠ G a p C l o s e d . \boxed{
RepairProposed
\neq
GapClosed.
} R e p ai r P r o p ose d = G a pC l ose d .
R e p a i r → O b l i g a t i o n s . \boxed{
Repair
\rightarrow
Obligations.
} R e p ai r → O b l i g a t i o n s .
M o r e C o m p u t e ⇏ B e t t e r P r o b l e m S t a t e . \boxed{
MoreCompute
\not\Rightarrow
BetterProblemState.
} M or e C o m p u t e ⇒ B e tt er P r o b l e m S t a t e .
P r o o f S u c c e s s ⇏ S e m a n t i c F a i t h f u l n e s s . \boxed{
ProofSuccess
\not\Rightarrow
SemanticFaithfulness.
} P r oo f S u ccess ⇒ S e man t i c F ai t h f u l n ess .
L o c a l D o m a i n C o m p l e t i o n ⇏ G l o b a l M a t h e m a t i c a l C l o s u r e . \boxed{
LocalDomainCompletion
\not\Rightarrow
GlobalMathematicalClosure.
} L oc a l D o main C o m pl e t i o n ⇒ Gl o ba l M a t h e ma t i c a l C l os u r e .
G e n e r a t i o n F r e e d o m ↑ ⟹ V e r i f i c a t i o n O b l i g a t i o n ↑ . \boxed{
GenerationFreedom\uparrow
\Longrightarrow
VerificationObligation\uparrow.
} G e n er a t i o n F r ee d o m ↑⟹ V er i f i c a t i o n O b l i g a t i o n ↑ .
67. 最終模型
完整 CMDC loop:
S t M → O b s e r v e O t → D i a g n o s e Δ t → P r i o r i t i z e δ t ∗ → R e t r i e v e / G e n e r a t e R t → G o v e r n r t ∗ → A p p l y C a n d i d a t e S t + 1 M , c a n d i d a t e → G e n e r a t e O b l i g a t i o n s O t → V e r i f y V t → A c c e p t / R e j e c t S t + 1 M → R e D i a g n o s e Δ t + 1 . \boxed{
\begin{aligned}
S_t^M
&\xrightarrow{Observe}
O_t\\
&\xrightarrow{Diagnose}
\Delta_t\\
&\xrightarrow{Prioritize}
\delta_t^*\\
&\xrightarrow{Retrieve/Generate}
\mathcal R_t\\
&\xrightarrow{Govern}
r_t^*\\
&\xrightarrow{ApplyCandidate}
S_{t+1}^{M,candidate}\\
&\xrightarrow{GenerateObligations}
\mathcal O_t\\
&\xrightarrow{Verify}
V_t\\
&\xrightarrow{Accept/Reject}
S_{t+1}^M\\
&\xrightarrow{ReDiagnose}
\Delta_{t+1}.
\end{aligned}
} S t M O b ser v e O t D ia g n ose Δ t P r i or i t i z e δ t ∗ R e t r i e v e / G e n er a t e R t G o v er n r t ∗ A ppl y C an d i d a t e S t + 1 M , c an d i d a t e G e n er a t e O b l i g a t i o n s O t V er i f y V t A cce pt / R e j ec t S t + 1 M R eD ia g n ose Δ t + 1 .
其中任何階段都可以:
R E F U S E , D E F E R , S T O P , I D L E , E S C A L A T E . REFUSE,
DEFER,
STOP,
IDLE,
ESCALATE. R E F U S E , D E F E R , S T O P , I D L E , E S C A L A T E .
68. 核心命題
本文將整篇壓縮為:
An autonomous mathematical system should not interpret every research failure as a request for more proof search. \boxed{
\textbf{
An autonomous mathematical system should not interpret every research failure as a request for more proof search.
}
} An autonomous mathematical system should not interpret every research failure as a request for more proof search.
更完整地:
It should diagnose which mathematical domain is incomplete, generate only typed candidate repairs, preserve the identity of the original problem, and require explicit evidence before declaring a gap closed. \boxed{
\textbf{
It should diagnose which mathematical domain is incomplete,
generate only typed candidate repairs,
preserve the identity of the original problem,
and require explicit evidence before declaring a gap closed.
}
} It should diagnose which mathematical domain is incomplete, generate only typed candidate repairs, preserve the identity of the original problem, and require explicit evidence before declaring a gap closed.
中文:
自主數學 AI 面對「解不出來」時,不應只有繼續算、繼續證或重新生成答案三種反應。它應先判定缺的是問題、定義、假設、判定、表示、方法、引理、算力、反例、驗證、橋接或新穎性中的哪一部分,再提出與該缺口相符的候選補全;而任何補全只有在問題身份、義務與驗證條件被保留後,才有資格成為新的數學研究狀態。
69. 結論
數學研究不是固定問題上的單一證明搜尋。
更一般地,它是一個:
problem state + theory state + gap state + repair state \boxed{
\text{problem state}
+
\text{theory state}
+
\text{gap state}
+
\text{repair state}
} problem state + theory state + gap state + repair state
共同演化的過程。
本文因此把:
S o l v e d / U n s o l v e d Solved/Unsolved S o l v e d / U n so l v e d
二分法展開為:
M a t h e m a t i c a l D o m a i n D i a g n o s i s . MathematicalDomainDiagnosis. M a t h e ma t i c a l D o main D ia g n os i s .
再把:
T r y A g a i n TryAgain T r y A g ain
展開為:
C o n s t r a i n e d M a t h e m a t i c a l D o m a i n C o m p l e t i o n . ConstrainedMathematicalDomainCompletion. C o n s t r ain e d M a t h e ma t i c a l D o main C o m pl e t i o n .
真正重要的轉變是:
AI no longer merely searches inside a given mathematical space; it can diagnose and cautiously repair the space required for research itself. \boxed{
\text{AI no longer merely searches inside a given mathematical space; it can diagnose and cautiously repair the space required for research itself.}
} AI no longer merely searches inside a given mathematical space; it can diagnose and cautiously repair the space required for research itself.
但「修復空間」並不代表 AI 可以自由重寫數學。
相反,CMDC 的基本交換是:
more generative freedom ⟺ more explicit mathematical obligations . \boxed{
\text{more generative freedom}
\Longleftrightarrow
\text{more explicit mathematical obligations}.
} more generative freedom ⟺ more explicit mathematical obligations .
因此,下一篇必須正式處理:
P r o b l e m I d e n t i t y , T h e o r y E x t e n s i o n , C o n s e r v a t i v i t y , M a t h e m a t i c a l O b l i g a t i o n . ProblemIdentity,
TheoryExtension,
Conservativity,
MathematicalObligation. P r o b l e m I d e n t i t y , T h eor y E x t e n s i o n , C o n ser v a t i v i t y , M a t h e ma t i c a l O b l i g a t i o n .
也就是回答:當 AI 可以自行改定義、補假設、建新方法甚至提出新公理時,究竟什麼條件下仍可說它在研究原問題,什麼時候已經建立另一個理論?
這將構成系列 Paper 03。
參考文獻
[1] Feng, T., Trinh, T. H., Bingham, G., et al. (2026). Towards Autonomous Mathematics Research . arXiv:2602.10177.
[2] Tsoukalas, G., Kovsharov, A., Shirobokov, S., et al. (2026). Advancing Mathematics Research with AI-Driven Formal Proof Search . arXiv:2605.22763.
[3] Min, M. J., He, M., Li, Z., Yi, Z., Malik, S., Gupta, A., Si, X., & Bastani, O. (2026). Theory-Level Autoformalization: From Isolated Statements to Unified Formal Knowledge Bases . arXiv:2607.13292. ICML 2026 Position Track, Spotlight.
[4] Zhang, L., Valentino, M., & Freitas, A. (2025). Autoformalization in the Wild: Assessing LLMs on Real-World Mathematical Definitions . Proceedings of EMNLP 2025, 1720–1738. DOI: 10.18653/v1/2025.emnlp-main.90.
[5] Zhang, J., & Tan, S.-C. (2026). Automated Conjecturing and Theorem Finding: A Survey . Journal of Computer Science and Technology, 41(1), 46–66. DOI: 10.1007/s11390-026-6040-0.
[6] Tsoukalas, G., Saha, R., Thakur, A., Reguyal, S., & Chaudhuri, S. (2025). Learning Interestingness in Automated Mathematical Theory Formation . Advances in Neural Information Processing Systems 38.
[7] Wang, E., Chess, S., Lee, D., Ge, S., Mallavarapu, A., Alper, J., & Ilin, V. (2026). Learning to Repair Lean Proofs from Compiler Feedback . arXiv:2602.02990.
[8] Ospanov, A., & Yousefzadeh, R. (2025). APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning . arXiv:2505.05758.
[9] Poiroux, A., Weiss, G., Kunčak, V., & Bosselut, A. (2025). Reliable Evaluation and Benchmarks for Statement Autoformalization . Proceedings of EMNLP 2025, 17947–17969. DOI: 10.18653/v1/2025.emnlp-main.907.
內部架構依賴文件
[I1] Neo.K. (2026). 從數學解題到自主數學研究:受約束數學域補全與自主數學研究 Runtime ,Autonomous Mathematical Research / Paper 01 of 04,v0.1.
[I2] Neo.K. (2026). 從自提示到自主認知閉環:持續目標型 AI 的基礎理論 ,系列 01/06,v0.1.
[I3] Neo.K. (2026). 可定址認知空間:Cognitive Affordance、Semantic Address 與認知算子 ,系列 02/06,v0.1.
[I4] Neo.K. (2026). 自我對話不是文字:AI-Native Cognitive Program 與 Zero-Rendering Runtime ,系列 03/06,v0.1.
[I5] Neo.K. (2026). 時間因果自我史:CTCL-ITR、Decision Receipt 與上下文壓縮後的可追溯性 ,系列 04/06,v0.1.
[I6] Neo.K. (2026). 契約邊界內的 AI 自主性:Execute、Refuse、Defer、Idle 與 Escalate ,系列 05/06,v0.1.
[I7] Neo.K. (2026). Addressable Cognitive Runtime × CTCL:統一技術白皮書與實作路線圖 v0.1 ,系列 06/06.
版本備註
v0.1 / 2026-08-23
本版正式固定:
Multi-Domain Mathematical Diagnosis;
Mathematical Research Domain Set v1;
Mathematical Domain Gap Map;
Repair Dependency Graph;
Typed Repair Taxonomy;
Repair Contract;
Gap Closure Condition;
Repair-Induced Obligation;
Repair Cascade;
Minimal Mathematical Repair Principle;
Local Domain Completion;
Diagnosis / Repair / Gap Closure Receipts;
diagnosis-aware cognitive routing;
CMDC falsification gates;
與 Paper 03 / Paper 04 的接口。
本版不宣稱十三域分類已是終極 taxonomy,也不宣稱 CMDC 已經由大規模 empirical benchmark 驗證。所有效能優勢均留待後續 AMRR MVP 與實驗論文測試。