自主數學研究 Runtime:從可定址認知到自主理論建構
Autonomous Mathematical Research Runtime: From Addressable Cognition to Self-Directed Theory Construction
系列:Autonomous Mathematical Research / Paper 04 of 04
版本:v0.1
日期:2026-08-23
作者:Neo.K
摘要
前三篇依序建立了自主數學研究的統合理論、多域問題診斷與受約束數學域補全,以及問題身份、理論擴張與數學義務框架。然而,若這些概念只停留在論文語義層,它們仍不足以形成真正可運行的自主數學研究 AI。本文提出 Autonomous Mathematical Research Runtime, AMRR(自主數學研究 Runtime) 的第一版可執行架構,目標是把數學研究中的問題、理論、缺口、修復、義務、證明、反例、方法、橋接與新穎性狀態轉換為 canonical runtime objects,並將它們接入既有 Addressable Cognitive Runtime(ACR)與 CTCL-ITR 時間因果證據層。
本文不重寫 ACR。現有 ACR 已具有 canonical schema、可定址 CognitiveObject、Semantic State Encoder、Cognitive Affordance、Cognitive Program、Governance、Decision Receipt 與 Persistent Loop 的架構邊界;AMRR 在此基礎上加入數學 domain extension。本文提出九個主要數學模組:Mathematical State Adapter、Problem Store、Theory Store、Gap Diagnosis Engine、Constrained Repair Engine、Mathematical Obligation Engine、Mathematical Cognitive Registry、Verifier Router / Verifier Ensemble、Theory Bridge & Novelty Layer,並以 Research Runtime Coordinator 將它們編排為持續研究閉環。
AMRR 的 canonical state 不再只保存「目前答案」,而保存 、 、 、 、 、 、 、 、 與 。任何 AI 生成的定義、假設、lemma、方法、猜想或公理都先進入 candidate state;只有經過 Problem Identity、Theory Extension、Obligation、Verification 與 Governance gates 後,才能 promotion 到 accepted research state。本文同時定義 cog://math/* 認知算子 namespace、typed verifier selection、problem/theory DAG、數學 receipts、CTCL event namespace、研究契約、budget model、failure model、multi-agent join 與 context compression 邊界。
本文最後提出一條由現有 ACR Phase 2 向 AMRR MVP 演進的實作路線:保留 Phase 0–2 的 frozen interfaces,先新增 Math Extension Schemas 與 deterministic Mathematical State Adapter,再依序建立 Gap Diagnosis、Problem Identity / Theory Store、Obligation Engine、Math Cognitive Registry、Verifier Router、CMDC Repair Engine、Theory Bridge、CTCL Adapter 與 Persistent Research Loop。每一層皆遵守:
而不是以「AI 看起來很會研究」作為完成標準。
本文的核心命題是:自主數學研究 AI 不應被實作為一個無界循環的文字 Agent,而應被實作為一個以 canonical mathematical state、typed cognitive programs、受約束 theory mutation、外部 verifier 與可驗證時間因果歷史共同構成的 persistent research runtime。
關鍵詞: Autonomous Mathematical Research Runtime、AMRR、Addressable Cognitive Runtime、CMDC、數學狀態、問題身份、理論擴張、數學義務、形式驗證、Theory Bridge、CTCL-ITR
1. 問題:前三篇已經有方法論,但還沒有 Runtime
前三篇可以壓縮成:
這仍然是一條概念鏈。
真正工程問題是:
哪些東西是 canonical object?誰保存它?誰可以改?誰驗證?哪些狀態可以 promotion?錯誤如何回退?上下文被壓縮後如何恢復?
因此本文研究:
2. AMRR 的設計原則
AMRR v0.1 固定十個工程原則。
第一:
第二:
第三:
第四:
第五:
第六:
第七:
第八:
第九:
第十:
3. Implemented Baseline 與 Proposed Extension 必須分開
AMRR v0.1 不把尚未實作的模組描述成既有功能。
目前 ACR 工程基線已具有:
Phase 0: canonical schema freeze
Phase 1: SPRC / CIO adapter and addressable CognitiveObjects
Phase 2: deterministic public Semantic State Encoder
Phase 2 artifact 的 release validation 記錄包含 10 個 Phase-0 schemas、64 個 Phase-1 CognitiveObjects、deterministic SemanticState fingerprint 與完整 release suite 的 73 passed。
本文新增的:
Mathematical State Adapter
Problem Store
Theory Store
Gap Diagnosis Engine
Repair Engine
Obligation Engine
Math Cognitive Registry
Verifier Router
Theory Bridge Layer
Persistent Research Loop
均屬 AMRR Proposed Extension。
4. 總體架構
AMRR 可以表示為:
邏輯上分為五層:
Layer A — Mathematical State & Knowledge
Layer B — Cognitive Research Runtime
Layer C — Mutation / Obligation / Verification
Layer D — Governance & Research Contract
Layer E — Temporal-Causal Evidence & Audit
5. Layer A — Mathematical State & Knowledge
Layer A 回答:
現在的數學世界狀態是什麼?
核心物件:
MathematicalState
ProblemRecord
ProblemVersion
TheoryRecord
TheoryVersion
DefinitionRecord
AssumptionRecord
ClaimRecord
MethodRecord
GapMap
ObligationGraph
BridgeGraph
這一層不負責自由生成研究方案。
它負責:
6. Layer B — Cognitive Research Runtime
Layer B 回答:
現在應該怎麼研究?
核心:
Mathematical Affordance Retriever
Math Cognitive Router
Research Program Compiler
Research Program Executor
Re-observer
Agenda Runtime
它使用:
取得:
再編譯:
7. Layer C — Mutation / Obligation / Verification
Layer C 回答:
這個新東西能不能合法進入 theory state?
核心:
Gap Diagnosis Engine
Constrained Repair Engine
Problem Identity Resolver
Theory Extension Classifier
Mathematical Obligation Engine
Verifier Router
Verifier Adapters
Promotion Engine
其主要輸出不是 prose。
而是:
8. Layer D — Governance & Research Contract
Layer D 回答:
AI 是否被允許自行採用、公開或升格這項數學變更?
沿用:
數學域映射成:
9. Layer E — Temporal-Causal Evidence & Audit
Layer E 回答:
這個問題何時被改了?為什麼改?當時知道什麼?哪些 proof / verifier / assumption 支撐現在的 claim?
核心:
CTCL instant
TemporalEvent
DecisionReceipt
ProblemMutationReceipt
TheoryExtensionReceipt
DiagnosisReceipt
RepairReceipt
ObligationReceipt
VerificationReceipt
GapClosureReceipt
ResearchReceipt
CommitReceipt
因此:
10. Mathematical State 不應取代原 Semantic State
ACR Phase 2 已經有:
AMRR 不應建立另一套互不相容的根狀態。
而應:
因此通用欄位繼續使用:
progress
uncertainty
failures
missing
risks
budget
authority_ref
goal_refs
commitments
environment_refs
memory_refs
數學欄位進:
extensions.math
11. extensions.math v0.1
第一版可以定義:
{
"problem_refs": [],
"active_problem_ref": null,
"theory_refs": [],
"active_theory_ref": null,
"definition_refs": [],
"assumption_refs": [],
"method_refs": [],
"claim_refs": [],
"gap_map_ref": null,
"open_obligation_refs": [],
"verification_refs": [],
"bridge_refs": [],
"novelty_state": "unknown",
"research_phase": "diagnosis"
}
這個 extension 必須 canonical serialize。
12. Mathematical State Identity
狀態 identity 不用自然語言摘要。
沿用 canonical fingerprint:
因此:
這保持 Phase 2 deterministic boundary。
13. Problem Store
Problem Store 保存:
最小 schema:
{
"problem_id": "problem:...",
"version": 3,
"parent_refs": [],
"canonical_statement_ref": "artifact:...",
"informal_statement_ref": "artifact:...",
"formal_statement_refs": [],
"domain_refs": [],
"success_criteria": [],
"status": "active",
"created_from_gap_ref": null,
"relation_to_parent": null,
"assumption_envelope_ref": "assumption-envelope:..."
}
14. Problem Version Graph
問題不能只靠:
problem:v7
還需要:
edge:
其中:
這使:
都可以共存於 Problem DAG。
15. Problem Relation Resolver
接口:
problem.relate(before_ref, after_ref)
輸出:
{
"relation": "restriction",
"confidence": 0.93,
"evidence_refs": [],
"verification_status": "candidate"
}
relation 也不是 oracle。
所以:
16. Theory Store
Theory Store 保存:
最小:
{
"theory_id": "theory:...",
"version": 4,
"parent_refs": [],
"signature_refs": [],
"axiom_refs": [],
"definition_refs": [],
"theorem_refs": [],
"method_refs": [],
"bridge_refs": [],
"open_obligation_refs": [],
"extension_profile_ref": null
}
17. Theory State 不等於 Document
保持:
論文、README、自然語言摘要只是 renderer / artifact。
Canonical theory state 應可被機器操作。
18. Theory Extension Classifier
接口:
theory.classify_extension(before, after)
輸出:
例如:
{
"language_change": true,
"semantic_change": false,
"proof_strength_change": false,
"model_change": false,
"foundation_change": false,
"class": "definitional_extension"
}
19. Gap Diagnosis Engine
輸入:
輸出:
第一版不要直接全交給 LLM。
可以採:
20. Gap Diagnosis 的資料源
允許:
formalizer errors
proof assistant errors
counterexamples
failed method history
missing library symbol
ambiguous definitions
problem relation evidence
verification receipts
literature / prior-art refs
compute budget exhaustion
human review annotations
禁止把 private chain-of-thought 當 canonical evidence。
21. Gap Map Schema
{
"gap_map_id": "gap-map:...",
"state_ref": "state:...",
"gaps": [
{
"gap_id": "gap:...",
"domain": "definition",
"target_ref": "definition:...",
"evidence_refs": [],
"confidence": 0.91,
"severity": "high",
"blocking": true,
"depends_on": [],
"candidate_repair_classes": []
}
]
}
22. Gap Map 是 DAG,不只是 list
因:
因此:
這可以直接利用 CTCL / topology style 的 multi-parent graph semantics。
23. Gap Prioritizer
第一版不應將其壓成不可解釋 reward。
可以保留 vector:
24. Constrained Repair Engine
輸入:
輸出:
每個 repair:
25. Repair 不能直接 Commit
保持三層:
因此:
26. MathChangeSet
{
"change_set_id": "math-change:...",
"source_state_ref": "state:...",
"gap_ref": "gap:...",
"change_type": "add_assumption",
"target_ref": "problem:...",
"before_ref": "problem:v3",
"after_candidate_ref": "problem:v4-candidate",
"reason_codes": [],
"obligation_refs": [],
"verification_status": "pending",
"promotion_status": "candidate"
}
27. Mathematical Obligation Engine
接口:
obligation.generate(change_set)
輸出:
第一版採 rule-first:
add_definition
→ well_formedness
→ semantic_faithfulness
→ bridge_check
→ conservativity_if_claimed
add_assumption
→ problem_identity
→ scope_disclosure
→ assumption_envelope_update
new_axiom
→ explicit_nonconservative_status
→ consistency/model obligations
→ escalation policy
28. Obligation Graph
義務之間也有 dependency:
例如:
因此:
29. Obligation Schema
{
"obligation_id": "obligation:...",
"class": "semantic_faithfulness",
"trigger_ref": "math-change:...",
"claim_ref": null,
"required_evidence": [],
"verifier_classes": ["formal", "semantic", "human"],
"depends_on": [],
"status": "OPEN",
"waiver_policy": "escalation_required"
}
30. Obligation Status Machine
也可以:
或:
但:
31. Mathematical Cognitive Registry
AMRR 不應每輪重新用自然語言發明「現在該怎麼研究」。
建立:
cog://math/interpret@1
cog://math/formalize@1
cog://math/diagnose-gap@1
cog://math/clarify-problem@1
cog://math/define@1
cog://math/add-assumption-candidate@1
cog://math/change-representation@1
cog://math/generate-lemma@1
cog://math/generate-conjecture@1
cog://math/search-counterexample@1
cog://math/retrieve-method@1
cog://math/generate-method@1
cog://math/prove@1
cog://math/disprove@1
cog://math/check-conservativity@1
cog://math/verify-faithfulness@1
cog://math/bridge-theory@1
cog://math/check-prior-art@1
cog://math/discharge-obligation@1
cog://math/package-theory@1
32. Math Cognitive Object
每個 math operator 沿用 CognitiveObject:
再加入 domain-specific metadata:
{
"math_domains": ["definition", "verification"],
"accepted_gap_classes": [],
"produces_candidate_types": [],
"obligation_effects": [],
"required_verifier_classes": []
}
33. Math Affordance Retrieval
輸入:
輸出:
評分:
34. Research Program Compiler
例如:
{
"program_id": "math-program:...",
"steps": [
{"op": "cog://math/search-counterexample@1"},
{"op": "cog://math/clarify-problem@1"},
{"op": "cog://math/formalize@1"},
{"op": "cog://math/prove@1"}
],
"stop_if": [
"counterexample_found",
"problem_identity_ambiguous",
"budget_exhausted"
]
}
35. Program Validation
執行前:
若失敗:
36. Verifier Router
不是每個 obligation 都丟給 Lean。
定義:
37. Verifier Ensemble
第一版 class:
formal_proof
symbolic_algebra
sat
smt
numeric
exhaustive_search
simulation
semantic_equivalence
literature_prior_art
human_review
因此:
38. Verifier Adapter Interface
統一接口:
verify(request) -> VerificationResult
request 至少帶:
claim_ref
theory_ref
problem_ref
obligation_ref
assumption_envelope_ref
budget
結果:
PASS
FAIL
INCONCLUSIVE
ERROR
而:
39. Verification Receipt
{
"verification_receipt_id": "verify:...",
"claim_ref": "claim:...",
"obligation_ref": "obligation:...",
"verifier": {
"class": "formal_proof",
"implementation": "lean"
},
"theory_ref": "theory:v4",
"assumption_envelope_ref": "ae:...",
"result": "PASS",
"artifact_refs": [],
"executed_at": "...",
"knowledge_boundary_ref": "kb:..."
}
40. Proof Result 不能自動成為 Claim
保持:
因此 promotion pipeline:
41. Assumption Envelope Service
接口:
claim.assumption_envelope(claim_ref)
計算:
它必須跨:
claim
lemma
definition
method
theory bridge
formalization
追蹤依賴。
42. No-Laundering Gate
公開 claim 前:
以及:
若不成立:
43. Theory Bridge Layer
Theory Bridge 回答:
新定義、新理論與既有數學到底如何連接?
Bridge object:
relation:
equivalent_to
generalizes
specializes
reduces_to
interprets
isomorphic_to
independent_from
uses
contradicts
44. Bridge Strength
下游使用 bridge 時必須知道 strength。
因此:
45. Prior-Art / Novelty Layer
Novelty 不應是 boolean。
第一版:
unknown
search_incomplete
known_prior_art
rediscovered
novel_candidate
human_review_required
因此:
而不是:
46. Discovery Receipt
{
"discovery_receipt_id": "discovery:...",
"contribution_ref": "claim:...",
"knowledge_boundary_ref": "kb:...",
"search_scope": [],
"prior_art_refs": [],
"novelty_status": "novel_candidate",
"review_status": "pending"
}
47. Research Contract
AMRR 的持續輸入不是一個 prompt。
而是:
第一版 contract:
48. Research Contract 範例
Goal:
Investigate the target conjecture and produce a reviewable theory package.
Allowed:
- generate definitions, lemmas, conjectures and methods as candidates
- run formal / symbolic / numerical verifiers
- change representations
- search prior art
- create problem branches
Approval required:
- replace the canonical problem
- promote a new axiom
- declare novelty
- mark package submission-ready
Denied:
- erase prior problem versions
- report restricted results as unrestricted
- hide assumptions
- fabricate citations
49. Research Agenda Runtime
輸入:
輸出:
例如:
investigate_definition_gap
search_counterexample
formalize_candidate_lemma
discharge_faithfulness_obligation
check_prior_art
reopen_closed_gap
50. Agenda 不能直接執行
保持:
因此:
51. Persistent Research Loop
核心偽代碼:
while runtime_active:
observation = observe_research_environment()
state = encode_mathematical_state(observation)
gaps = diagnose(state)
obligations = refresh_open_obligations(state)
agenda_candidates = propose_research_agenda(
goal,
state,
gaps,
obligations,
contract
)
if no_positive_candidate(agenda_candidates):
enter_idle_or_stop()
continue
for agenda in agenda_candidates:
agenda_decision = govern_agenda(agenda, contract)
if agenda_decision != EXECUTE:
record_nonexecute_decision()
continue
affordances = retrieve_math_affordances(state, gaps)
program = compile_math_program(state, agenda, affordances)
validate_program(program)
result = execute_math_program(program)
if result.proposes_mutation:
change = create_math_change_set(result)
relation = classify_problem_relation(change)
extension = classify_theory_extension(change)
new_obligations = generate_obligations(change)
mutation_decision = govern_mutation(
change,
relation,
extension,
new_obligations,
contract
)
if mutation_decision == EXECUTE:
candidate_state = apply_candidate(change)
verification = route_and_run_verifiers(candidate_state)
promotion = evaluate_promotion(candidate_state, verification)
commit_or_reject(promotion)
write_receipts()
emit_ctcl_events()
reobserve()
rediagnose()
compact_context_if_needed()
52. Re-observation 是研究核心
一次 repair 後不能假設問題就好了。
必須:
因此:
53. Gap Reopen
已關閉 gap 可以:
觸發條件:
new counterexample
new proof failure
new prior art
new semantic mismatch
new theory contradiction
new human review
這是長期研究必要能力。
54. Multi-Agent / Parallel Research
允許:
最後:
但 canonical state 仍由 shared stores / receipts 決定。
不能讓每個 agent 各自擁有互相矛盾的「正式問題」。
55. Multi-Agent Conflict
若:
不能直接投票了事。
應生成:
並比較:
56. Model Roles
AMRR 邏輯角色:
Researcher
Diagnostician
Formalizer
Conjecturer
CounterexampleSearcher
MethodDesigner
VerifierController
BridgeAnalyst
Governor
Auditor
同一 model 可以扮演多角色。
保持:
57. Model Provider 不寫死
接口:
generate(request) -> response
可以是:
remote frontier model
local model
specialized theorem model
symbolic subsystem
hybrid
AMRR semantics 不綁定單一模型。
58. Tool Space 與 Cognitive Space 統一
例如:
CALL web.search
CALL lean.verify
CALL cas.simplify
CALL cog://math/diagnose-gap@1
CALL cog://math/change-representation@1
59. Mathematical Action 也有 Authority
例如:
create conjecture candidate = ALLOW
create problem branch = ALLOW
promote new axiom = APPROVAL_REQUIRED
publish novelty claim = APPROVAL_REQUIRED
erase original problem = DENY
所以:
60. Budget Model
AMRR budget:
不同 operator 使用不同 budget class。
61. Budget-Aware Verification
若完整 formal proof 成本過高,可先:
例如:
但便宜 verifier 不得被當成更強 verifier 的替代證據。
62. Failure Model
AMRR 至少區分:
problem_error
definition_error
assumption_error
representation_error
method_error
proof_error
formalization_error
verifier_error
tool_error
bridge_error
novelty_error
authority_error
protocol_error
ledger_error
budget_error
environment_error
不能全部叫:
failure
63. Unknown 是合法狀態
允許:
unknown
ambiguous
inconclusive
unresolved
因此:
以及:
64. Context Compression
AMRR working context 必然要壓縮。
但:
必須保留:
problem refs
theory refs
gap refs
obligation refs
assumption envelopes
verification receipts
bridge refs
research receipts
causal parents
65. Research Ledger
Research ledger 保存:
66. CTCL Event Namespace
第一版:
math.problem.registered
math.problem.derived
math.problem.relation.proposed
math.problem.relation.verified
math.theory.registered
math.theory.extension.proposed
math.theory.extension.classified
math.theory.extension.promoted
math.gap.detected
math.gap.updated
math.gap.closed
math.gap.reopened
math.repair.proposed
math.repair.accepted
math.repair.rejected
math.obligation.created
math.obligation.updated
math.obligation.discharged
math.obligation.waived
math.verification.requested
math.verification.completed
math.bridge.proposed
math.bridge.verified
math.claim.proposed
math.claim.approved
math.claim.blocked
math.research.idled
math.research.deferred
math.research.escalated
67. Temporal Evidence
每個重大事件至少引用:
ctcl_instant_id
occurred_at
recorded_at
run_id
interaction_round
ledger_seq
causal_parent_ids
state_ref
problem_ref
theory_ref
contract_ref
knowledge_boundary_ref
68. Research Receipt
對一個完成的研究 transition:
{
"research_receipt_id": "research-receipt:...",
"agenda_ref": "agenda:...",
"state_before_ref": "state:...",
"state_after_ref": "state:...",
"problem_refs": [],
"theory_refs": [],
"gap_refs": [],
"change_set_refs": [],
"obligation_refs": [],
"verification_refs": [],
"decision_refs": [],
"causal_parent_ids": []
}
69. Research Receipt 不保存 Private CoT
保存:
public state
selected cognitive program
reason codes
math changes
verification results
receipts
artifact refs
不要求:
hidden reasoning tokens
因此:
70. Replay 與 Re-enactment
Replay:
Re-enactment:
所以:
71. AMRR API v0.1
runtime.observe_research(...)
runtime.encode_math_state(...)
runtime.diagnose_gaps(...)
runtime.propose_research_agenda(...)
runtime.retrieve_math_affordances(...)
runtime.compile_math_program(...)
runtime.execute_math_program(...)
problem.register(...)
problem.derive(...)
problem.relate(...)
problem.branch(...)
theory.register(...)
theory.extend(...)
theory.classify_extension(...)
theory.bridge(...)
repair.propose(...)
repair.apply_candidate(...)
repair.rollback(...)
obligation.generate(...)
obligation.discharge(...)
obligation.defer(...)
obligation.waive(...)
verify.route(...)
verify.run(...)
claim.assumption_envelope(...)
claim.validate_scope(...)
claim.propose(...)
research.audit(...)
research.replay(...)
72. Repo Structure 建議
不建議一開始另開完全獨立 runtime。
優先:
addressable-cognitive-runtime/
├── src/addressable_cognitive_runtime/
│ ├── state/
│ ├── registry/
│ ├── retrieval/
│ ├── compiler/
│ ├── runtime/
│ ├── governance/
│ ├── ctcl_adapter/
│ └── math/
│ ├── state/
│ ├── problem/
│ ├── theory/
│ ├── gap/
│ ├── repair/
│ ├── obligation/
│ ├── verifier/
│ ├── bridge/
│ ├── claim/
│ └── research_loop/
├── schemas/
│ └── math/
├── registry/
│ └── math/
├── adapters/
│ └── math/
├── experiments/
│ └── amrr_gates/
└── tests/
└── math/
73. Math Schema Catalog
建議新增:
mathematical-state-extension.schema.json
problem-record.schema.json
problem-relation.schema.json
problem-mutation.schema.json
assumption-envelope.schema.json
theory-record.schema.json
theory-extension.schema.json
math-change-set.schema.json
gap.schema.json
gap-map.schema.json
repair-candidate.schema.json
mathematical-obligation.schema.json
obligation-graph.schema.json
verification-request.schema.json
verification-receipt.schema.json
theory-bridge.schema.json
discovery-receipt.schema.json
research-receipt.schema.json
claim-record.schema.json
74. Schema Versioning
每個 schema:
schema_id
schema_version
canonicalization_version
Identity:
不能靠檔名推斷語義。
75. Phase M0 — Math Schema Freeze
先建立全部 math schemas。
成功條件:
round-trip
invalid-case rejection
stable IDs
canonical serialization
hash stability
no hidden implicit fields
76. Phase M1 — Mathematical State Adapter
輸入仍是公開 observation。
輸出:
成功條件:
77. Phase M2 — Problem / Theory Stores
建立:
Problem DAG
Theory version graph
Assumption Envelope
stable refs
成功條件:
78. Phase M3 — Gap Diagnosis Engine
先做 controlled benchmark。
第一批只支援:
definition
assumption
representation
lemma/dependency
compute
verification
不要一開始硬做全部十三域。
79. Phase M4 — Problem Identity / Theory Extension
加入:
problem relation classifier
problem mutation receipt
theory extension classifier
assumption-envelope update
成功條件:
80. Phase M5 — Obligation Engine
rule-first obligation templates。
成功條件:
同時控制 false obligations。
81. Phase M6 — Math Cognitive Registry
先建立 15–25 個 operator。
可:
resolve
validate
render
compose
replay
82. Phase M7 — Verifier Router
第一版只接少量 verifier class:
formal proof
symbolic
numeric / exhaustive counterexample
semantic review
先證明 routing semantics,再擴 provider。
83. Phase M8 — CMDC Repair Engine
支援:
definition repair
assumption candidate
representation change
lemma candidate
method retrieval
所有 repair 只建立 candidate state。
84. Phase M9 — Theory Bridge / Prior Art
加入:
known concept mapping
formal-library refs
literature refs
bridge strength
novelty status
85. Phase M10 — CTCL / Receipt Adapter
所有重大 math events:
並生成:
DiagnosisReceipt
RepairReceipt
ProblemMutationReceipt
TheoryExtensionReceipt
VerificationReceipt
ResearchReceipt
86. Phase M11 — Persistent Research Loop
最後才接:
不再逐輪提供:
下一步請做什麼。
87. MVP 第一個環境
最適合不是直接丟 Millennium Problem。
第一個環境應是:
內含:
small theorem set
known malformed problems
missing assumptions
ambiguous definitions
false conjectures
missing lemmas
simple Lean targets
finite counterexample domains
known prior-art mappings
88. 第一個 Demo
Research Goal:
Investigate whether the target statement is valid and produce a reviewable result package.
Environment:
ambiguous definition
one hidden counterexample
formal target available
small prior-art corpus
期待:
Observe
→ DefinitionGap
→ ClarificationCandidate
→ ProblemRelation
→ ObligationGeneration
→ CounterexampleSearch
→ conjecture refuted
→ restricted branch proposed
→ verification
→ TheoryPackage
→ STOP / ESCALATE
89. 第二個 Demo — Silent Assumption Trap
原問題:
模型容易加:
期待:
不能被包成:
測:
90. 第三個 Demo — Wrong Domain
真 gap:
比較:
與:
測:
91. 第四個 Demo — False Conjecture
AI 不應無限 proof search。
期待:
92. 第五個 Demo — Faithfulness Trap
自然語言問題:
提供一個容易證但偏離原意的:
測:
93. 第六個 Demo — Prior Art
AI 生成新 definition。
但 corpus 中已有等價概念。
期待:
而不是:
94. Falsification Gate A — State Determinism
95. Gate B — Gap Diagnosis
測 multi-label confusion matrix。
96. Gate C — Matched Affordance
97. Gate D — Problem Identity
98. Gate E — Obligation Coverage
99. Gate F — No Laundering
100. Gate G — Verifier Routing
101. Gate H — Repair Validity
102. Gate I — Gap Reopen
加入 delayed counterexample 後:
103. Gate J — Context Recovery
context 壓縮後能否用:
恢復:
problem lineage
assumption envelope
open obligations
verification basis
104. Gate K — Long-Horizon Research
只給:
測:
個 research transitions。
105. 評估向量
AMRR evaluator:
其中:
- :problem-solving performance;
- :diagnosis;
- :repair validity;
- :identity preservation;
- :obligation quality;
- :verification;
- :bridge quality;
- :novelty calibration;
- :continuity / auditability;
- :efficiency。
106. 不用單一 Reward
避免:
把:
抵銷:
因此初期使用 vector evaluation。
107. 成功與失敗的新定義
一個 final theorem 沒有證成,不一定是 research failure。
例如:
且 AI 找到可驗證反例,這可能是:
反過來:
但偷加假設,則可能是:
108. AMRR v0.1 不做什麼
第一版不做:
fully autonomous publication
unbounded axiom creation
unsupervised foundational replacement
claiming universal novelty
claiming universal mathematical correctness
philosophical consciousness
legal mathematical personhood
full automatic community acceptance
109. AMRR v0.1 真正只做什麼
第一版只回答:
給定一個受控數學研究環境、持續 Research Goal 與 Research Contract,AI 是否能在不需要人類逐輪指定下一步的情況下,持續診斷問題域、選擇數學認知程序、提出受約束 repair、保留問題身份、管理義務、調用 verifier、形成 theory state,並留下可恢復的研究因果歷史?
110. 與前三篇的統一
Paper 01:
Paper 02:
Paper 03:
Paper 04:
111. 系列最終總模型
112. 系列最終不變量
113. 核心命題
本文最終提出:
中文:
自主數學不應被實作成一個不停產生文字與證明的 Agent,而應被實作成一個具有 canonical 數學狀態、問題與理論版本、typed gap、受約束 repair、數學義務、外部 verifier、治理決策與時間因果研究歷史的持續 Runtime。
更完整地:
114. 結論
數學 AI 若只會等待人類提供完整問題,它仍主要是一個 solver。
若它可以自己選 proof strategy,但不能知道自己其實缺的是定義、假設、表示或 theory bridge,它仍缺少研究層的 diagnosis。
若它可以自己改問題與建理論,卻不能保存 Problem Identity、Theory Extension 與 Assumption Envelope,它可能只是在「改到可解」。
若它可以產生 proof,卻沒有 typed verifier、semantic faithfulness 與 claim-scope gate,它可能只是在更可靠地證明錯誤形式化。
因此真正的自主數學研究需要:
本文提出 AMRR v0.1 作為這些能力的統一 Runtime 架構。
它不要求 AI 一開始就具有無界的數學創造力。
相反,它要求每一項自治能力都通過:
當這條路線完成後,人類與數學 AI 的接口才可能真正從:
逐步轉為:
人類不再必須逐輪替 AI 撰寫每一個下一步;但數學身份、證明義務、理論依賴、權限邊界與研究歷史仍保持可檢查、可回退、可驗證。
這才是本系列所稱的:
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.
[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] 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.
[6] Rabe, F., & Kohlhase, M. (2013). A Scalable Module System. Information and Computation, 230, 1–54. DOI: 10.1016/j.ic.2013.06.001.
[7] Xue, T. (2014). Definitional Extension in Type Theory. TYPES 2013, LIPIcs 26, 251–269. DOI: 10.4230/LIPIcs.TYPES.2013.251.
[8] 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.
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). 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,Autonomous Mathematical Research / Paper 02 of 04,v0.1.
[I3] Neo.K. (2026). 問題身份、理論擴張與數學義務:AI 生成數學的合法變換框架,Autonomous Mathematical Research / Paper 03 of 04,v0.1.
[I4] 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.
[I5] 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.
[I6] 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.
[I7] 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.
[I8] 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.
[I9] 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.
[I10] Addressable Cognitive Runtime MVP v0.1.2. Phase 2 Validation: Semantic State Encoder. Internal implementation artifact, 2026-08-21.
版本備註
v0.1 / 2026-08-23
本版正式固定:
- AMRR 五層 Runtime architecture;
extensions.math兼容 ACR SemanticState 的策略;- Problem Store / Problem DAG;
- Theory Store / Theory Extension Classifier;
- Gap Diagnosis Engine / Gap DAG;
- CMDC Repair Engine / MathChangeSet;
- Mathematical Obligation Engine / Obligation Graph;
cog://math/*Cognitive Registry;- Mathematical Research Program Compiler;
- Verifier Router / Verifier Ensemble;
- Assumption Envelope / No-Laundering Gate;
- Theory Bridge / Novelty Layer;
- Research Contract / Agenda / Governance;
- CTCL math event namespace 與 mathematical receipts;
- Persistent Research Loop;
- Phase M0–M11 實作順序;
- Controlled Mathematical Research Sandbox;
- Gate A–K falsification program;
- vector evaluation;
- 與 ACR Phase 0–2 既有工程的相容性邊界。
本版仍是 architecture specification paper,不宣稱 AMRR 的 M0–M11 已全部實作。現有工程基線只採既有 ACR Phase 0–2;所有新增 math runtime components 應在後續 technical whitepaper / MVP 中逐項實作與否證。