空間域證明包圍論 VIII
Runtime、Benchmark 與 Proof-Space Observatory
Spatial-Domain Proof Enclosure VIII: Runtime Semantics, Benchmarks, and a Proof-Space Observatory
Version: v0.1
Date: 2026-08-14
Status: Phase-I integration paper / executable reference architecture; not a complete proof assistant and not a claim of universal proof-search acceleration
Canonical source: UTF-8 Markdown; canonical mathematics uses $...$ and $$...$$ only.
摘要
前七篇空間域證明包圍論依序建立:sound survivor contraction、route representation faithfulness、global coverage / closure certificates、proof-trace compilation、Discovery–Verification cost dynamics、measure-zero exceptional-core analysis,以及 cost-aware enclosure routing。這些結果共同提出一個新的工程問題:若研究真的被視為一個持續收縮、可回滾、可重播的 counterexample survivor process,那麼其正式 runtime 狀態、authority boundaries、event history、benchmark semantics 與 observability contract 應如何設計?
本文給出 SDPE Phase I 的第一個整合 runtime。其 canonical state 為
本文採用 append-only event ledger 與 deterministic reducer。令
為事件序列,則 authoritative state 由
重建。由此證明 Event-Sourced Replay Determinism 與 Checkpoint–Tail Equivalence:只要 reducer、schema、environment fingerprint 與 prefix checkpoint 一致,完整重播與 checkpoint 加 tail replay 產生相同 proof state。
本文進一步建立 proof-authority separation。Proposer、Router、Observatory 只可產生 proposal、policy decision 與 telemetry;它們都無權直接縮小 或設定 。只有通過 verifier、route / representation contract、coverage、boundary、必要 gluing、dependency / version 等 gate 的 certificate,才可由 commit controller 激活並改變 authoritative survivor state。這給出 Certificate-Gated Commit Preservation:若初始 survivor sound,且每個 activated exclusion certificate 都是 counterexample-safe,則所有 committed states 永遠保持
本文同時把 Paper 04 的 support-aware rollback 實作成 runtime invariant。certificate 失效時,真正 reopen 的區域只包含失去最後一個 active sound support 的 states,而不是所有曾被該 theorem 觸碰的區域。
在 benchmark 層,本文採納一個更嚴格的分層:kernel proof validity、formal statement fidelity、route / coverage fidelity、runtime reproducibility 與 observatory metrics 分屬不同 authority channels。2026 年對 Lean theorem-proving benchmarks 的 audit 顯示,即使 solution kernel-check 通過,benchmark 本身仍可能包含 vacuity、錯誤 formalization、缺 hypothesis 或 unsound axiom。因此 SDPE benchmark 不允許以「Lean 編譯成功」替代 problem-fidelity audit。
本文提出八條 benchmark tracks:finite closure、representation、global coverage、rollback / replay、DVI、exceptional-core、routing,以及 domain adapter。每個 benchmark 必須 pin problem fingerprint、formalization / adapter version、runtime schema、verifier / kernel version、dependency snapshot、routing policy、model / prover version、hardware / concurrency metadata、random seed 與 audit version。Proof-Space Observatory 可以讀取並顯示 survivor geometry、DVI telemetry、gap debt、routing values、dirty replay 與 support multiplicity,但 observatory 是非權威的:任何 dashboard 指標都不能自行提交 theorem cut 或 global closure。
本文最後提供一個標準 Python prototype、JSON state / event / benchmark schemas、append-only example ledger、finite structural checker 與 Hard-Zeta case-study adapter specification。這使 SDPE Phase I 從一組概念論文轉成一個可重播、可 benchmark、可擴充的 proof-space runtime contract。
關鍵詞
空間域證明包圍;proof runtime;event sourcing;proof-space observatory;Global Closure Certificate;proof provenance;rollback;incremental replay;benchmark fidelity;formal verification;AI theorem proving;proof telemetry;survivor space
1. Phase I 的整合問題
前七篇回答的問題可以排列為:
但若這些對象存在於不同文件、不同 AI session、不同 theorem prover、不同 notebook 或不同版本,它們仍不能形成一個可靠長時程研究系統。
Paper 08 的問題是:
如何建立一個最小 runtime,使上述所有 proof objects 具有一致 state semantics、event provenance、authority separation、rollback、replay、benchmark 與 observability?
本文不嘗試建立完整 theorem prover。SDPE Runtime 的角色是proof-space control plane:管理證明域、證書、依賴、coverage、routing 與研究 telemetry,而底層 proof checking 可由 Lean、Coq、Isabelle、SMT/SAT certificate checker、domain-specific checker 或人工形式化 audit 承擔。
2. Fresh literature grounding
2.1 Scalable verifier infrastructure
AXLE 在 2026 年把 Lean proof verification、proof manipulation、metadata extraction、多版本支援與 per-request isolation 提升成 scalable cloud infrastructure。這說明 AI mathematics runtime 不應把 verifier 視為單一 compile 指令,而應把版本、隔離、request identity、metadata 與 strict verification 當成 first-class runtime state。
SDPE 取其工程原則:
2.2 Formal proof validity is not benchmark fidelity
2026 年對五個 Lean theorem-proving benchmarks 的 audit 顯示,kernel-check 只能證明 proof term 證成了 formal statement,不能證明 formal statement 忠實對應 intended problem,也不能保證 benchmark harness 沒有 vacuity、unsound axiom、missing hypotheses 或其它 specification defects。
因此 SDPE 將 benchmark authority 拆成:
2.3 Dependency-rich benchmark design
TheoremBench 同時發布 main theorem 與 premise-expanded theorem families,並用 theorem-level coverage / token-efficiency 描述 proof behavior。這支持 SDPE benchmark 不只記錄「最後成功或失敗」,還要保存 dependency / partial progress / closure-obligation structure。
2.4 Observable agent loops
OpenProver 將 Planner、Workers、Whiteboard / Repository 與 formal Verifier 分離,並提供 reproducible automatic verification。這與 SDPE authority separation 相容:規劃、生成、儲存與驗證不應由同一個未區分的 agent authority 承擔。
2.5 Operational state reuse
Proof-state snapshotting 顯示,重建 elaborated state 可佔 branching search 的主要成本,而直接 reuse state 可以大幅降低 wall time。SDPE 將 snapshot / checkpoint 視為 operational accelerator;它不自動繼承 theorem validity authority。
3. Canonical SDPE Runtime State
Definition 3.1 — SDPE State
定義 epoch 的 runtime state:
其中:
- :authoritative survivor envelope;
- :active route / representation contracts;
- :typed closure gaps;
- :Global Closure Certificate state;
- :proof-history dependency DAG;
- :closure basis、support index、snapshots、dirty / reopen state;
- :Paper 05 cost telemetry;
- :Paper 06 survivor geometry profile;
- :Paper 07 routing decision / policy provenance;
- :runtime / checker / dependency / model / hardware fingerprints。
3.2 Authoritative vs diagnostic fields
將 state fields 分成兩類。
Proof-authoritative
Diagnostic / policy
後者可以改變研究排序,不能直接改變數學真值或 closure bit。
4. Proof Authority Separation
Definition 4.1 — Authority channels
SDPE Runtime 至少區分:
- Proposer:產生 theorem / refinement / repair candidate;
- Router:排序或選擇 action;
- Verifier:驗證 local certificate;
- Coverage Auditor:驗證 route / branch cover;
- Boundary Auditor:驗證 equality / singular / degenerate strata ownership;
- Glue Auditor:只在 ConstructiveGluing mode 驗證 compatibility / gluing;
- Compiler:建立 proof index / closure basis / snapshot / support map;
- Commit Controller:在 gates 通過後激活 certificate;
- Observatory:只讀 telemetry / visualization。
Proposition 4.2 — Proposal Non-Authority
若 runtime mutation contract 規定只有 Commit / Invalidate 類 authority events 能改變 authoritative survivor state,則任意 proposal 或 route-value prediction 都不能單獨推出:
這是系統規格命題,不是數學 theorem:它必須由 runtime enforcement 實作。
No-Go 4.3 — Model Confidence as Proof Authority
以下都不能激活 theorem cut:
- LLM confidence;
- route score;
- retrieval similarity;
- sampled empirical support;
- observatory anomaly score;
- benchmark historical success rate。
只有 registered proof / audit certificate 可進 commit gate。
5. Event-Sourced Proof State
Definition 5.1 — Event Ledger
令 append-only event ledger 為:
每個 event 至少帶有:
Definition 5.2 — Deterministic reducer
令
為 deterministic reducer。
狀態定義為:
Theorem 5.3 — Event-Sourced Replay Determinism
若:
- initial state 相同;
- event ledger byte-equivalent / canonically equivalent;
- reducer version 相同;
- reducer deterministic;
則 full replay 的 final state 唯一。
Proof
對 event sequence length induction。零事件時 state 相同。若前 個 events 後 state 唯一,由 deterministic ,第 個 event 對相同 state 產生唯一 successor。故成立。
Corollary 5.4 — State fingerprint
若 state 使用 canonical serialization,可定義:
不同 machine / session replay 應得到相同 fingerprint,否則 runtime / environment drift 必須被視為 verification event。
6. Checkpoint–Tail Replay
對長時程研究,不應每次從 event 0 完整重播。
令 prefix ledger:
產生 checkpoint:
Theorem 6.1 — Checkpoint–Tail Equivalence
若 checkpoint 經完整 prefix replay 驗證,且 reducer / schema / environment fingerprint 未變,則:
Proof
由 fold associativity 與 checkpoint 定義立即成立。
No-Go 6.2 — Snapshot Equals Certificate
checkpoint 只證明:「在指定 reducer / environment 下可以重建這個 state」。
它不證明 checkpoint 內每個 theorem / route / coverage certificate 的數學正確性。certificate authority 仍由其原始 checker / dependency provenance 承擔。
7. Certificate-Gated Commit
Definition 7.1 — Commit gate
令 candidate certificate 要進 active state,至少要求:
若 conclusion 要提交 global closure,還需要 Paper 03 的 GCC obligations。
Theorem 7.2 — Certificate-Gated Survivor Preservation
設初始 state 滿足:
若每一個 committed exclusion certificate 排除的 region 都已 soundly 證明:
且 committed survivor update 為:
則所有 committed epochs 保持:
Proof
若 ,由 induction hypothesis 。又對所有 committed , 。故 。
7.3 GCC bit 的語義
GCC.Valid = true 本身不是 proof artifact。它只是一個 runtime result bit,其可置信性來自:
因此最終 publication / submission 應附 closure basis / proof artifacts,而不是只附一個 runtime JSON flag。
8. Invalidation, Dirty State, and Precise Reopening
Paper 04 定義 support multiplicity:
certificate 失效後:
Proposition 8.1 — Precise support-aware reopen
若 support index complete,則 certificate invalidation 後:
- 的 point 仍有 sound active exclusion support,不需 reopen;
- 且 previously excluded 的 point 必須 reopen。
這是 stale theorem 局部化修復的 runtime realization。
8.2 Dependency dirty closure
對失效 proof node :
只有 dirty closure 與其受影響 coverage / GCC nodes 需要 incremental replay。
9. Typed Gap API
Paper 03 / 06 / 07 的 gaps 在 runtime 中至少分為:
典型 API:
register_gap(type, scope, witness, mandatory, dependencies);resolve_gap(gap_id, certificate_id);reopen_gap(gap_id, stale_dependency);query_persistent_gaps();query_starvation_age(gap_id)。
Gap priority 是 routing policy;gap closure 是 certificate semantics。二者不得混淆。
10. Route API
每個 route 至少帶有:
runtime 應提供:
register_route_cert;invalidate_route_cert;refine_representation;query_singular_fibers;query_route_coverage。
任何 route representation 改版都應使依賴舊 representation fingerprint 的 certificates 進 dirty state。
11. Proof DAG and Closure-Basis API
proof DAG node 至少保存:
runtime API:
add_verified_certificate;active_ancestors;closure_basis(target);invalidate_certificate;dirty_closure;incremental_replay;support_index(region)。
12. Proof-Space Observatory
Definition 12.1 — Observatory
Proof-Space Observatory 是 state / event 的只讀 projection:
其輸出可以包括:
以及:
Theorem 12.2 — Observatory Non-Authority
若 runtime mutation rules 不接受 observatory output 作 proof-authoritative event,則 observatory telemetry 的任意改變不能單獨改變:
這是安全架構原則:
13. Runtime Pipeline
Phase-I reference pipeline:
這不是強迫所有 domain 逐步執行每一 stage。RefutationOnly problem 可以讓 GlueAudit 返回 not-required certificate;沒有 representation change 的 epoch 可以略過 refine action;但所有被略過的 authority gate 必須有明確的 not-applicable semantics,不能靠 silent omission。
14. Benchmark Authority Separation
Formal theorem benchmark 至少有四層 correctness:
14.1 Proof validity
底層 kernel / checker 接受 proof artifact。
14.2 Statement fidelity
formal statement 是否忠實代表 intended mathematical problem。
14.3 Closure / route fidelity
local cases 是否真正 cover domain、representation 是否 sound、boundaries 是否有 ownership。
14.4 Evaluation fidelity
metric / harness 是否沒有 leakage、vacuity、shortcut、version mismatch 或不公平 cost accounting。
因此:
這一點不是理論上的杞人憂天;2026 benchmark audit 已在現有 Lean benchmarks 中機械發現多種問題。
15. SDPE Benchmark Tracks
本文定義八條最小 tracks。
Track A — Finite Closure
有限 universe 有 hidden oracle。測:
- survivor soundness;
- theorem cut;
- cover certificate;
- GCC;
- false closure rejection。
Track B — Representation
測:
- fiber saturation;
- mixed-fiber retention;
- route refinement;
- lift certificate。
Track C — Global Coverage
測:
- branch union;
- boundary omission;
- overlap;
- RefutationOnly / ConstructiveGluing mode。
Track D — Rollback / Replay
測:
- stale dependency;
- dirty closure;
- support-aware reopen;
- full vs incremental replay;
- checkpoint equivalence。
Track E — DVI
沿用 Paper 05:
- Queue A known / replay;
- Queue B near-frontier transfer;
- Queue C true frontier;
- cold vs compiled;
- build / maintenance / replay cost。
Track F — Exceptional Core
測:
- measure-zero survivor;
- singleton localization;
- theorem-language irreducible core;
- representation-singular residue。
Track G — Routing
測:
- volume greedy;
- gap-aware;
- bridge-aware;
- Pareto routing;
- fairness;
- time-to-GCC。
Track H — Domain Adapter
將真實數學研究映射成 SDPE objects,例如:
- Hard-Zeta / Collatz;
- SAT / cube-and-conquer;
- finite graph classification;
- repository-scale Lean verification。
Domain adapter benchmark 不應假裝 hidden oracle 存在;它主要測 provenance、state reconstruction、routing、certificate integration 與 research telemetry。
16. Benchmark Fingerprints
每次 run 至少 pin:
定義 benchmark identity:
若這些 fingerprint 不同,runtime 可以比較結果,但不得把它們默認當成同一 experimental condition。
17. Benchmark Metrics
17.1 Correctness metrics
- false exclusion count;
- false GCC count;
- missed reopen count;
- stale certificate activation count;
- route / boundary omission count。
在 sound benchmark 中,proof-authoritative false-positive 應要求:
17.2 Replay metrics
- full replay time;
- checkpoint-tail replay time;
- dirty replay fraction;
- state fingerprint mismatch rate。
17.3 Compilation metrics
- compiled hit rate ;
- closure-basis compression ;
- support multiplicity;
- reopen precision;
- build / maintenance cost。
17.4 Closure metrics
- active mandatory gap count;
- boundary debt;
- certificate debt;
- time-to-GCC;
- accepted nonredundant cut count。
17.5 DVI metrics
沿用:
17.6 Routing metrics
- nonredundant closure yield;
- action cost;
- routing regret;
- mandatory-gap starvation age;
- bridge unlock rate。
18. Runtime Replay Theorems and Their Boundaries
Event-sourced determinism 證明的是:
它不證明:
同樣:
只在 dependency metadata、checker determinism、version fingerprints 完整的 runtime contract 下成立。
因此 runtime theorem 與 mathematical theorem 必須分層記帳。
19. Benchmark-Fidelity Gate
在正式 benchmark acceptance 前,至少要求:
- statement fidelity audit;
- proof-kernel / checker version pinned;
- banned axioms / shortcuts policy declared;
- dependency graph audited;
- test harness replayable;
- environment fingerprint complete;
- hidden oracle 只在 toy / synthetic track 使用;
- metric authority 不超過其定義範圍。
No-Go 19.1 — Kernel-Checked Means Benchmark-Correct
錯。
kernel 只回答:
這個 proof term 是否證成這個 formal statement?
它不回答:
這個 formal statement 是否忠實翻譯 intended problem?
也不回答:
benchmark route / coverage / cost protocol 是否公平?
20. Hard-Zeta Adapter as a Case-Study Interface
SDPE Runtime 不重新證明 Hard-Zeta 的任何 theorem。adapter 只規定如何映射已存在的研究 artifacts。
可將 B-side frontier branches 映射成:
與 route nodes;把每個 theorem / no-go / reduction 放入 proof DAG;把 survivor parameter inequalities 存成 symbolic constraint payload;把 checker / JSON / literature provenance 綁到 certificate nodes;把 Sparse / Huge-PQ / structured survivor ladder 等 branch frontier 當作 typed gaps / route obligations。
任何 numerical exponent、checker PASS 或 paper statement 都只有在其 upstream proof certificate 被正式接入 runtime 後,才可對 authoritative survivor state 產生作用。
因此 adapter 不是「把聊天紀錄直接變 proof」,而是:
21. Reference Prototype
本 bundle 提供:
sdpe_runtime.py:deterministic event reducer;SDPE_State_Schema_v1.json;SDPE_Event_Schema_v1.json;SDPE_Benchmark_Schema_v1.json;example_finite_closure_run.jsonl;example_final_state.json;verify_SDPE_Paper08_runtime.py;- Hard-Zeta adapter specification。
prototype 刻意保持簡單:finite explicit survivor universe 用於 structural checker;真正 infinite / symbolic mathematics 必須由 domain adapter 與外部 proof checker 提供 sound symbolic region semantics。
22. Structural Checker Results
companion checker 驗證:
- reference ledger replay 到 empty survivor + valid GCC;
- 組 random event histories deterministic replay;
- 組 checkpoint-tail equivalence;
- 組 hidden-oracle survivor soundness;
- proposal non-authority;
- telemetry non-authority;
- unverified-certificate commit rejection;
- redundant-support rollback precision;
- GCC authority separation。
checker 只驗證 finite runtime semantics,不驗證 arbitrary mathematical certificate 的內容真偽。
23. No-Go Ledger
No-Go 23.1 — Runtime State Equals Mathematical Truth
runtime 只能攜帶 certificate-backed belief state;真理不由 database bit 創造。
No-Go 23.2 — Proposer Can Commit Its Own Result
proposal / verification / commit 必須 authority-separated。
No-Go 23.3 — Observatory Metric Can Close Proof
telemetry 沒有 proof authority。
No-Go 23.4 — Kernel Acceptance Guarantees Statement Fidelity
formal proof validity 與 intended-problem fidelity 是不同 obligations。
No-Go 23.5 — Checkpoint Is a New Proof
checkpoint 是 replay accelerator,不是新增 theorem。
No-Go 23.6 — Event Replay Repairs Missing Dependencies
若 dependency metadata 本身不完整,deterministic replay 只能穩定重現錯誤 state。
No-Go 23.7 — JSON Schema Proves Semantics
schema 只保證 shape;route / theorem / coverage soundness 需要各自 certificate。
No-Go 23.8 — A Benchmark Score Is a Proof-Space Law
benchmark 只能支持被其 design / controls 覆蓋的 empirical claim。
No-Go 23.9 — Domain Adapter Imports Proof Automatically
adapter 只做 typed mapping;原 artifacts 必須驗證後才能 commit。
No-Go 23.10 — Phase-I Runtime Is a Complete Proof Assistant
不是。它是 proof-space orchestration / audit layer。
24. Theorem / System-Assumption / External-Input Ledger
24.1 Internal runtime theorems / propositions
- Event-Sourced Replay Determinism;
- Checkpoint–Tail Equivalence;
- Proposal Non-Authority under enforced mutation rules;
- Certificate-Gated Survivor Preservation;
- Precise Support-Aware Reopen under complete support index;
- Observatory Non-Authority;
- Benchmark Identity under canonical fingerprints。
24.2 Inherited SDPE theorems
- P01 survivor soundness;
- P02 representation faithfulness;
- P03 global cover / GCC;
- P04 closure basis / rollback / incremental replay;
- P05 DVI cost architecture;
- P06 exceptional-core semantics;
- P07 enclosure routing。
24.3 Runtime correctness assumptions
- event ledger append-only / tamper-evident;
- reducer deterministic;
- canonical serialization stable;
- certificate checker identity / version pinned;
- dependency metadata complete;
- support index complete;
- authority boundaries enforced;
- benchmark formalization audit honest and versioned。
24.4 External grounding
- AXLE scalable strict Lean infrastructure;
- 2026 formal benchmark fault audit;
- TheoremBench dependency-rich benchmark structure;
- OpenProver Planner / Worker / Verifier runtime;
- Lean proof-state snapshotting;
- testing-style semantic evaluation of formal statements。
25. Phase-I Closure
八篇形成第一階段完整 stack:
Phase I 現在不再只是「一套想法」;它具有:
- proof-space semantics;
- closure obligations;
- representation contracts;
- event provenance;
- certificate / dependency state;
- rollback / replay;
- cost telemetry;
- routing semantics;
- benchmark protocol;
- executable finite prototype。
26. Final Status
Paper 08 的核心原則可以壓縮成五句。
第一:
第二:
第三:
第四:
第五:
因此,空間域證明包圍第一階段現在形成一個完整迴圈:
它仍然不能保證任意猜想會被證明;但它把「長期 AI 輔助數學研究如何留下可驗證痕跡、如何安全收縮反例域、如何避免局部正確冒充全域閉合、如何重用歷史、如何辨認 exceptional hard core、如何選下一刀、如何 benchmark 整個過程」統一成了一個可執行的 proof-space architecture。
References
- Jimmy Xin, Alex Schneidman, Chris Cummins, Karun Ram, Srihari Ganesh, Jannis Limperg, AXLE: A Cloud Infrastructure for Lean 4 Theorem Proving Utilities, arXiv:
2606.26442, 2026. - Pawan Sasanka Ammanamanchi, Siddharth Bhat, Stella Biderman, Faults in Our Formal Benchmarking: Dataset Defects and Evaluation Failures in Lean Theorem Proving, arXiv:
2606.29493, 2026. - QuocViet Pham, Elvir Karimov, Andrey Galichin, Ivan Oseledets, TheoremBench: Evaluating LLMs on Theorem Proving in Formal Mathematics, arXiv:
2606.09450, 2026. - Matěj Kripner, Milan Straka, OpenProver: Agentic and Interactive Theorem Proving with Lean 4, arXiv:
2607.09217, 2026. - Austin Shen, Yunong Shi, Keep the Proof State Live: Snapshotting for Efficient Tactic Search in Lean 4, arXiv:
2605.25556, 2026. - Jongyoon Kim, Hojae Han, Seung-won Hwang, Benchmarking Testing in Automated Theorem Proving, arXiv:
2604.23698, 2026. - Prior SDPE artifacts: Papers 01--07.