← Archive
lm-002895 · 2026-08

空間域證明包圍論 VIII

下載 MD 檔 ⬇
📎 附件 · Companion files — 隨文交付的程式 / 證明 / 資料,可獨立下載重驗

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

St=Ωt,RouteCertt,Gt,GCCt,Ht,Compiledt,DVIt,SurvProft,RouteDecisiont,Envt.\boxed{ \mathbb S_t = \langle \Omega_t, RouteCert_t, \mathbf G_t, GCC_t, \mathcal H_t, Compiled_t, DVI_t, SurvProf_t, RouteDecision_t, Env_t \rangle. }

本文採用 append-only event ledger 與 deterministic reducer。令

Lt=(e0,e1,,et)\mathcal L_t=(e_0,e_1,\ldots,e_t)

為事件序列,則 authoritative state 由

St=Fold(R,S1,Lt)\boxed{ \mathbb S_t = \operatorname{Fold}(\mathcal R,\mathbb S_{-1},\mathcal L_t) }

重建。由此證明 Event-Sourced Replay DeterminismCheckpoint–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 Ωt\Omega_t 或設定 GCC.ValidGCC.Valid。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

CΩt.\boxed{ \mathcal C\subseteq\Omega_t. }

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:

P01: 如何安全縮小反例候選域?P02: 如何避免 representation 先丟失 proof-relevant fibers?P03: 如何把 local routes 關閉成 global closure?P04: 如何編譯 proof history 並做 incremental replay?P05: 如何區分 discovery acceleration 與 verification dominance?P06: 如何辨認 measure-zero / singular exceptional core?P07: 如何選擇下一個研究 action?\boxed{ \begin{aligned} &\text{P01: 如何安全縮小反例候選域?}\\ &\text{P02: 如何避免 representation 先丟失 proof-relevant fibers?}\\ &\text{P03: 如何把 local routes 關閉成 global closure?}\\ &\text{P04: 如何編譯 proof history 並做 incremental replay?}\\ &\text{P05: 如何區分 discovery acceleration 與 verification dominance?}\\ &\text{P06: 如何辨認 measure-zero / singular exceptional core?}\\ &\text{P07: 如何選擇下一個研究 action?} \end{aligned} }

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:

verification serviceproposal service.\boxed{ \text{verification service} \neq \text{proposal service}. }

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:

Proof ValidityStatement FidelityCoverage FidelityPerformance Telemetry.\boxed{ \text{Proof Validity} \neq \text{Statement Fidelity} \neq \text{Coverage Fidelity} \neq \text{Performance Telemetry}. }

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 tt 的 runtime state:

St=Ωt,RouteCertt,Gt,GCCt,Ht,Compiledt,DVIt,SurvProft,RouteDecisiont,Envt.\boxed{ \mathbb S_t = \langle \Omega_t, RouteCert_t, \mathbf G_t, GCC_t, \mathcal H_t, Compiled_t, DVI_t, SurvProf_t, RouteDecision_t, Env_t \rangle. }

其中:

  • Ωt\Omega_t:authoritative survivor envelope;
  • RouteCerttRouteCert_t:active route / representation contracts;
  • Gt\mathbf G_t:typed closure gaps;
  • GCCtGCC_t:Global Closure Certificate state;
  • Ht\mathcal H_t:proof-history dependency DAG;
  • CompiledtCompiled_t:closure basis、support index、snapshots、dirty / reopen state;
  • DVItDVI_t:Paper 05 cost telemetry;
  • SurvProftSurvProf_t:Paper 06 survivor geometry profile;
  • RouteDecisiontRouteDecision_t:Paper 07 routing decision / policy provenance;
  • EnvtEnv_t:runtime / checker / dependency / model / hardware fingerprints。

3.2 Authoritative vs diagnostic fields

將 state fields 分成兩類。

Proof-authoritative

Ωt,RouteCertt,Gt,GCCt,Htactive,Compiledtsupport.\boxed{ \Omega_t, RouteCert_t, \mathbf G_t, GCC_t, \mathcal H_t^{\rm active}, Compiled_t^{\rm support}. }

Diagnostic / policy

DVIt,SurvProft,RouteDecisiont,predicted values,dashboard metrics.\boxed{ DVI_t, SurvProf_t, RouteDecision_t, \text{predicted values}, \text{dashboard metrics}. }

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:

  1. Proposer:產生 theorem / refinement / repair candidate;
  2. Router:排序或選擇 action;
  3. Verifier:驗證 local certificate;
  4. Coverage Auditor:驗證 route / branch cover;
  5. Boundary Auditor:驗證 equality / singular / degenerate strata ownership;
  6. Glue Auditor:只在 ConstructiveGluing mode 驗證 compatibility / gluing;
  7. Compiler:建立 proof index / closure basis / snapshot / support map;
  8. Commit Controller:在 gates 通過後激活 certificate;
  9. 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:

Ωt+1Ωt.\Omega_{t+1}\subsetneq\Omega_t.

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:

Lt=(e0,e1,,et).\boxed{ \mathcal L_t=(e_0,e_1,\ldots,e_t). }

每個 event 至少帶有:

seq,kind,payload,actor,version,fingerprint.\langle seq, kind, payload, actor, version, fingerprint \rangle.

Definition 5.2 — Deterministic reducer

R:S×ES\mathcal R: \mathbb S\times\mathcal E\to\mathbb S

為 deterministic reducer。

狀態定義為:

St=Fold(R,S1,Lt).\boxed{ \mathbb S_t = \operatorname{Fold}(\mathcal R,\mathbb S_{-1},\mathcal L_t). }

Theorem 5.3 — Event-Sourced Replay Determinism

若:

  1. initial state 相同;
  2. event ledger byte-equivalent / canonically equivalent;
  3. reducer version 相同;
  4. 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 kk 個 events 後 state 唯一,由 deterministic R\mathcal R,第 k+1k+1 個 event 對相同 state 產生唯一 successor。故成立。 \square

Corollary 5.4 — State fingerprint

若 state 使用 canonical serialization,可定義:

Ft=H(Canon(St)).\boxed{ F_t = H(\operatorname{Canon}(\mathbb S_t)). }

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:

L0:k\mathcal L_{0:k}

產生 checkpoint:

S^k.\widehat{\mathbb S}_k.

Theorem 6.1 — Checkpoint–Tail Equivalence

若 checkpoint 經完整 prefix replay 驗證,且 reducer / schema / environment fingerprint 未變,則:

Fold(R,S1,L0:t)=Fold(R,S^k,Lk+1:t).\boxed{ \operatorname{Fold}(\mathcal R,\mathbb S_{-1},\mathcal L_{0:t}) = \operatorname{Fold}(\mathcal R,\widehat{\mathbb S}_k,\mathcal L_{k+1:t}). }

Proof

由 fold associativity 與 checkpoint 定義立即成立。 \square

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 cc 要進 active state,至少要求:

ProposalExists(c),LocalVerify(c)=pass,DepsActive(c),VersionMatch(c),RouteLiftValid(c),RequiredAuditsValid(c).\boxed{ \begin{aligned} &\mathsf{ProposalExists}(c),\\ &\mathsf{LocalVerify}(c)=\mathrm{pass},\\ &\mathsf{DepsActive}(c),\\ &\mathsf{VersionMatch}(c),\\ &\mathsf{RouteLiftValid}(c),\\ &\mathsf{RequiredAuditsValid}(c). \end{aligned} }

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:

CΩ0.\mathcal C\subseteq\Omega_0.

若每一個 committed exclusion certificate cc 排除的 region EcE_c 都已 soundly 證明:

EcC=,E_c\cap\mathcal C=\varnothing,

且 committed survivor update 為:

Ωt+1=ΩtcAtEc,\Omega_{t+1} = \Omega_t\setminus \bigcup_{c\in A_t}E_c,

則所有 committed epochs 保持:

CΩt.\boxed{ \mathcal C\subseteq\Omega_t. }

Proof

xCx\in\mathcal C,由 induction hypothesis xΩtx\in\Omega_t。又對所有 committed ccxEcx\notin E_c。故 xΩt+1x\in\Omega_{t+1}\square

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:

event replay+active certificate basis+cover / boundary / glue obligations+environment fingerprints.\boxed{ \text{event replay} + \text{active certificate basis} + \text{cover / boundary / glue obligations} + \text{environment fingerprints}. }

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:

κt(x)=#{c:xEc,  c active}.\kappa_t(x) = \#\{c:x\in E_c,\;c\text{ active}\}.

certificate 失效後:

R={x:κold(x)>0,  κnew(x)=0}.\boxed{ R = \{x:\kappa_{\rm old}(x)>0,\;\kappa_{\rm new}(x)=0\}. }

Proposition 8.1 — Precise support-aware reopen

若 support index complete,則 certificate invalidation 後:

  • κnew(x)>0\kappa_{\rm new}(x)>0 的 point 仍有 sound active exclusion support,不需 reopen;
  • κnew(x)=0\kappa_{\rm new}(x)=0 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 vv

Dirty(v)={v}Desc(v).Dirty(v) = \{v\}\cup\operatorname{Desc}(v).

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:

G=(GD,GB,G,GC,GG,GR,Gcore,Grepr).\boxed{ \mathbf G = (G_D,G_B,G_{\partial},G_C,G_G,G_R,G_{\rm core},G_{\rm repr}). }

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:

RouteCert=Scope,Representation,FiberPolicy,Lift,Boundary,Dependencies,Version,Replay.\boxed{ RouteCert = \langle Scope, Representation, FiberPolicy, Lift, Boundary, Dependencies, Version, Replay \rangle. }

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:

Stmt,Scope,Deps,Route,Version,Fingerprint,Checker,Payload,Region,Status.\boxed{ \langle Stmt, Scope, Deps, Route, Version, Fingerprint, Checker, Payload, Region, Status \rangle. }

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:

Obst=Ψ(St,Lt).\boxed{ \mathsf{Obs}_t = \Psi(\mathbb S_t,\mathcal L_t). }

其輸出可以包括:

Dtresolve,Dtfrontier,Wt,σt,ht,χt,Dirtyt,Rt,D_t^{\rm resolve}, \quad D_t^{\rm frontier}, \quad W_t, \quad \sigma_t, \quad h_t, \quad \chi_t, \quad |Dirty_t|, \quad |R_t|,

以及:

SurvProft,Gt,RouteValuet,support multiplicity,certificate debt.SurvProf_t, \quad \mathbf G_t, \quad RouteValue_t, \quad support\ multiplicity, \quad certificate\ debt.

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:

ΩtGCCt.Valid.\boxed{ \Omega_t \quad\text{或}\quad GCC_t.Valid. }

這是安全架構原則:

measurementmathematical authority.\boxed{ \text{measurement} \neq \text{mathematical authority}. }

13. Runtime Pipeline

Phase-I reference pipeline:

DetectProfileGapExtractActionGenerateRouteProposeVerifyCoverageAuditBoundaryAuditGlueAuditCompileCommit.\boxed{ \begin{aligned} &\mathsf{Detect}\\ &\to\mathsf{Profile}\\ &\to\mathsf{GapExtract}\\ &\to\mathsf{ActionGenerate}\\ &\to\mathsf{Route}\\ &\to\mathsf{Propose}\\ &\to\mathsf{Verify}\\ &\to\mathsf{CoverageAudit}\\ &\to\mathsf{BoundaryAudit}\\ &\to\mathsf{GlueAudit}\\ &\to\mathsf{Compile}\\ &\to\mathsf{Commit}. \end{aligned} }

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:

proof checks⇏benchmark is sound.\boxed{ \text{proof checks} \not\Rightarrow \text{benchmark is sound}. }

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:

ProblemHash,StatementHash,AdapterVersion,RuntimeSchema,ReducerVersion,VerifierVersion,DependencySnapshot,RoutePolicyVersion,ModelVersion,HardwareProfile,Seed,AuditVersion.\boxed{ \begin{aligned} &ProblemHash,\\ &StatementHash,\\ &AdapterVersion,\\ &RuntimeSchema,\\ &ReducerVersion,\\ &VerifierVersion,\\ &DependencySnapshot,\\ &RoutePolicyVersion,\\ &ModelVersion,\\ &HardwareProfile,\\ &Seed,\\ &AuditVersion. \end{aligned} }

定義 benchmark identity:

BenchID=H(ProblemHashEnvHashSchemaHashPolicyHashSeed).\boxed{ BenchID = H( ProblemHash\Vert EnvHash\Vert SchemaHash\Vert PolicyHash\Vert Seed ). }

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:

0.\boxed{0.}

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 hh
  • closure-basis compression χ\chi
  • 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:

Dresolve,Dfrontier,W,σ,Gcompile,Pdrift.D^{\rm resolve}, D^{\rm frontier}, W, \sigma, G^{\rm compile}, P^{\rm drift}.

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:

same validated historysame runtime state.\boxed{ \text{same validated history} \to \text{same runtime state}. }

不證明

the history itself is mathematically valid.\boxed{ \text{the history itself is mathematically valid}. }

同樣:

incremental replay = full replay\boxed{ \text{incremental replay = full replay} }

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:

  1. statement fidelity audit;
  2. proof-kernel / checker version pinned;
  3. banned axioms / shortcuts policy declared;
  4. dependency graph audited;
  5. test harness replayable;
  6. environment fingerprint complete;
  7. hidden oracle 只在 toy / synthetic track 使用;
  8. 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:

GBG_B

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:

research artifactstyped candidatesverificationcertificate nodescommitted SDPE state.\boxed{ \text{research artifacts} \to \text{typed candidates} \to \text{verification} \to \text{certificate nodes} \to \text{committed SDPE state}. }

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:

  1. reference ledger replay 到 empty survivor + valid GCC;
  2. 15001500 組 random event histories deterministic replay;
  3. 15001500 組 checkpoint-tail equivalence;
  4. 15001500 組 hidden-oracle survivor soundness;
  5. proposal non-authority;
  6. telemetry non-authority;
  7. unverified-certificate commit rejection;
  8. redundant-support rollback precision;
  9. 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

  1. Event-Sourced Replay Determinism;
  2. Checkpoint–Tail Equivalence;
  3. Proposal Non-Authority under enforced mutation rules;
  4. Certificate-Gated Survivor Preservation;
  5. Precise Support-Aware Reopen under complete support index;
  6. Observatory Non-Authority;
  7. 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

  1. event ledger append-only / tamper-evident;
  2. reducer deterministic;
  3. canonical serialization stable;
  4. certificate checker identity / version pinned;
  5. dependency metadata complete;
  6. support index complete;
  7. authority boundaries enforced;
  8. 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:

P01 Survivor SoundnessP02 Representation FaithfulnessP03 Global Coverage / ClosureP04 Trace Compilation / Incremental ReplayP05 Discovery–Verification DynamicsP06 Exceptional-Core GeometryP07 Cost-Aware Enclosure RoutingP08 Runtime / Benchmark / Proof-Space Observatory.\boxed{ \begin{aligned} &\text{P01 Survivor Soundness}\\ &\downarrow\\ &\text{P02 Representation Faithfulness}\\ &\downarrow\\ &\text{P03 Global Coverage / Closure}\\ &\downarrow\\ &\text{P04 Trace Compilation / Incremental Replay}\\ &\downarrow\\ &\text{P05 Discovery--Verification Dynamics}\\ &\downarrow\\ &\text{P06 Exceptional-Core Geometry}\\ &\downarrow\\ &\text{P07 Cost-Aware Enclosure Routing}\\ &\downarrow\\ &\text{P08 Runtime / Benchmark / Proof-Space Observatory}. \end{aligned} }

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:

SDPE state 應由可重播 event history 重建,而不是依賴聊天記憶。\boxed{ \text{SDPE state 應由可重播 event history 重建,而不是依賴聊天記憶。} }

第二:

proposal、prediction、telemetry 都沒有 proof commit authority。\boxed{ \text{proposal、prediction、telemetry 都沒有 proof commit authority。} }

第三:

verified certificate + global obligations 才能改變 authoritative survivor state。\boxed{ \text{verified certificate + global obligations 才能改變 authoritative survivor state。} }

第四:

benchmark 必須同時驗證 proof、statement、coverage 與 environment fidelity。\boxed{ \text{benchmark 必須同時驗證 proof、statement、coverage 與 environment fidelity。} }

第五:

proof-space observability 應提高研究可理解性,但不能反過來成為數學真理來源。\boxed{ \text{proof-space observability 應提高研究可理解性,但不能反過來成為數學真理來源。} }

因此,空間域證明包圍第一階段現在形成一個完整迴圈:

DetectEncloseRouteProposeVerifyAuditCompileCommitObserveDetect again.\boxed{ \text{Detect} \to \text{Enclose} \to \text{Route} \to \text{Propose} \to \text{Verify} \to \text{Audit} \to \text{Compile} \to \text{Commit} \to \text{Observe} \to \text{Detect again}. }

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

  1. 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.
  2. 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.
  3. QuocViet Pham, Elvir Karimov, Andrey Galichin, Ivan Oseledets, TheoremBench: Evaluating LLMs on Theorem Proving in Formal Mathematics, arXiv:2606.09450, 2026.
  4. Matěj Kripner, Milan Straka, OpenProver: Agentic and Interactive Theorem Proving with Lean 4, arXiv:2607.09217, 2026.
  5. Austin Shen, Yunong Shi, Keep the Proof State Live: Snapshotting for Efficient Tactic Search in Lean 4, arXiv:2605.25556, 2026.
  6. Jongyoon Kim, Hojae Han, Seung-won Hwang, Benchmarking Testing in Automated Theorem Proving, arXiv:2604.23698, 2026.
  7. Prior SDPE artifacts: Papers 01--07.