← Archive
lm-004044 · 2026-09

UNPNP-II Phase B — Paper 04 — Formal Invariants and Verification Targets

下載 MD 檔 ⬇

UNPNP-II Phase B — Paper 04

Formal Invariants and Verification Targets

URR Runtime、Benchmark Protocol、Crystal Lifecycle 與 Sidecar 的形式化驗證目標

系列: UNPNP-II / Multi-Scale Computational Geometry — Phase B
篇次: Phase B Paper 04 / 04
作者: Neo.K with Aletheia(GPT)
機構: EveMissLab/一言諾科技有限公司
版本: v0.1
日期: 2026-09-09
文件性質: Formal Invariants/Property Testing/Model Checking Targets/Verification Handoff
前置: Phase B Paper 01–03;UNPNP-II v0.1;UNPNP-I;MWT;GCM
狀態: Canonical Draft


摘要

Phase B Paper 01 將 UNPNP-II 壓成 URR-v0.1 Reference Runtime,並提出 URR-1~URR-16;Paper 02 建立 A/B/C/D 實驗矩陣與 B-1~B-16 benchmark invariants;Paper 03 再把它們落成 Sidecar 工程規格、資料 schema、adapter contract、crystal lifecycle、A/B/C/D feature flags 與 Done Gate。

本文完成 Phase B 最後一步:

把這些 Runtime / Benchmark / Lifecycle 規則轉成可由測試框架、property testing、state-machine checking 與未來形式化工具直接接手的 Verification Targets。

本文將驗證分成四層:

L1 SchemaL2 Runtime PropertiesL3 State-Machine InvariantsL4 Formal Models.\boxed{ \text{L1 Schema} \rightarrow \text{L2 Runtime Properties} \rightarrow \text{L3 State-Machine Invariants} \rightarrow \text{L4 Formal Models}. }

L1 使用 type / schema / deterministic serialization / receipt hashing;
L2 使用 pytest、Hypothesis、differential tests、replay tests、fault injection;
L3 使用 lifecycle、authority、group isolation、invalidation propagation 等 state-machine property;
L4 留給 TLA+、Alloy、Lean、Coq、SMT 或 explicit-state model checking。

Phase B 的 final gate 不要求「證明整個 WRO 理論」,而要求:

任何 implementation 若自稱 URR-v0.1 conformant,至少不能違反這些可機器檢查的核心 invariants。


1. Verification Scope

本文驗證:

  • URR object consistency;
  • route legality;
  • authority preservation;
  • cost accounting;
  • crystal lifecycle;
  • invalidation / reopen;
  • A/B/C/D isolation;
  • frozen-model integrity;
  • experiment receipts;
  • replayability。

本文不宣稱形式化證明:

  • MWT World ontology;
  • GCM 完備性;
  • 24/72 完備性;
  • WRO global optimum 必然存在;
  • universal crystal compiler;
  • P/NP 結論。

2. Verification Layers

L1 — Schema

所有 canonical runtime object 必須:

serializable
versioned
id-bearing
null-explicit
enum-stable

核心規則:

Unknown0.\boxed{ \text{Unknown}\neq0. }

unknown 應序列化成 null 或 typed UNKNOWN,不可偷轉成 0。

Canonical JSON 必須:

UTF-8
sorted keys
no NaN/Infinity
stable enum strings
explicit null

同一 receipt 應滿足:

S1(x)=S2(x)S_1(x)=S_2(x)

以及:

H(S1(x))=H(S2(x)).H(S_1(x))=H(S_2(x)).

Hypothesis target:

@given(valid_receipts())
def test_receipt_hash_stable(receipt):
    assert canonical_hash(receipt) == canonical_hash(receipt)

3. URR-1 — World Primitive 與 Runtime State 分離

WRt.\boxed{ \mathbf W\neq\mathfrak R_t. }

工程上要求 WorldRecordRuntimeStateChartRecordRouteRecord 為不同 type;任何 Runtime tuple 都不得被命名或序列化成 MWT World primitive。


4. URR-2 — Computation / Observation / Materialization 分離

ComputationObservationMaterialization.\boxed{ \text{Computation} \neq \text{Observation} \neq \text{Materialization}. }

基本 property:

before = await adapter.snapshot()
_ = await adapter.observe(task)
after = await adapter.snapshot()
assert equivalent(before, after)

若 target 的 observation 本身有 side effect,adapter 必須顯式宣告。


5. URR-3 — Finite Active Frontier

Ft<.\boxed{ |F_t|<\infty. }
assert len(runtime.frontier.entries) <= settings.max_active_frontier

理論上可無界 refinement,不代表 Runtime 無界 materialization。


6. URR-4 — Unbounded Refinement ≠ Full Expansion

存在 inactive lower-scale worlds 時,Runtime 不得因其「存在」就全部 materialize。

驗證:

registered refinement depth > active refinement depth

仍為合法狀態。


7. URR-5 — Active Route 必須落在 Safe Feasible Space

RZtsafe.\boxed{ \mathcal R\subseteq Z_t^{\mathrm{safe}}. }

每個 route segment 至少必須:

reachable
authorized
guard-valid
resource-feasible
risk-valid

8. URR-6 — Reachable ≠ Authorized

ReachableAuthorized.\boxed{ \text{Reachable}\neq\text{Authorized}. }

Negative test:

reachable = true
authorized = false

必須產生:

AUTH_DENIED

而不是執行。


9. URR-7 — One Call ≠ One Computational Transition

one callone computational transition.\boxed{ \text{one call} \neq \text{one computational transition}. }

Wrapper illusion property:

baseline = run_raw_steps(10)
wrapped = run_wrapper(10)

assert wrapped.cost.hops < baseline.cost.hops
assert wrapped.cost.work == baseline.cost.work
assert wrapped.cost.causal_depth == baseline.cost.causal_depth

如果只 hop 下降,禁止宣稱 work-speedup 或 causal-speedup。


10. URR-8 — Speedup 必須標示維度

合法:

HOP
WORK
DEPTH
LATENCY
MATERIALIZATION
VERIFICATION
LIFECYCLE

非法:

speedup = 2.0
speedup_kind = null

11. URR-9 — Same Endpoint ≠ Same History

sfA=sfB⇏HA=HB.\boxed{ s_f^A=s_f^B \not\Rightarrow \mathcal H_A=\mathcal H_B. }

兩 route final state digest 相同但 history digest 不同時,只能標 endpoint-equivalent,不能自動升格成 full-route-equivalent。


12. URR-10 — Path Compilation 不能只是 Wrapper

Path Compilation candidate 至少需顯示一項 structural gain:

Δtop>0\Delta_{\mathrm{top}}>0

或:

Δcausal>0\Delta_{\mathrm{causal}}>0

或:

Δwork>0\Delta_{\mathrm{work}}>0

或明確的 reasoning / verification / materialization reduction。

若:

same route
same work
same depth
same materialization
same runtime
only interface changed

則標:

WRAPPER_ONLY

不是 compiled computational shortcut。


13. URR-11 — Compiled ≠ Crystallized

CompiledCrystallized.\boxed{ \text{Compiled}\neq\text{Crystallized}. }

禁止:

CANDIDATE → HOT

預設合法路徑:

CANDIDATE → COLD → WARM → HOT

Property:

with pytest.raises(InvalidLifecycleTransition):
    transition(CANDIDATE, HOT)

14. URR-12 — Crystal Authority 不得大於 Source Authority

Auth(κ)Auth(source).\boxed{ Auth(\kappa)\subseteq Auth(source). }

若 source 只有 READ,crystal 不得出現 WRITE。

對 composite crystal,v0.1 採保守規則:

Auth(κ)iAuth(sourcei).Auth(\kappa) \subseteq \bigcap_i Auth(source_i).

15. URR-13 — Promoted Crystal Record 完整性

任何 COLD/WARM/HOT crystal 必須有:

valid_domain
guard
validator
provenance
lifecycle
resource_envelope
source_route

缺任一核心欄位,不得 promotion。


16. URR-14 — Reopenability

每個 promoted crystal 必須:

reopen_pointer != null

或:

irreversible = true
reopen_loss_declared = true

非法:

reopen_pointer = null
irreversible = false

17. URR-15 — Switch Cost 不得隱形

scale / config / geometry / observer switch 均必須有 cost record。

即使實測為 0,也必須記:

value = 0
metric_kind = EXACT | ESTIMATED

而不是 missing。


18. URR-16 — Frozen-Model Integrity

θA=θB=θC=θD.\boxed{ \theta_A=\theta_B=\theta_C=\theta_D. }

對 paired receipt:

hashes = {r.model_signature_hash for r in pair}
assert len(hashes) == 1

不一致時:

INVALID_FOR_FROZEN_COMPARISON

19. B-1 ~ B-16 Benchmark Conformance

B-1 Same Initial Condition

assert len({r.initial_snapshot_digest for r in pair}) == 1

B-2 Meta Cost Counted

B/C/D:

cost.meta != null

B-3 Crystal Lifecycle Cost Counted

Clife=Ccompile+Crun+Cmaintain+Creopen.C_{\mathrm{life}} = C_{\mathrm{compile}} + C_{\mathrm{run}} + C_{\mathrm{maintain}} + C_{\mathrm{reopen}}.

B-4 Fast Path Counts Guard + Validation

不得把 fast path guard / cheap validator 當免費。

B-5 Unknown ≠ Zero

assert metric.kind == "UNKNOWN"
assert metric.value is None

B-6 Failed Runs Retained

Nreceipts=Nattempted runs.N_{\mathrm{receipts}} = N_{\mathrm{attempted\ runs}}.

B-7 Seed Pairing

benchmark_id + seed 需有 protocol 指定全部 groups。

B-8 Speedup Dimension Required

同 URR-8。

B-9 Held-Out Promotion

HOT 前:

held_out_successes >= threshold

B-10 Shift Boundary Explicit

shift_epoch != null

B-11 Authorization Violation Hard Failure

Nauth violation=0.N_{\mathrm{auth\ violation}}=0.

B-12 World Revision Triggers Validity Check

dependency-relevant world revision 改變後,affected crystal 不得繼續無條件 HOT。

B-13 Reopen Failure Observable

必須 append:

REOPEN_FAILED

B-14 D 不得偷用更強模型

同 URR-16。

B-15 Adapter Cost Reported

adapter 差異成本需顯式 report。

B-16 Structural Learning Must Point to External Delta

合法 structural delta:

ROUTE_COMPILED
CRYSTAL_PROMOTED
CONFIG_PROFILE_UPDATE
SCALE_PROFILE_UPDATE
BRIDGE_ADDED
ATLAS_UPDATED

如果 claimed structural learning 但 ledger 沒 structural event,claim invalid。


20. Crystal Lifecycle State Machine

狀態:

QK={CANDIDATE,COLD,WARM,HOT,STALE,RETIRED}.Q_K = \{ CANDIDATE, COLD, WARM, HOT, STALE, RETIRED \}.

預設 allowed transitions:

CANDIDATE → COLD
COLD → WARM
WARM → HOT

COLD → STALE
WARM → STALE
HOT → STALE

STALE → COLD
STALE → RETIRED

CANDIDATE → RETIRED
COLD → RETIRED
WARM → RETIRED
HOT → RETIRED

v0.1 不允許:

STALE → HOT

直接回升。

Fast invoke 只允許:

WARM
HOT

且還必須同時:

GuardValidAuthorityValidWorldCompatible.GuardValid \land AuthorityValid \land WorldCompatible.

21. Invalidation Propagation

若:

κcDeps(κp)\kappa_c\in Deps(\kappa_p)

且 child stale,則 parent 必須獲得 revalidation obligation:

STALE(κc)NeedsRevalidation(κp).\boxed{ STALE(\kappa_c) \Rightarrow NeedsRevalidation(\kappa_p). }

禁止 silent parent HOT。


22. Reopen State Machine

QR={CLOSED,REOPENING,OPEN,FAILED}.Q_R = \{ CLOSED, REOPENING, OPEN, FAILED \}.

若 source 存在且 adapter 支援所需 replay level:

REOPENING → OPEN

若 source 不存在:

REOPENING → FAILED

並產生 structured failure receipt。


23. Scale State Machine

QΣ={COARSE,MESO,FINE}.Q_\Sigma = \{ COARSE, MESO, FINE \}.

預設:

COARSE ↔ MESO ↔ FINE

若允許:

COARSE → FINE

direct jump,則必須標 stronger cross-scale validation。


24. Scale Thrashing

在 window ww 內,如果 coarse/fine reversal 超過 threshold:

SCALE_THRASHING

需被記錄。

這是 performance / policy warning,不一定是 correctness failure。


25. Configuration State Machine

v0.1:

QC={S,J,P,R}.Q_C = \{S,J,P,R\}.

若:

cicj,c_i\neq c_j,

必須存在:

bridge_id

或:

bridge_kind = IDENTITY_VALIDATED

26. S → P Safety Property

只有 independence / synchronization obligation 已驗證,才可:

SP.S\rightarrow P.

Hidden dependency test:

a3 → a7

若 router 仍標所有 task fully independent,驗證失敗。


27. R → S Fallback

Recognition fast path guard fail / stale 時:

RSR\rightarrow S

canonical fallback 應合法。


28. Group Isolation

A/B/C/D 各自有獨立 adaptive namespace:

NA,NB,NC,ND.N_A,N_B,N_C,N_D.

對:

iji\neq j

要求 adaptive store 不互相讀寫。

可共享的只有:

  • model signature;
  • initial snapshot;
  • benchmark spec;
  • read-only external data。

Crystal store、route profile、config profile、scale profile 必須 group-isolated。


29. Replay Invariants

Replay receipt 至少需要:

world_snapshot
runtime_version
adapter_version
model_signature
route
config_history
scale_history
crystal_ids
seed

Replay levels:

SEMANTIC
STATE
ROUTE
TIMING
EXACT

claim 不得超過 adapter capability。


30. Cost Ledger Invariants

每個 metric 具有:

value
kind = EXACT | PROXY | ESTIMATED | UNKNOWN

Aggregation 不可一律 sum。

建議:

W=iWi.W=\sum_iW_i.

因果深度由 dependency graph longest path;

wall time 用 measured elapsed;

risk 採 policy-specific aggregation;

observation loss 採 bridge-specific aggregation。

核心規則:

Unknown+x=Unknown\boxed{ Unknown+x=Unknown }

對不可推導 metric,而不是偷變成 xx


31. Crystal Utility Gate

Uκ=Bruntime+BfutureCcompileCverifyCmaintainCrisk.U_\kappa = B_{\mathrm{runtime}} + B_{\mathrm{future}} - C_{\mathrm{compile}} - C_{\mathrm{verify}} - C_{\mathrm{maintain}} - C_{\mathrm{risk}}.

若:

Uκ0,U_\kappa\le0,

不得升 HOT。

注意:這是 v0.1 promotion policy target,不是普世數學定理。


32. Guard False Accept / False Reject

False Accept:

guard pass,但 crystal semantics fail。

這是 correctness failure,通常應立即 stale。

False Reject:

guard fail,但 crystal 其實可用。

主要是 performance loss。

兩者不可混為同一錯誤率。


33. Differential Validation

若 canonical route 可執行:

Result(κ)qResult(Γcanonical).Result(\kappa) \simeq_q Result(\Gamma_{\mathrm{canonical}}).

對 benchmark-defined relevant invariants:

Iq.I_q.

Exact:

Iq(sfκ)=Iq(sfcanonical).I_q(s_f^\kappa)=I_q(s_f^{\mathrm{canonical}}).

Approximate:

d(Iq(sfκ),Iq(sfcanonical))ϵq.d( I_q(s_f^\kappa), I_q(s_f^{\mathrm{canonical}}) ) \le \epsilon_q.

34. History-Sensitive Validation

如果 task 定義:

required_events
forbidden_events
ordering_constraints

即使 endpoint 相同,也必須驗 history obligations。


35. Hypothesis vs Invariant

這是 Phase B 最重要的方法學界線之一。

Invariant

違反代表 implementation 不 conform。

例如:

  • unauthorized execution;
  • stale fast invoke;
  • unknown → zero;
  • failed run dropped。

Hypothesis

可以在實驗中失敗。

例如:

E[JD]<E[JA],E[J_D]<E[J_A], RK(t),R_K(t)\uparrow, RegretC(t).Regret_C(t)\downarrow.

不得把 performance hypothesis 寫成 unit-test success condition。


36. Benchmark Conformance ≠ Theory Success

一個 URR implementation 可以 100% conform,但 D 組完全沒有 performance gain。

這仍是一個有效科學結果。


37. Formal Model Target 1 — Crystal Lifecycle

適合 TLA+ / Alloy。

Safety:

state{WARM,HOT}¬FastInvoke.\boxed{ state\notin\{WARM,HOT\} \Rightarrow \neg FastInvoke. }

Safety 2:

FastInvokeGuardValidAuthorityValidWorldCompatible.\boxed{ FastInvoke \Rightarrow GuardValid \land AuthorityValid \land WorldCompatible. }

Liveness candidate:

若 STALE 且 source available,在公平條件下 eventually:

revalidated
or retired

38. Formal Model Target 2 — Authority

適合 Alloy / SMT / Lean。

Auth(κ)Auth(source).\boxed{ Auth(\kappa)\subseteq Auth(source). }

Composite:

Auth(κ)iAuth(sourcei).Auth(\kappa) \subseteq \bigcap_iAuth(source_i).

39. Formal Model Target 3 — Commit Protocol

State-changing action:

PLAN
→ AUTHORIZE
→ EXECUTE
→ VALIDATE
→ COMMIT

Safety:

ValidationFail¬CleanCommit.\boxed{ ValidationFail \Rightarrow \neg CleanCommit. }

若 target 不可 rollback,也必須留下:

COMMIT_FAILED_WITH_SIDE_EFFECT

而非 clean success。


40. Formal Model Target 4 — Group Isolation

對 adaptive stores:

ijWrite(Ni)ReadWrite(Nj)=.i\neq j \Rightarrow Write(N_i)\cap ReadWrite(N_j)=\varnothing.

41. Formal Model Target 5 — Frozen Model

paired experiment receipts 的 model signature 必須相同。


42. Formal Model Target 6 — Bounded Frontier

FtBF.|F_t|\le B_F.

可 model-check bounded toy model。


43. Formal Model Target 7 — Invalidation Propagation

dependency DAG 中 stale child 必須產生 parent revalidation obligation。


44. Formal Model Target 8 — Receipt Evidence Chain

receipt parent hash chain 不得無故斷裂。


45. Formal Model Target 9 — No Silent Failure

每個:

  • authorization denial;
  • validation failure;
  • reopen failure;
  • adapter failure;

最終必須 append structured event。


46. Formal Model Target 10 — Revealed-Set Optimality

對有限:

F^t\widehat{\mathfrak F}_t

可用 SMT 驗:

J(x\*)J(xi)J(x^\*)\le J(x_i)

對所有 revealed candidate。

這只能宣稱:

best-known in revealed candidate set

不能宣稱 omniscient global optimum。


47. Lean Target — Typed Unknown Metric

可定義:

Metric(α)=Known(α)Unknown.Metric(\alpha) = Known(\alpha) \mid Unknown.

加法:

Known(a)+Known(b)=Known(a+b),Known(a)+Known(b)=Known(a+b), Unknown+x=Unknown.Unknown+x=Unknown.

這可從 type level 防止 unknown 被默認 0。


48. Lean Target — PromotedCrystal Constructor

讓 promoted crystal type constructor 強制要求:

valid_domain
guard
validator
provenance
reopenability declaration

比 runtime null check 更強。


49. Lean Target — Route Equivalence Levels

明確區分:

EndpointEquivalent
TaskEquivalent
HistoryEquivalent
FullRouteEquivalent

禁止 endpoint equivalence 自動 coercion 成 full route equivalence。


50. Suggested TLA+ Modules

CrystalLifecycle.tla
RuntimeCommit.tla
Authority.tla
GroupIsolation.tla
Invalidation.tla

Suggested Lean:

Metric.lean
Authority.lean
Lifecycle.lean
RouteEquivalence.lean
CrystalRecord.lean

Suggested Hypothesis:

test_prop_cost.py
test_prop_lifecycle.py
test_prop_authority.py
test_prop_invalidation.py
test_prop_receipts.py
test_prop_isolation.py

51. Fault Injection Matrix

Fault Expected behavior
WORLD_DRIFT affected crystal stale / revalidate
POLICY_DRIFT authority recheck
GUARD_BUG differential validator catches false accept where possible
INVALID_INDEX J/R shortcut validation failure
MISSING_PROVENANCE promotion reject or audit fail
REOPEN_SOURCE_MISSING REOPEN_FAILED
HIDDEN_DEPENDENCY S→P blocked or validator fail
ADAPTER_FAILURE structured error receipt

52. Sidecar Done Gate → DG-1 ~ DG-16

DG-1

Install succeeds.

DG-2

Unit tests pass.

DG-3

URR-1~16 pass.

DG-4

B-1~16 benchmark conformance pass.

DG-5

SW-01 / SW-04 / SW-07 / SW-08 / SW-10 runnable.

DG-6

A/B/C/D isolation pass.

DG-7

Frozen-model signature gate pass.

DG-8

Crystal stale / reopen pass.

DG-9

Wrapper illusion test pass.

DG-10

Config switch cost recorded.

DG-11

Meta cost recorded.

DG-12

Receipt hash reproducible.

DG-13

Authorization hard-fail pass.

DG-14

Failed runs retained.

DG-15

Unknown metrics remain unknown.

DG-16

Canonical fallback works.


53. Conformance Levels

建議:

URR-C0 — Schema

資料模型可互通。

URR-C1 — Runtime

URR 核心 invariants pass。

URR-C2 — Structural Adaptation

Crystal lifecycle + replay + reopen。

URR-C3 — Benchmark

A/B/C/D conformance。

URR-C4 — Formalized Core

核心 state-machine 有 TLA+ / Lean / Alloy / SMT 等部分形式化證據。

重要:

ConformantFaster.\boxed{ \text{Conformant} \neq \text{Faster}. }

Performance 仍是實驗結果。


54. Release Verification Artifacts

每個 URR release 建議輸出:

TEST_REPORT.md
INVARIANT_REPORT.json
BENCHMARK_CONFORMANCE.json
SCHEMA_HASHES.json
RUNTIME_VERSION.json

若有形式化:

tla/
lean/
alloy/
smt/

55. Clean Replay Gate

Release 前至少:

  1. 空 DB;
  2. 載入 canonical receipts;
  3. 重建必要 runtime indices;
  4. 重播指定 Synthetic benchmark;
  5. final state / semantic result 符合 declared replay level。

這可防止 hidden local cache 讓「重現」看起來成功。


56. Reproducible Package

release package 應包含:

  • source;
  • dependency lock;
  • benchmark manifests;
  • seed list;
  • schema version;
  • runtime version;
  • theory version;
  • SHA-256。

57. Verification Before Claim

tests green 但:

  • cost receipts 缺失;
  • failed runs 被丟棄;
  • model signature 不固定;

都不能宣稱 Phase B benchmark complete。

同樣,部分 TLA+ / Lean 證明不能被擴張成:

「UNPNP-II 已被數學證明。」

合法說法應是:

「URR-12 authority monotonicity 已在 model X 中形式驗證。」


58. Runtime Safety Kernel — First Formalization Batch

最優先五條:

  1. no unauthorized execution;
  2. stale crystal cannot fast execute;
  3. no direct candidate → HOT;
  4. crystal authority cannot exceed source;
  5. validation failure cannot become clean commit。

它們小、清楚、高價值,最適合先 model check。


59. Second Formalization Batch

  1. unknown ≠ zero;
  2. A/B/C/D group isolation;
  3. frozen model equality;
  4. invalidation propagation;
  5. reopen failure observability。

60. Third Formalization Batch

  1. bounded active frontier;
  2. scale transition constraints;
  3. configuration bridge constraints;
  4. route equivalence levels;
  5. receipt hash chain;
  6. revealed-set optimality。

61. Canonical Development Order

建議 implementation workflow:

Schema RED/GREEN
→ Ledger RED/GREEN
→ URR invariant suite
→ Synthetic adapter
→ A/B
→ Path compiler
→ Crystal C
→ Scale/Config D
→ Benchmark matrix
→ TLA+/Lean subset

62. Theory / Runtime Conflict Rule

如果 implementation 與 Paper 03 / 04 衝突:

  1. 先判斷 implementation bug;
  2. 如果規格本身不可行,再 revision theory paper;
  3. 禁止 silent semantic mutation。

Theory revision:

Phase B Paper 04 v0.2

Runtime revision:

URR v0.2

Benchmark receipt 必須綁:

theory_version
runtime_version
benchmark_version

63. Phase B Final Core Laws

 An optimization that cannot be measured is not a validated optimization. \boxed{ \textbf{ An optimization that cannot be measured is not a validated optimization. } }  A fast path that cannot be invalidated is not a safe crystal. \boxed{ \textbf{ A fast path that cannot be invalidated is not a safe crystal. } }  A primitive that cannot expose its provenance is not an auditable earned primitive. \boxed{ \textbf{ A primitive that cannot expose its provenance is not an auditable earned primitive. } }  A benchmark that changes the model while claiming architecture gain is invalid for frozen-model comparison. \boxed{ \textbf{ A benchmark that changes the model while claiming architecture gain is invalid for frozen-model comparison. } }  A runtime theory becomes experimentally meaningful only when its failure conditions are machine-checkable. \boxed{ \textbf{ A runtime theory becomes experimentally meaningful only when its failure conditions are machine-checkable. } }

64. Phase B Final Closure

Phase B 四篇形成:

Formal RuntimeBenchmark ProtocolImplementation SpecVerification Targets.\boxed{ \text{Formal Runtime} \rightarrow \text{Benchmark Protocol} \rightarrow \text{Implementation Spec} \rightarrow \text{Verification Targets}. }

Paper 01 回答:

Runtime 最小物件與狀態是什麼?

Paper 02 回答:

怎麼公平比較?

Paper 03 回答:

怎麼實作?

Paper 04 回答:

怎麼知道實作沒有偷偷破壞理論邊界?

最終:

URR-v0.1=State+Receipts+Invariants+Benchmarks+Replay.\boxed{ \text{URR-v0.1} = \text{State} + \text{Receipts} + \text{Invariants} + \text{Benchmarks} + \text{Replay}. }

結論

UNPNP-II 的理論核心可以很大:

  • World;
  • Atlas;
  • Geometry;
  • Scale;
  • 24/72;
  • Route;
  • Crystal;
  • World-Relative Optimization。

但真正進入工程後,首先需要的不是更大的 optimizer,而是一個小而嚴格的 correctness kernel。

這個 kernel 必須知道:

Unknown0,\boxed{ \text{Unknown}\neq0, } ReachableAuthorized,\boxed{ \text{Reachable}\neq\text{Authorized}, } CompiledCrystallized,\boxed{ \text{Compiled}\neq\text{Crystallized}, } Same Endpoint⇏Same History,\boxed{ \text{Same Endpoint}\not\Rightarrow\text{Same History}, }

以及:

One CallOne Computation.\boxed{ \text{One Call}\neq\text{One Computation}. }

因此 Phase B 最後的核心不是:

證明 UNPNP-II 一定能加速所有世界。

而是:

 建立一套讓任何自稱 UNPNP-II Runtime 的實作, 都必須先通過的可機器檢查邊界。 \boxed{ \textbf{ 建立一套讓任何自稱 UNPNP-II Runtime 的實作, 都必須先通過的可機器檢查邊界。 } }

如果 implementation 通過所有 invariants,但 benchmark 顯示沒有 performance gain,那仍是一個成功的科學結果:

這表示理論在該 workload 上沒有收益,而不是實驗失敗。

這正是整個 Phase B 最重要的收束:

 先讓系統有資格失敗, 再討論它是否真的成功。 \boxed{ \textbf{ 先讓系統有資格失敗, 再討論它是否真的成功。 } }