← Archive
lm-003584 · 2026-09

APSC Runtime Technical Architecture v0.1

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

APSC Runtime Technical Architecture v0.1

Adaptive Possibility-Space Cognition Runtime — Canonical Engineering Architecture, Contracts, Control Loop, and MVP Roadmap

文件類型:技術白皮書 01 / 02
系列:Adaptive Possibility-Space Cognition(APSC)
版本:v0.1
日期:2026-08-24
作者:Neo.K
機構:EveMissLab / EVEMISS Technology


0. 文件定位

本文件是 APSC 六篇理論論文的第一份統一工程母文件。

它不新增平行母理論,而是把下列理論物件轉成可執行 Runtime:

  1. Constrained Possibility Space;
  2. Multi-Scale Future Expansion;
  3. Counterfactual Observation Expansion(COE);
  4. Finite Cognitive Spacetime;
  5. Adaptive Computation Control;
  6. Cognitive Operating Profile(COP);
  7. Replayable Cognitive Evidence。

本文件的直接工程目標是:

將「AI 怎麼想」從隱式行為轉成可定址、可約束、可調度、可重播的 Runtime 狀態。\boxed{ \text{將「AI 怎麼想」從隱式行為轉成可定址、可約束、可調度、可重播的 Runtime 狀態。} }

1. 一句話架構

APSC Runtime 的第一版可以收斂為:

Neural Adapter+State Kernel+Possibility Graph+Constraint Engine+Operator Runtime+COE Engine+Budget Controller+COP Manager+Receipt / Replay Layer\boxed{ \text{Neural Adapter} + \text{State Kernel} + \text{Possibility Graph} + \text{Constraint Engine} + \text{Operator Runtime} + \text{COE Engine} + \text{Budget Controller} + \text{COP Manager} + \text{Receipt / Replay Layer} }

其核心控制流為:

ObservationStatePossibilitySpaceCandidateCognitionValueEstimateBudgetGateExecuteReceiptUpdateStopOrContinue.Observation \rightarrow State \rightarrow PossibilitySpace \rightarrow CandidateCognition \rightarrow ValueEstimate \rightarrow BudgetGate \rightarrow Execute \rightarrow Receipt \rightarrow Update \rightarrow StopOrContinue.

2. 核心工程不變量

2.1 Neural Core 不是 Runtime State

神經模型只是一個:

ProposalEngine.ProposalEngine.

它可以提出:

  • 候選狀態;
  • 候選分支;
  • 候選觀察;
  • 候選策略;
  • 候選抽象;
  • 候選驗證方式。

但不能直接改寫 canonical state。

因此:

NeuralProposalCanonicalMutation.\boxed{ NeuralProposal \neq CanonicalMutation. }

2.2 Observation 不等於 Truth

任何外部觀察都先形成:

ObservationRecord.ObservationRecord.

不得直接等於:

WorldTruth.WorldTruth.

因此:

ObservationTruth.\boxed{ Observation \neq Truth. }

2.3 Candidate 不等於 Active Branch

候選分支進入 Runtime 後,必須通過約束與資源檢查。

CandidateBranchActiveBranch.CandidateBranch \neq ActiveBranch.

2.4 Decision 不等於 Commit

控制器提出決策後,仍可能因:

  • verifier;
  • budget;
  • safety;
  • user preference;
  • deadline;
  • rule;

被拒絕或延後。

因此:

DecisionCommit.\boxed{ Decision \neq Commit. }

2.5 Profile Label 不等於 Canonical Policy

UI 顯示:

Fast
Accurate
Research
Safe
Explore

只是 preset。

Runtime 必須保存:

Θ.\Theta.

所以:

ModeLabelCanonicalCOP.ModeLabel \neq CanonicalCOP.

3. Runtime 分層

第一版分成九層:

L0  External / Neural Adapters
L1  Observation & Input Normalization
L2  Canonical State Kernel
L3  Possibility Graph & Constraint Engine
L4  Cognitive Operator Runtime
L5  COE / Active Observation Engine
L6  Adaptive Computation Controller
L7  Cognitive Operating Profile Manager
L8  Receipt / Replay / Audit Layer

4. L0:External / Neural Adapters

4.1 Neural Adapter

每個模型供應商只需實作統一介面:

propose_state()
propose_branches()
propose_observations()
propose_abstraction()
propose_action()

Runtime 不依賴某一特定模型。


4.2 Tool Adapter

外部工具統一為:

observe()
verify()
simulate()
execute()

每個工具必須宣告:

  • input schema;
  • output schema;
  • cost estimate;
  • latency class;
  • side-effect class;
  • retry policy;
  • trust / reliability metadata。

5. L1:Observation & Input Normalization

5.1 ObservationRecord

第一版 canonical object:

{
  "observation_id": "obs://...",
  "source": "sensor|tool|user|model|world",
  "timestamp": "...",
  "payload": {},
  "reliability": 0.0,
  "cost": {},
  "side_effect": "none|possible|confirmed",
  "provenance": {}
}

5.2 Observation Status

觀察狀態:

RAW
NORMALIZED
SUPPORTED
CONTRADICTED
UNRESOLVED
SUPERSEDED

6. L2:Canonical State Kernel

6.1 StateObject

核心狀態:

St=(Xt,Rt,Ct,Ot,Bt).S_t = ( X_t, R_t, C_t, O_t, B_t ).

工程表示可拆為:

StateSnapshot
EntityState
RuleState
CausalState
ObservationState
BudgetState

6.2 Unknown 必須顯式表示

欄位不可用:

null

隨意混合:

  • unknown;
  • unavailable;
  • not applicable;
  • not observed。

應使用:

KNOWN
UNKNOWN
UNCERTAIN
NOT_APPLICABLE
UNAVAILABLE

6.3 State Versioning

每次 canonical mutation:

StSt+1S_t \rightarrow S_{t+1}

必須產生:

state_version
parent_state_version
mutation_receipt

7. L3:Possibility Graph

7.1 PossibilityNode

每個可能節點至少包含:

{
  "node_id": "poss://...",
  "parent_ids": [],
  "state_ref": "state://...",
  "horizon": 0,
  "resolution": "near|mid|far",
  "probability": null,
  "risk": null,
  "decision_value": null,
  "reachability": "supported|contradicted|undetermined",
  "status": "candidate|active|dormant|pruned|merged|committed"
}

7.2 PossibilityEdge

邊表示:

ACTION
EVENT
OBSERVATION_UPDATE
COUNTERFACTUAL_INTERVENTION
ABSTRACTION
REFINEMENT
MERGE

7.3 Graph 不要求樹

當兩條路徑進入等價狀態:

SaGSb,S_a \sim_G S_b,

可以執行:

Merge.Merge.

所以 Runtime 使用 DAG / general graph,而不是固定 tree。


8. Constraint Engine

8.1 Rule Classes

HARD
SOFT
PROBABILISTIC
UNKNOWN

8.2 ConstraintResult

{
  "constraint_id": "rule://...",
  "target": "poss://...",
  "status": "pass|fail|undetermined",
  "severity": "hard|soft",
  "reason_code": "...",
  "evidence_refs": []
}

8.3 Constraint Projection

候選空間:

Ω~\widetilde{\Omega}

經過:

ΠK\Pi_{\mathcal K}

得到:

Ω^.\widehat{\Omega}.

工程流程:

candidate
→ hard-rule check
→ causal compatibility
→ reachability
→ risk annotation
→ budget relevance
→ active / dormant / prune

9. Reachability Engine

第一版不要求通用規劃器。

只提供統一 contract:

check_reachability(state, target, horizon, budget)

回傳:

SUPPORTED
CONTRADICTED
UNDETERMINED

及可選:

witness_path
blocking_constraints
required_resources

10. Multi-Scale Horizon Manager

10.1 三層 Horizon

NEAR
MID
FAR

10.2 Resolution Policy

一般規則:

hResolution.h\uparrow \Rightarrow Resolution\downarrow.

但可局部 refine。


10.3 Near

保留:

  • explicit state;
  • explicit action;
  • local causal transition。

10.4 Mid

轉為:

ScenarioFamily

10.5 Far

轉為:

Attractor
Region
MacroState

11. L4:Cognitive Operator Runtime

11.1 Operator Registry

第一版 operator:

expand
counterfactual_intervene
counterfactual_observe
observe
verify
prune
merge
abstract
refine
backtrack
reframe
profile_shift
commit

11.2 OperatorSpec

每個 operator 必須宣告:

{
  "operator_id": "cog://expand@0.1",
  "input_types": [],
  "output_types": [],
  "cost_model": {},
  "side_effect_class": "none",
  "preconditions": [],
  "postconditions": [],
  "interruptible": true,
  "checkpointable": true
}

12. Operator Invocation

任何認知操作都建立:

CognitiveOperation

包含:

operation_id
operator_id
state_ref
goal_ref
profile_ref
budget_before
estimated_value
estimated_cost
status

13. Operator Status

PROPOSED
QUEUED
RUNNING
PAUSED
PREEMPTED
COMPLETED
FAILED
CANCELLED

14. L5:COE Engine

14.1 Candidate Observation Generation

Q:(Bt,Gt,Ω^t)Qt.\mathfrak Q : (B_t,G_t,\widehat{\Omega}_t) \rightarrow \mathcal Q_t.

工程輸出:

ObservationCandidate[]

14.2 ObservationCandidate

{
  "query_id": "query://...",
  "target": "...",
  "expected_outcomes": [],
  "estimated_information_gain": null,
  "estimated_decision_value": null,
  "estimated_risk_reduction": null,
  "estimated_cost": {},
  "observer_effect": "none|possible|likely"
}

15. COE Expansion

對觀察 qq

OqCF(Ω^t)={(oi,pi,Ω^toi)}.\mathfrak O_q^{CF} ( \widehat{\Omega}_t ) = \{ (o_i,p_i,\widehat{\Omega}_{t|o_i}) \}.

第一版允許:

  • discrete outcomes;
  • bounded continuous bins;
  • symbolic outcomes。

16. Observation Value

統一值函數:

Vobs(q)=wIIG(q)+wDDROV(q)+wRRiskReduction(q)+wCCompression(q)wKCost(q).V_{obs}(q) = w_I IG(q) + w_D DROV(q) + w_R RiskReduction(q) + w_C Compression(q) - w_K Cost(q).

權重由 COP 提供。


17. L6:Adaptive Computation Controller

17.1 Candidate Cognitive Queue

每輪建立:

Utcandidate.\mathcal U_t^{candidate}.

然後過濾:

Utfeasible={uC(u)Bt}.\mathcal U_t^{feasible} = \{ u | C(u)\preceq B_t \}.

17.2 Value Estimator

每個候選:

ΔQ^(u).\widehat{\Delta Q}(u).

計算:

Score(u)=ΔQ^(u)λC(u).Score(u) = \widehat{\Delta Q}(u) - \lambda^\top C(u).

18. Marginal Value of Cognition

MVC(u)=ΔQ^(u)ScalarCost(u).MVC(u) = \frac{ \widehat{\Delta Q}(u) }{ ScalarCost(u) }.

MVC 是第一版 scheduler 的主要排序訊號之一。


19. Dynamic Shadow Price

每種資源:

λi=f(Bi,Deadline,COP,Risk).\lambda_i = f( B_i, Deadline, COP, Risk ).

當:

Bi,B_i\downarrow,

通常:

λi.\lambda_i\uparrow.

20. Scheduler

第一版 scheduler:

filter infeasible
→ apply hard gates
→ score candidates
→ urgency adjustment
→ risk adjustment
→ choose top action

21. Preemption

若新操作:

Score(ub)Score(ua)>τp,Score(u_b)-Score(u_a)>\tau_p,

可以搶佔:

ua.u_a.

前提:

interruptible = true

22. Checkpoint

可搶佔操作應支援:

{
  "checkpoint_id": "ckpt://...",
  "operation_id": "...",
  "state_ref": "...",
  "open_branches": [],
  "evidence_refs": [],
  "resume_token": "..."
}

23. Stop Policy

主動停止:

maxuScore(u)ϵ.\max_u Score(u) \leq \epsilon.

則:

STOP_COGNITION

24. Commit Policy

提交至少考慮:

StabilityStability RiskRisk ExpectedFurtherGain.ExpectedFurtherGain.

第一版:

Commit=1Commit=1

若:

Stabilityτs,Stability\geq\tau_s, Riskτr,Risk\leq\tau_r, ExpectedFurtherGainτg.ExpectedFurtherGain\leq\tau_g.

25. L7:COP Manager

25.1 Canonical COP

第一版十維:

Θ=(θd,θb,θv,θcf,θo,θr,θl,θa,θc,θn).\Theta = ( \theta_d, \theta_b, \theta_v, \theta_{cf}, \theta_o, \theta_r, \theta_l, \theta_a, \theta_c, \theta_n ).

25.2 COP Object

{
  "cop_id": "cop://...",
  "depth": 0.5,
  "breadth": 0.5,
  "verification": 0.5,
  "counterfactual": 0.5,
  "observation": 0.5,
  "risk": 0.5,
  "latency": 0.5,
  "abstraction": 0.5,
  "commitment": 0.5,
  "novelty": 0.5
}

26. COP Sources

有效 Profile 由:

user preference
task profile
system hard constraints
adaptive proposal

合成。


27. Adaptive Region

AI 自調範圍:

A=[Θmin,Θmax].\mathcal A = [\Theta_{min},\Theta_{max}].

28. Profile Shift

重要 shift 必須:

PROPOSE
→ PROJECT_TO_FEASIBLE_REGION
→ APPLY
→ RECEIPT

29. Hysteresis

避免頻繁切換:

ΔTrigger>τh.|\Delta Trigger| > \tau_h.

否則不 shift。


30. L8:Receipt Layer

30.1 DecisionReceipt

{
  "receipt_id": "receipt://decision/...",
  "state_ref": "...",
  "selected_operation": "...",
  "alternatives": [],
  "budget_before": {},
  "budget_after": {},
  "cop_ref": "...",
  "commit": false
}

30.2 ProfileReceipt

{
  "receipt_id": "receipt://profile/...",
  "old_cop": "...",
  "new_cop": "...",
  "trigger": "...",
  "reason_code": "...",
  "authorized_region": {}
}

30.3 ObservationReceipt

保存:

query
expected outcomes
actual outcome
estimated value
realized value
cost

31. Replay Layer

Runtime 應輸出:

run_manifest.json
events.jsonl
operations.jsonl
observations.jsonl
budgets.jsonl
profiles.jsonl
receipts.jsonl
final_state.json

32. Private Reasoning Boundary

Runtime 不需要保存:

private chain-of-thought
hidden model scratchpad

需要保存的是:

selected operator
public justification code
state transition
budget transition
external observation
receipt

因此:

Replayability⇏PrivateReasoningDisclosure.Replayability \not\Rightarrow PrivateReasoningDisclosure.

33. Canonical Runtime Objects

v0.1 最小 canonical objects:

RunManifest
Goal
ObservationRecord
StateSnapshot
Rule
ConstraintResult
PossibilityNode
PossibilityEdge
CognitiveOperation
ObservationCandidate
BudgetState
CognitiveOperatingProfile
Checkpoint
DecisionReceipt
ProfileReceipt
ObservationReceipt
FinalCommit

34. ID Namespace

建議:

run://
goal://
obs://
state://
rule://
poss://
cog://
query://
budget://
cop://
ckpt://
receipt://
commit://

35. Runtime State Machine

主 Runtime 狀態:

INIT
OBSERVE
STATE_BUILD
GENERATE
CONSTRAIN
ROUTE_COGNITION
EXECUTE
UPDATE
EVALUATE_STOP
COMMIT
DONE
FAILED

36. 主循環

INIT
→ ingest goal / COP / budget
→ OBSERVE
→ normalize observations
→ STATE_BUILD
→ construct canonical state
→ GENERATE
→ propose possibilities
→ CONSTRAIN
→ project to working possibility space
→ ROUTE_COGNITION
→ select operator
→ EXECUTE
→ UPDATE
→ emit receipts
→ EVALUATE_STOP
→ continue / commit

37. API Surface

第一版建議:

POST /runs
GET  /runs/{id}
POST /runs/{id}/observe
POST /runs/{id}/step
POST /runs/{id}/pause
POST /runs/{id}/resume
POST /runs/{id}/commit
GET  /runs/{id}/state
GET  /runs/{id}/possibilities
GET  /runs/{id}/budget
GET  /runs/{id}/profile
PATCH /runs/{id}/profile
GET  /runs/{id}/receipts
GET  /runs/{id}/replay

38. Step Contract

POST /runs/{id}/step

輸入:

{
  "max_operations": 1,
  "allow_external_observation": true
}

輸出:

{
  "selected_operation": "...",
  "status": "...",
  "state_ref": "...",
  "budget_ref": "...",
  "receipt_ref": "..."
}

39. Budget Contract

{
  "wall_ms": 10000,
  "compute_units": 100,
  "memory_units": 50,
  "observation_units": 10,
  "external_units": 5,
  "risk_ceiling": 0.2
}

40. COP Presets

Preset 只是 convenience:

single_turn_accuracy
exploration
low_latency
high_risk_assurance
research
balanced

Runtime 仍保存展開後的完整 Θ\Theta


41. Failure Model

41.1 State Failure

STATE_UNRESOLVED
STATE_CONTRADICTION
STATE_SCHEMA_ERROR

41.2 Constraint Failure

HARD_RULE_FAIL
CAUSAL_CONFLICT
REACHABILITY_FAIL

41.3 Cognitive Failure

OPERATOR_FAILED
OPERATOR_TIMEOUT
CHECKPOINT_FAILED
PREEMPTION_FAILED

41.4 Observation Failure

OBSERVATION_UNAVAILABLE
OBSERVATION_UNRELIABLE
OBSERVER_EFFECT_BLOCKED

41.5 Budget Failure

BUDGET_EXHAUSTED
DEADLINE_REACHED
EXTERNAL_QUOTA_EXHAUSTED

42. Graceful Degradation

當資源下降:

reduce breadth
→ raise abstraction
→ reduce far-horizon detail
→ defer low-value verify
→ freeze dormant branches
→ minimal viable commit

43. Safety / Governance Gate

第一版 Runtime 至少應具備:

external action allowlist
side-effect classification
budget ceiling
observation permission
COP hard minimum / maximum
manual pause
manual cancel
manual commit override

44. Model Independence

核心 Kernel 不應要求某家模型 SDK。

Agent Adapter 使用:

provider-neutral interface

45. Deterministic Core

能 deterministic 的部分應 deterministic:

  • schema validation;
  • budget accounting;
  • state versioning;
  • rule checking;
  • receipt generation;
  • replay manifest;
  • scheduler tie-breaking with seed。

46. Stochastic Boundary

模型 proposal 可不確定,但必須被包進:

ProposalRecord

並保存:

model_id
model_version
seed if available
request hash
response hash

47. Metrics

Runtime 本身輸出:

Qdecision,Q_{\mathrm{decision}}, Ccompute,C_{\mathrm{compute}}, Twall,T_{\mathrm{wall}}, Cobs,C_{\mathrm{obs}}, Recallcritical,Recall_{\mathrm{critical}}, Ralloc,R_{\mathrm{alloc}}, Robs,R_{\mathrm{obs}}, Rcommit.R_{\mathrm{commit}}.

48. Telemetry

本地 telemetry 至少包含:

active nodes
dormant nodes
pruned nodes
current horizon
current COP
remaining budget
current operation
decision stability
expected further gain

49. Dashboard

第一版 Dashboard 不需華麗。

頁面:

Overview
State
Possibility Graph
Operators
Observations
Budget
COP
Receipts
Replay

50. Possibility Graph View

至少顯示:

  • active;
  • dormant;
  • pruned;
  • merged;
  • near / mid / far;
  • risk;
  • reachability;
  • decision value。

51. COP Control View

UI 支援:

preset selector
advanced sliders
adaptive toggle
adaptive min / max
profile shift history

52. Minimal Neural Integration

MVP 可以只接一個模型。

要求:

JSON structured proposal
bounded branch count
bounded observation candidate count
bounded abstraction candidates

53. Branch Cap

第一版必須硬限制:

NactiveNmax.N_{\mathrm{active}} \leq N_{\max}.

避免模型一次生成無限候選。


54. Horizon Cap

hhmax.h \leq h_{\max}.

但可以用 abstract node 表示更遠期。


55. COE Cap

QtQmax|\mathcal Q_t| \leq Q_{\max}

以及:

O(q)Omax.|\mathcal O(q)| \leq O_{\max}.

56. Verification Cap

每輪:

NverifyVmax.N_{verify} \leq V_{\max}.

57. Runtime Acceptance Gates

Gate A:Schema Integrity

所有 canonical object 必須通過 schema。

Gate B:State Versioning

任何 mutation 都必須有 parent version。

Gate C:No Direct Neural Mutation

模型 proposal 不可直接改 canonical state。

Gate D:Budget Monotonicity

已消耗資源不可無故增加回來。

Gate E:Constraint Enforcement

Hard Rule fail 不可進 active space。

Gate F:COE Boundedness

觀察候選與 outcome 必須有上限。

Gate G:Preemption Safety

搶佔後可恢復或明確標記不可恢復。

Gate H:COP Boundary

Adaptive shift 不可離開授權區域。

Gate I:Receipt Completeness

重要決策必須有 receipt。

Gate J:Replay Integrity

同一 deterministic core 輸入可重播。

Gate K:Commit Gate

未通過 commit policy 不得自動 final commit。


58. MVP v0.1 的完成定義

MVP v0.1 不要求完整 AGI。

只需證明:

  1. 可以建立 canonical state;
  2. 可以生成有限 possibility graph;
  3. 可以執行 constraint projection;
  4. 可以在至少 5 種 cognition operator 間選擇;
  5. 可以計算 budget;
  6. 可以執行 COE 的簡化版本;
  7. 可以切換 COP;
  8. 可以主動 stop / commit;
  9. 可以輸出 receipts;
  10. 可以完整 replay。

59. MVP Operator Scope

v0.1 建議只做:

expand
observe
verify
prune
abstract
commit

第二輪再加入:

counterfactual_intervene
merge
refine
backtrack
reframe
profile_shift

60. MVP COE Scope

第一版只支援:

discrete observation candidates
bounded outcomes
heuristic DROV
simple cost

不要求完整 causal Bayesian model。


61. MVP Controller

第一版可以使用:

Score(u)=wqΔQ^(u)λC(u).Score(u) = w_q \widehat{\Delta Q}(u) - \lambda^\top C(u).

先用 heuristic estimator。

之後再學習:

ΔQ^.\widehat{\Delta Q}.

62. MVP COP

先提供 5 個 preset:

balanced
single_turn_accuracy
exploration
low_latency
high_risk_assurance

外加 advanced JSON。


63. MVP Demo Task

建議第一個 demo 不直接做複雜遊戲。

先做:

Hidden-State Route Planning

世界:

graph map
hidden hazard
limited observation
limited verify
multiple routes
deadline

AI 必須決定:

  • 先走;
  • 先觀察;
  • 先驗證;
  • 剪掉哪些路;
  • 何時 commit。

64. Demo 可測理論

同一任務可以比較:

Fixed Depth
Fixed Search
APSC without COE
APSC with COE
APSC with COP

65. Phase Plan

Phase 0 — Schema Freeze

完成:

canonical objects
JSON schemas
ID namespace
event types
receipt types

Phase 1 — State / Budget Kernel

完成:

StateStore
BudgetManager
RunManifest
versioning
receipts

Phase 2 — Possibility Graph

完成:

nodes
edges
constraint engine
reachability interface
prune
abstract

Phase 3 — Operator Runtime

完成:

registry
invocation
queue
status
checkpoint

Phase 4 — COE

完成:

observation candidates
outcome expansion
value estimate
observe / update

Phase 5 — Adaptive Controller

完成:

MVC
scheduler
shadow price
stop
commit
preemption

Phase 6 — COP Manager

完成:

profile
presets
adaptive region
shift
receipt

Phase 7 — Replay / Dashboard

完成:

events
replay
graph view
budget view
COP view

Phase 8 — Benchmark Harness

連接下一份白皮書:

AI Tension Arena Protocol & Benchmark Specification


66. Repository Layout

建議:

apsc-runtime/
├── README.md
├── pyproject.toml
├── schemas/
├── src/apsc/
│   ├── state/
│   ├── possibility/
│   ├── constraints/
│   ├── operators/
│   ├── coe/
│   ├── controller/
│   ├── cop/
│   ├── budget/
│   ├── receipts/
│   ├── replay/
│   ├── adapters/
│   └── api/
├── tests/
├── examples/
├── benchmarks/
└── docs/

67. 測試策略

至少包含:

unit tests
schema tests
property tests
golden replay tests
budget monotonicity tests
profile boundary tests
preemption tests
failure injection

68. Property Tests

重要 property:

Budgett+1BudgettBudget_{t+1} \preceq Budget_t

除非有明確 external top-up。

以及:

HardFail(ω)ωActive.HardFail(\omega) \Rightarrow \omega\notin Active.

以及:

AdaptiveCOPA.AdaptiveCOP \in \mathcal A.

69. Golden Replay

固定:

seed
world
observations
operator outcomes

應得到相同:

state versions
budget trajectory
receipts
final commit

70. 性能目標

v0.1 不追求最大吞吐。

優先順序:

correctness
replayability
boundedness
auditability
then performance

71. 非目標

v0.1 不做:

  • 通用世界模型;
  • 通用因果發現;
  • 完整 POMDP solver;
  • 自主訓練大模型;
  • 私有 chain-of-thought 儲存;
  • 無限制長期自主;
  • 現實高風險 actuator。

72. 可否證工程標準

若 APSC Runtime:

  1. 無法比固定 depth baseline 降低資源浪費;
  2. COE 無法改善任何觀察選擇;
  3. COP 無法穩定改變可預測行為;
  4. Replay 無法重現 deterministic core;
  5. Budget accounting 不可靠;

則 v0.1 架構需要被修正,而不是只增加更多模組。


73. 第一版最小成功條件

可以用一個最小公式表示:

same model+same task+different COP / controllermeasurably different, replayable cognitive trajectory.\boxed{ \text{same model} + \text{same task} + \text{different COP / controller} \Rightarrow \text{measurably different, replayable cognitive trajectory}. }

並且:

adaptive controller 在至少一類任務中>fixed cognition baseline.\boxed{ \text{adaptive controller} \text{ 在至少一類任務中} > \text{fixed cognition baseline}. }

74. 與第二份白皮書的邊界

本文件只定義:

APSC Runtime.APSC\ Runtime.

下一份:

AI Tension Arena Protocol & Benchmark Specification

負責:

  • arena world contract;
  • multi-agent episode;
  • fairness;
  • scoring;
  • benchmark tasks;
  • replay comparison;
  • leaderboard;
  • experimental matrix。

75. 最終架構總結

APSC Runtime 的工程母式:

OtStΩ~tΠKΩ^tUtCutSt+1\boxed{ O_t \rightarrow S_t \rightarrow \widetilde{\Omega}_t \rightarrow \Pi_{\mathcal K} \rightarrow \widehat{\Omega}_t \rightarrow \mathcal U_t \rightarrow \mathcal C \rightarrow u_t^\ast \rightarrow S_{t+1} }

其中:

C=f(Budget,COP,Risk,Deadline,ExpectedValue).\mathcal C = f( Budget, COP, Risk, Deadline, ExpectedValue ).

觀察候選則透過:

OCF\mathfrak O^{CF}

加入 cognition queue。


76. 結論

APSC Runtime 的目的不是重新建立一個神經模型,也不是建立一個與 LLM 平行的「邏輯 AI」。

它建立的是一個位於模型與世界之間的:

cognitive control runtime.\boxed{ \text{cognitive control runtime}. }

神經模型負責提出候選。

Runtime 負責:

哪些候選可以成立?
哪些值得展開?
哪些應觀察?
哪些應驗證?
哪些應剪枝?
哪些應抽象?
下一單位資源該花在哪裡?
Profile 是否需要改變?
現在是否已經值得停止?
最終決策是否可以提交?

因此:

Neural intelligence generates possibilities; APSC Runtime governs cognition over them.\boxed{ \text{Neural intelligence generates possibilities; APSC Runtime governs cognition over them.} }

這份白皮書固定了第一版工程骨架。

下一份技術白皮書將進入:

AI Tension Arena Protocol & Benchmark Specification v0.1

並把本 Runtime 放入真正會反作用的多智能體世界中進行可重播驗證。


版本記錄

v0.1 — 2026-08-24

本版首次固定:

  1. APSC Runtime 九層工程架構;
  2. Neural Proposal / Canonical Mutation 分離;
  3. Observation / Truth 分離;
  4. Candidate / Active Branch 分離;
  5. Decision / Commit 分離;
  6. Canonical Runtime Objects;
  7. State Kernel;
  8. Possibility Graph;
  9. Constraint Engine;
  10. Reachability Interface;
  11. Multi-Scale Horizon Manager;
  12. Cognitive Operator Registry;
  13. COE Engine;
  14. Observation Value;
  15. Adaptive Computation Controller;
  16. MVC;
  17. Dynamic Shadow Price;
  18. Scheduler;
  19. Preemption;
  20. Checkpoint;
  21. Stop / Commit Policy;
  22. COP Manager;
  23. Adaptive Region;
  24. Profile Shift / Hysteresis;
  25. Decision / Profile / Observation Receipts;
  26. Replay Layer;
  27. Private Reasoning Boundary;
  28. Runtime State Machine;
  29. API Surface;
  30. Budget Contract;
  31. Failure Model;
  32. Graceful Degradation;
  33. Governance Gate;
  34. Deterministic Core;
  35. Telemetry / Dashboard;
  36. Branch / Horizon / COE / Verification Caps;
  37. Acceptance Gate A–K;
  38. MVP Completion Definition;
  39. MVP Operator Scope;
  40. MVP COE Scope;
  41. MVP Controller;
  42. MVP COP;
  43. Hidden-State Route Planning Demo;
  44. Phase 0–8 Implementation Plan;
  45. Repository Layout;
  46. Test Strategy;
  47. Property Tests;
  48. Golden Replay;
  49. Non-goals;
  50. Falsification Criteria。