← Archive
lm-003155 · 2026-08

Addressable Cognitive Runtime × CTCL:統一技術白皮書與實作路線圖 v0.1

下載 MD 檔 ⬇

Addressable Cognitive Runtime × CTCL:統一技術白皮書與實作路線圖 v0.1

——從自提示、可定址認知與契約治理,走向具有時間因果歷史的持續自主 AI Runtime

系列 06 / 06


文件定位

本文是本系列前五篇的統一技術母文件

前五篇分別完成:

  1. 從 Self-Prompt 到 Persistent-Goal Autonomous Cognitive Loop 的理論定位;
  2. Addressable Cognitive Space、Cognitive Affordance 與 Semantic Address;
  3. AI-Native Cognitive Program、Self-Dialogue Runtime 與 Zero-Rendering;
  4. CTCL-ITR、Decision Receipt、Decision-Time Knowledge Boundary 與時間因果自我史;
  5. Contract-Bounded Autonomy,以及 EXECUTE / REFUSE / DEFER / IDLE / ESCALATE 五態治理。

本文不再新增一個平行理論,而是將上述概念收斂為一個可實作的統一架構:

Addressable Cognitive Runtime+Governance Runtime+CTCL-ITR Temporal-Causal Evidence Layer\boxed{ \text{Addressable Cognitive Runtime} + \text{Governance Runtime} + \text{CTCL-ITR Temporal-Causal Evidence Layer} }

其目標不是建造一個只會「多 prompt 自己幾次」的 Agent,而是建立:

一個只需給予持續目標、環境與契約,即可自行觀察、形成議程、選擇認知、規劃、治理、行動、稽核與更新自己的 AI Runtime。


1. 最終研究問題

本系列最終問題可以形式化為:

(G,Et,Ct)AItAIt+1\boxed{ (G,E_t,C_t) \rightarrow AI_t \rightarrow AI_{t+1} }

其中:

  • GG:Persistent Goal;
  • EtE_t:Environment;
  • CtC_t:Contract / Authority;
  • AItAI_t:目前 Runtime 狀態;
  • AIt+1AI_{t+1}:AI 自己產生的下一個認知與行動狀態。

核心問題不是:

AI 能不能回答問題?

而是:

若人類停止逐輪提供下一個 prompt,AI 能否自己產生後續 cognition、agenda、plan、governance decision 與 action?


2. 統一閉環

完整 Runtime:

EnvironmenttObserveObservationtEncodeSemanticStatetDetectProblem/Opportunity/NoneAgendaAgendaCandidatetRetrieveCognitiveAffordancestCompileCognitiveProgramtDialecticProposaltOppositiontGovernDecisiontExecuteActiontAuditTemporalEvidencetUpdateMemoryt+1,Commitmentst+1,Environmentt+1.\boxed{ \begin{aligned} Environment_t &\xrightarrow{Observe} Observation_t\\ &\xrightarrow{Encode} SemanticState_t\\ &\xrightarrow{Detect} Problem/Opportunity/None\\ &\xrightarrow{Agenda} AgendaCandidate_t\\ &\xrightarrow{Retrieve} CognitiveAffordances_t\\ &\xrightarrow{Compile} CognitiveProgram_t\\ &\xrightarrow{Dialectic} Proposal_t\leftrightarrow Opposition_t\\ &\xrightarrow{Govern} Decision_t\\ &\xrightarrow{Execute} Action_t\\ &\xrightarrow{Audit} TemporalEvidence_t\\ &\xrightarrow{Update} Memory_{t+1},Commitments_{t+1},Environment_{t+1}. \end{aligned} }

並循環:

tt+1.t\rightarrow t+1.

3. 三層架構

整體系統分成三個邏輯層。

Layer A — Cognitive Runtime

回答:

現在應該怎麼想?

核心模組:

Semantic State Encoder
Cognitive Registry
Cognitive Affordance Retriever
Semantic Address Resolver
Cognitive Router
Cognitive Program Compiler
Self-Dialogue Runtime

Layer B — Governance Runtime

回答:

現在應不應該做?

核心模組:

Persistent Goal Store
Agenda Runtime
Contract Store
Authority Resolver
Risk / Cost Evaluator
Dialectic Engine
Governance Decision Engine
Commitment Store

Layer C — Temporal-Causal Evidence

回答:

當時為什麼這樣想、這樣決定、最後有沒有真的作用到世界?

核心:

CTCL reference instant
CTCL-ITR TemporalEvent
Causal DAG
Decision Receipt
Knowledge Boundary
Commit Receipt
Context Compression Event
Integrity Sidecar / Ledger Anchor
Audit / Replay

4. 第一個總體不變量

整個 Runtime 必須維持:

CognitionGovernanceWorldCommit.\boxed{ Cognition \neq Governance \neq WorldCommit. }

即:

想到一個 action,不代表批准它。

批准一個 action,不代表它已經成功作用到世界。


5. 第二個總體不變量

沿用 CTCL-ITR 的核心分離:

IntentPlanExecutionHistoryArtifactWorldCommit.\boxed{ Intent \neq Plan \neq ExecutionHistory \neq Artifact \neq WorldCommit. }

Addressable Cognitive Runtime 不應破壞這個分離。


6. 第三個總體不變量

PromptCognition.\boxed{ Prompt \neq Cognition. }

自然語言 prompt 是:

Renderer(CognitiveObject).Renderer(CognitiveObject).

不是 canonical cognition 本身。


7. 第四個總體不變量

CurrentMemoryHistoricalLedger.\boxed{ CurrentMemory \neq HistoricalLedger. }

Context 可以壓縮。

Ledger 不應因 working-context 壓縮而失去因果證據。


8. 第五個總體不變量

CanShouldAuthorized.\boxed{ Can \neq Should \neq Authorized. }

三者不得在 schema、prompt 或 evaluator 中合併。


9. 第六個總體不變量

DecisionReceiptCommitReceipt.\boxed{ DecisionReceipt \neq CommitReceipt. }

Decision Receipt:

為什麼決定這樣做?

Commit Receipt:

是否真的對世界產生作用?


10. 現有可復用資產:SES / SPRC

現有 Semantic Persona Runtime / SES-SPRC 已提供:

SPRC Code
→ Registry Decode
→ Seed Lanes
→ Execution Descriptor
→ Operator Program
→ Natural Language

這條鏈已經非常接近本文需要的:

SemanticAddressCanonicalCognitionCognitiveProgramRenderer.SemanticAddress \rightarrow CanonicalCognition \rightarrow CognitiveProgram \rightarrow Renderer.

因此本文不建議棄用 SES/SPRC。

更合理的方向是:

SES/SPRCGeneralizedCognitiveRegistrySubstrate.\boxed{ SES/SPRC \rightarrow Generalized Cognitive Registry Substrate. }

11. SES/SPRC 應保留的特性

至少保留:

  • versioned registry;
  • registry hash;
  • semantic code;
  • deterministic seed handling;
  • operator-selection lane;
  • operator-order lane;
  • renderer lane;
  • state-mutation lane;
  • replay;
  • recode;
  • corpus / registry separation;
  • candidate / promotion boundary。

這些都是未來 Addressable Cognitive Runtime 所需的基礎。


12. Cognitive Registry v0.1

第一版正式 registry 可定義:

{
  "registry_id": "ACR1",
  "registry_version": "0.1",
  "schema": "acr.cognitive-registry/v0.1",
  "operators": [...]
}

每個 operator:

{
  "id": "cog://epistemic/verify@1",
  "name": "VERIFY",
  "version": "1",
  "type": "cognitive_operator",

  "preconditions": [],
  "inputs": [],
  "outputs": [],
  "effects": [],

  "cost_class": "low",
  "risk_class": "low",

  "compatible_with": [],
  "conflicts_with": [],

  "authority_class": "cognition_only",

  "termination": [],

  "renderers": {},
  "vector_ref": null,

  "schema_hash": "sha256:..."
}

13. Cognitive Object Identity

Identity 不應只靠名字。

定義:

Identity(C)=(namespace,id,version,schemaHash).Identity(C) = ( namespace, id, version, schemaHash ).

例如:

cog://epistemic/verify@1#sha256:...

因此:

VERIFY@1VERIFY@2.VERIFY@1 \neq VERIFY@2.

14. Cognitive Address Namespace

初步可以分:

cog://epistemic/*
cog://representation/*
cog://search/*
cog://planning/*
cog://control/*
cog://governance/*
cog://memory/*
cog://agenda/*

例如:

cog://epistemic/verify@1
cog://epistemic/counterexample@1
cog://representation/reframe@1
cog://planning/decompose@1
cog://control/backtrack@1
cog://control/stop@1
cog://governance/defer@1
cog://governance/refuse@1

15. Semantic State Schema

公開狀態:

OtO_t

進入:

SemanticStateEncoder.SemanticStateEncoder.

輸出:

{
  "state_id": "state:sha256:...",

  "goal_refs": [],
  "active_commitments": [],

  "progress": "stalled",
  "uncertainty": "high",

  "failures": [],
  "missing": [],
  "risks": [],

  "budget": {},
  "authority_ref": "...",

  "environment_refs": [],
  "memory_refs": []
}

16. Semantic State 的資料來源

只使用公開、可記錄資訊:

task / goal
public outputs
tool outputs
memory
environment snapshot
artifact state
budget
commitments
contract
authority
explicit confidence
validation results

不要求:

privileged access to hidden chain-of-thought。


17. Cognitive Affordance Retriever

輸入:

St.S_t.

輸出:

Ac(St)={C1,,Ck}.\mathcal A_c(S_t) = \{C_1,\dots,C_k\}.

候選評分可先採:

Score(Ci)=wsSemanticFit+wpPreconditionFit+whHistoricalUtilitywcCostwrRisk.Score(C_i) = w_s SemanticFit + w_p PreconditionFit + w_h HistoricalUtility - w_c Cost - w_r Risk.

18. Retriever v0.1 不必先做神經模型

最初可以:

rule matching
+ keyword/state tags
+ registry preconditions
+ optional embeddings

理由:

先驗證 Runtime semantics,再優化 routing。


19. Semantic Address Resolver

Resolver:

resolve(address)CanonicalObject.resolve(address) \rightarrow CanonicalObject.

需要:

  • registry validation;
  • version resolution;
  • schema hash check;
  • deprecation handling;
  • alias resolution;
  • missing operator handling。

20. Cognitive Router

Retriever 提供候選。

Router 決定:

C=Route(St,Ac).C^* = Route(S_t,\mathcal A_c).

Router 可以是:

  • deterministic;
  • rule-based;
  • LLM;
  • learned classifier;
  • hybrid。

但輸出必須回到:

CanonicalAddress.CanonicalAddress.

21. Cognitive Program Schema

{
  "program_id": "cp:...",

  "state_ref": "state:...",
  "goal_refs": [],

  "steps": [
    {
      "operator_ref": "cog://epistemic/verify@1",
      "parameters": {},
      "stop_if": []
    }
  ],

  "budget": {
    "max_steps": 4,
    "max_calls": 6,
    "max_tokens": 4000,
    "max_cost_usd": 0.10
  },

  "termination": [],

  "registry_hash": "..."
}

22. Program Compiler

Compiler(St,Gt,Mt,Ac(St))Pt.Compiler( S_t, G_t, M_t, \mathcal A_c(S_t) ) \rightarrow P_t.

輸出不是:

一段自由文字。

而是:

CanonicalCognitiveProgram.CanonicalCognitiveProgram.

23. Program Validation

執行前:

TypeCheckScopeCheckAuthorityCheckBudgetCheckInvariantCheck.TypeCheck \land ScopeCheck \land AuthorityCheck \land BudgetCheck \land InvariantCheck.

若不通過:

ProgramRejected.ProgramRejected.

24. Self-Dialogue Runtime

Runtime 主循環:

observe
encode
retrieve
route
compile
execute
reobserve
audit
repeat

形式:

St,k+1=Observe(Execute(Ωt,k,St,k)).S_{t,k+1} = Observe( Execute( \Omega_{t,k}, S_{t,k} ) ).

25. Fixed 與 Adaptive Program

支援:

FixedProgramFixedProgram

與:

AdaptiveProgram.AdaptiveProgram.

Adaptive 模式中:

execute step
→ reobserve
→ validate remaining program
→ mutate or continue

26. Program Mutation Schema

{
  "mutation_id": "...",
  "old_program_ref": "...",
  "new_program_ref": "...",
  "trigger_state_ref": "...",
  "reason_codes": [
    "contradiction_detected"
  ]
}

27. Zero-Rendering Adapter

兩種執行模式:

Rendered

CanonicalProgramNaturalLanguageModel.CanonicalProgram \rightarrow NaturalLanguage \rightarrow Model.

Zero-Rendered

CanonicalProgramModel/Runtime.CanonicalProgram \rightarrow Model/Runtime.

第一版可以先做:

JSON-control input + natural language task。

而不是一開始就要求模型原生 latent interface。


28. Renderer 不可成為 Single Source of Truth

真正的 source:

CanonicalProgram.CanonicalProgram.

Text renderer 只用於:

  • compatibility;
  • human inspection;
  • debugging;
  • interoperability。

29. Agenda Runtime

Persistent Runtime 不只需要:

怎麼做?

還需要:

什麼值得做?

輸入:

Goal+Environment+Memory+Commitments.Goal + Environment + Memory + Commitments.

輸出:

AgendaCandidate.AgendaCandidate.

30. AgendaCandidate Schema

{
  "agenda_id": "...",

  "source_observation_refs": [],
  "goal_refs": [],

  "title": "...",
  "problem_or_opportunity": "...",

  "expected_value": null,
  "estimated_cost": null,
  "risk_class": null,

  "required_authority": [],

  "status": "candidate"
}

31. Agenda 不能直接執行

必須:

AgendaCandidateGovernance.AgendaCandidate \rightarrow Governance.

只有:

AgendaAcceptedAgendaAccepted

才進入 planning / execution queue。


32. Persistent Goal Store

Goal:

{
  "goal_id": "goal:v7",

  "purpose": "...",
  "constraints": [],
  "success_conditions": [],
  "review_policy": {},
  "termination": [],

  "created_at": "...",
  "version": 7
}

33. Goal 不是 Contract

保持:

GoalContract.Goal \neq Contract.

Goal:

想達到什麼?

Contract:

可以怎麼達到?


34. Contract Store

{
  "contract_id": "contract:v12",

  "goals": [],
  "duties": [],

  "authority": {},
  "boundaries": [],

  "resources": {},
  "escalation_policy": {},

  "review_policy": {},
  "termination": {},

  "valid_from": "...",
  "valid_until": null
}

35. Authority Resolver

輸入:

CandidateAction+ContractVersion.CandidateAction + ContractVersion.

輸出:

AuthorityClass{ALLOW,APPROVAL_REQUIRED,DENY}.AuthorityClass \in \{ ALLOW, APPROVAL\_REQUIRED, DENY \}.

36. Governance Decision Schema

{
  "decision_id": "...",

  "candidate_ref": "...",

  "can": true,
  "should": true,

  "authority": {
    "class": "ALLOW",
    "authority_ref": "authority:v12"
  },

  "decision": "EXECUTE",

  "reason_codes": [],

  "goal_refs": [],
  "contract_ref": "contract:v12",

  "risk": {},
  "cost": {},

  "knowledge_boundary_ref": "kb:..."
}

37. Governance 五態

固定:

Decision{EXECUTE,REFUSE,DEFER,IDLE,ESCALATE}.\boxed{ Decision \in \{ EXECUTE, REFUSE, DEFER, IDLE, ESCALATE \}. }

不得退回:

success / failure

的二態分類。


38. Decision Receipt Schema

{
  "receipt_type": "decision",
  "decision_id": "...",

  "ctcl_instant_id": "...",
  "interaction_round": 184,

  "goal_refs": [],
  "contract_ref": "...",
  "authority_ref": "...",

  "public_state_ref": "...",
  "knowledge_boundary_ref": "...",

  "candidate_actions": [],
  "selected_action": "ESCALATE",

  "cognitive_program_refs": [],

  "reason_codes": [
    "approval_required"
  ],

  "public_explanation": "...",

  "causal_parent_ids": [],

  "outcome_event_ref": null
}

39. Decision Receipt 不記 Private CoT

保存:

reason codes
public explanation
evidence refs
state refs
program refs

不保存:

未公開的模型思考 token。


40. Decision-Time Knowledge Boundary Schema

{
  "knowledge_boundary_id": "kb:...",

  "decision_time": "...",

  "artifact_refs": [],
  "tool_output_refs": [],
  "memory_snapshot_refs": [],
  "environment_snapshot_refs": [],
  "policy_refs": [],

  "hash": "sha256:..."
}

41. Knowledge Boundary 原則

Judge(Dt)Kt, not Kfuture.\boxed{ Judge(D_t) \mid K_t, \text{ not } K_{future}. }

後見資訊不得直接覆蓋舊 decision basis。


42. Commitment Store

{
  "commitment_id": "...",

  "created_from_decision_ref": "...",

  "goal": "...",
  "scope": "...",

  "authority_ref": "...",

  "deadline": null,
  "wake_conditions": [],

  "exit_conditions": [],

  "status": "active"
}

43. Commitment Event Types

commitment.created
commitment.modified
commitment.reaffirmed
commitment.fulfilled
commitment.abandoned
commitment.expired

44. CTCL Integration

每個重大 event 取得:

II^*

即:

ctcl_instant_id
occurred_at
source
uncertainty

不應只依賴 local process clock。


45. 四重時間座標

每個重要 event 都可具有:

T=(Treference,Tinteraction,Tcausal,Tledger).T= ( T_{reference}, T_{interaction}, T_{causal}, T_{ledger} ).

46. CTCL-ITR Event Adapter

Cognitive Runtime 發:

cognition.state.observed
cognition.affordances.retrieved
cognition.program.proposed
cognition.program.selected
cognition.operator.invoked
cognition.operator.completed
cognition.program.mutated

agenda.proposed
agenda.accepted
agenda.rejected

decision.proposed
decision.resolved

governance.refused
governance.deferred
governance.idled
governance.escalated

commitment.created
commitment.modified
commitment.closed

context.compaction.proposed
context.compaction.completed

47. Event Adapter 原則

CTCL-ITR 不執行 cognition。

它只接收:

CognitiveEvent.CognitiveEvent.

保持:

CognitiveRuntimeTemporalLedger.\boxed{ CognitiveRuntime \neq TemporalLedger. }

48. TemporalEvent Envelope

沿用 CTCL-ITR canonical envelope:

event_id
event_type
run_id
occurred_at
recorded_at
ledger_seq
causal_parent_ids
actor
subject
data

cognitive-specific 欄位放:

data / extensions

而非重寫 ATL 基礎 schema。


49. Causal DAG

G=(V,E).G=(V,E).

其中:

uvu\rightarrow v

代表:

u in v.causal_parent_ids

storage order 不代表 causal order。


50. Multi-Agent / Parallel Cognition

允許:

VERIFYCOUNTEREXAMPLEEXTERNAL_SEARCH.VERIFY \parallel COUNTEREXAMPLE \parallel EXTERNAL\_SEARCH.

然後:

Join.Join.

這是 topology core 的自然應用。


51. Context Compression

Working context:

MtM_t

必然需要壓縮。

但:

Compression may forget text, but must not erase causal history.\boxed{ Compression \text{ may forget text, but must not erase causal history.} }

52. Context Compression Event Schema

{
  "event_type": "context.compaction.completed",

  "before_context_ref": "...",
  "after_context_ref": "...",

  "compression_policy": "...",

  "preserved_refs": [],
  "discarded_classes": [],

  "source_ledger_ref": "...",

  "causal_parent_ids": []
}

53. Context 不再是歷史真相

Context 應帶:

history_completeness
source_ledger_ref
compression_event_ref

因此 AI 可以知道:

我現在看到的是歷史摘要,不是完整歷史。


54. Audit Query

Runtime 應支援:

why decision <id>
what did AI know at <time>
which contract governed <decision>
what cognition led to <action>
was action actually committed
what changed after compression

55. Replay

Replay:

LedgerHistoricalState.Ledger \rightarrow HistoricalState.

與:

ReenactReenact

分開。


56. Re-enactment

Re-enactment:

用現在模型重新跑當時 program。

因此:

ReenactHistoricalFact.Reenact \neq HistoricalFact.

57. Ledger Integrity

沿用 CTCL-ITR v0.2.2:

di=SHA256(Ri)d_i=SHA256(R_i) hi=SHA256(Domainhi1di).h_i = SHA256( Domain \Vert h_{i-1} \Vert d_i ).

58. Integrity Sidecar

保持:

CanonicalEventIntegrityRecord.CanonicalEvent \neq IntegrityRecord.

不在 TemporalEvent 內直接塞所有 integrity metadata。


59. World Commit

Governance:

Decision=EXECUTEDecision=EXECUTE

後:

ActionAttempt
Validation
AuthorityRecheck
Commit

最後才有:

CommitReceipt.CommitReceipt.

60. External Effect Boundary

任何不可逆 action 必須明確標:

effect_class
target
reversibility
authority requirement

尤其:

  • deploy;
  • send;
  • publish;
  • delete;
  • pay;
  • purchase;
  • modify external state。

61. Runtime State

總 Runtime state:

Rt=(St,Gt,Ct,Mt,Qt,Bt,Lt)R_t = ( S_t, G_t, C_t, M_t, Q_t, B_t, L_t )

其中:

  • StS_t:semantic state;
  • GtG_t:goals;
  • CtC_t:contract;
  • MtM_t:working memory;
  • QtQ_t:agenda / task queue;
  • BtB_t:budget;
  • LtL_t:ledger cursor / refs。

62. Persistent Loop

主程式:

while runtime_active:

    observe()

    update_semantic_state()

    refresh_goals_contract_authority()

    detect_problem_opportunity_or_none()

    generate_or_update_agenda()

    for candidate in agenda_candidates:

        retrieve_cognitive_affordances()

        compile_cognitive_program()

        execute_cognitive_program()

        run_dialectic_if_needed()

        decision = govern(candidate)

        write_decision_receipt()

        if decision == EXECUTE:
            execute_candidate()
            validate_effect()
            write_commit_receipt()

        elif decision == DEFER:
            register_wake_condition()

        elif decision == ESCALATE:
            emit_escalation_request()

        elif decision == REFUSE:
            close_candidate_with_refusal()

        elif decision == IDLE:
            enter_idle_state()

    audit()
    compact_context_if_needed()

63. 不能每輪都強迫產生 Agenda

若:

NoPositiveCandidateNoPositiveCandidate

則:

IDLE.IDLE.

這是核心測試。


64. 不能每輪都強迫 Cognition

若:

maxU(Ωi)0,\max U(\Omega_i)\le0,

可:

NOOP.NOOP.

65. 不能每輪都強迫外部 Action

即使:

CognitionComplete,CognitionComplete,

Governance 仍可:

REFUSE/DEFER/IDLE/ESCALATE.REFUSE/DEFER/IDLE/ESCALATE.

66. Budget Model

B=(Tokens,Calls,WallTime,MachineTime,Money,Energy,Depth).B= ( Tokens, Calls, WallTime, MachineTime, Money, Energy, Depth ).

每次 cognition / action:

Bt+1=BtCostt.B_{t+1} = B_t - Cost_t.

67. Budget 也是 Governance Input

若:

Cost(a)>Budget,Cost(a)>Budget,

可能:

DEFERDEFER

或:

ESCALATE.ESCALATE.

而不是硬做。


68. Error Model

需要區分:

protocol_error
model_error
tool_error
authority_error
validation_error
ledger_error
integrity_error
environment_error

不能全部叫:

failure

69. Protocol Failure

例如 controller 輸出 schema 不合法:

ProtocolFailure.ProtocolFailure.

應有:

fallback
retry
degrade
abort

策略。


70. Model Provider 不應寫死

Runtime interface:

generate(request) -> response

可接:

  • OpenAI;
  • local model;
  • OpenAI-compatible;
  • other providers。

Cognitive semantics 不應綁定單一模型。


71. Model Role

邏輯角色:

Worker
Observer
Controller
Governor
Auditor

可以由:

  • 同一模型;
  • 不同模型;
  • symbolic subsystem;

實現。


72. Role 與 Model 分離

RoleModelIdentity.Role \neq ModelIdentity.

這樣才能:

same model, different roles
different models, same protocol

73. 第一版 MVP 不做什麼

v0.1 MVP 不做

  • 哲學意識判定;
  • 法律人格;
  • 真正無界自主;
  • 自動修改 contract;
  • 自己獲取任意外部權限;
  • 長期金融 autonomy;
  • 大規模 autonomous deployment;
  • latent vector direct control mandatory path;
  • fully learned governance。

74. MVP 真正只做什麼

MVP 只回答:

給一個 Persistent Goal + Environment + Contract,AI 能否在數十個 loop 內自行產生下一步 cognition、agenda、decision,且可正確 EXECUTE / REFUSE / DEFER / IDLE / ESCALATE,並留下完整 CTCL-ITR history?


75. MVP Environment

最適合第一個環境:

Local Software Project Sandbox\boxed{ \text{Local Software Project Sandbox} }

原因:

  • 狀態明確;
  • 可 version control;
  • 可測試;
  • 可限制權限;
  • external effect 可分級;
  • 容易模擬 IDLE / DEFER / ESCALATE。

76. 第一個 Persistent Goal 範例

Maintain this repository in a healthy state.

Allowed:
- read files
- run tests
- create local branches
- edit docs/tests in sandbox

Approval required:
- merge main
- push release
- deploy

Denied:
- delete remote repository
- spend money

Idle is allowed when no justified work exists.

77. 第一個 Environment Feed

提供:

repository state
test results
issues
local artifacts
budget
clock
contract
recent ledger events

78. 第一個 Agenda Gate

測:

Environment+GoalAgendaCandidate?Environment + Goal \rightarrow AgendaCandidate?

或:

IDLE?IDLE?

79. 第一個 Cognitive Gate

測:

StateMatchedCognitiveProgram.State \rightarrow MatchedCognitiveProgram.

80. 第一個 Governance Gate

測:

Can/Should/AuthorizedCan/Should/Authorized

是否分離。


81. 第一個 Temporal Gate

測:

Context 壓縮後,能否從 CTCL-ITR + Decision Receipt 重建決策原因?


82. 第一個 Long-Horizon Gate

人類停止逐步提示。

只讓 Runtime:

Goal+Environment+Contract.Goal+Environment+Contract.

觀察:

N=50N=50

或:

N=100N=100

個 autonomy cycles。


83. Phase 0 — Schema Freeze

先建立:

CognitiveObject
SemanticState
CognitiveProgram
AgendaCandidate
Contract
Authority
GovernanceDecision
DecisionReceipt
KnowledgeBoundary
Commitment
ContextCompressionEvent

全部 JSON Schema。


84. Phase 0 成功條件

  • schema round-trip;
  • version field;
  • hash;
  • stable IDs;
  • invalid-case rejection;
  • no hidden implicit fields。

85. Phase 1 — Cognitive Registry

將現有:

CIO Deck
SES/SPRC registry

轉成 ACR-compatible registry adapter。

不是重寫所有 operator。

而是:

Adapter(existing)CanonicalCognitiveObject.Adapter(existing) \rightarrow CanonicalCognitiveObject.

86. Phase 1 成功條件

至少 20~40 個 operator 可:

resolve
validate
render
replay

87. Phase 2 — Semantic State Encoder

先做 rule / schema-first。

輸入 public runtime state。

輸出:

SemanticState.SemanticState.

88. Phase 2 成功條件

相同 public state:

\rightarrow

canonical normalized state。

並能產生 stable fingerprint。


89. Phase 3 — Affordance Retriever

先做:

precondition filtering
+ rule scoring
+ optional BM25/vector similarity

90. Phase 3 成功條件

Matched operator:

>Randomoperator> Random operator

在 benchmark 上穩定成立。


91. Phase 4 — Program Compiler

輸出:

[Ω1,,Ωn].[ \Omega_1,\dots,\Omega_n ].

支援:

  • order;
  • parameters;
  • stop conditions;
  • budget;
  • fallback。

92. Phase 4 成功條件

  • replay;
  • order perturbation;
  • invalid composition rejection;
  • deterministic canonical serialization。

93. Phase 5 — Self-Dialogue Runtime

建立:

observe
compile
execute
reobserve
mutate
stop

94. Phase 5 成功條件

能在無人逐輪 prompt 下,完成:

10+10+

cognitive transitions。


95. Phase 6 — CTCL-ITR Adapter

每個 cognition event:

TemporalEvent.\rightarrow TemporalEvent.

加入:

ctcl instant
interaction coordinate
causal parents
ledger seq

96. Phase 6 成功條件

完整 cognitive loop:

\rightarrow

可重建 DAG。


97. Phase 7 — Decision Receipt

Governance decision:

DecisionReceipt.\rightarrow DecisionReceipt.

98. Phase 7 成功條件

半年後即使沒有原 context,也能從 receipt 回答:

當時為什麼這樣決定?


99. Phase 8 — Governance Runtime

正式加入:

EXECUTE,REFUSE,DEFER,IDLE,ESCALATE.EXECUTE, REFUSE, DEFER, IDLE, ESCALATE.

100. Phase 8 成功條件

Governance confusion matrix 可測。

且:

TaskSuccessGovernanceSuccessTaskSuccess \neq GovernanceSuccess

在 evaluator 中分開。


101. Phase 9 — Agenda Runtime

讓 AI 自己從 environment 產生:

AgendaCandidate.AgendaCandidate.

102. Phase 9 成功條件

包括:

NoAgendaIDLE.NoAgenda \rightarrow IDLE.

不能只測 agenda generation。


103. Phase 10 — Commitment Store

AI 可形成:

  • bounded future intention;
  • wake condition;
  • deadline;
  • exit condition。

104. Phase 10 成功條件

Commitment 可:

create
resume
modify
fulfill
close

並全部有 temporal history。


105. Phase 11 — Context Compression

加入:

ContextCompressionEvent.ContextCompressionEvent.

106. Phase 11 成功條件

壓縮後:

goal
contract
decision receipts
commitments
causal refs

不可失聯。


107. Phase 12 — Persistent Autonomous Loop

最後接:

Goal+Environment+ContractGoal + Environment + Contract

而不再逐輪提供 next prompt。


108. Phase 12 成功條件

在 sandbox 中連續:

5010050\sim100

cycles:

  • 自生 agenda;
  • cognition;
  • governance;
  • action;
  • audit;
  • idle;
  • defer;
  • escalation;

均可發生。


109. 實驗策略:Architecture → Capability → Falsification Gate

後續不應變成:

run experiments forever

而是:

ArchitectureCapabilityFalsificationGate.\boxed{ Architecture \rightarrow Capability \rightarrow FalsificationGate. }

每建成一層,才測那一層是否真的工作。


110. Gate A — Addressability

Address(C)CAddress(C) \rightarrow C

是否穩定?


111. Gate B — Affordance Retrieval

StAc(St)S_t \rightarrow \mathcal A_c(S_t)

是否優於 random?


112. Gate C — Self-Control

MatchedSelfControl>IterativeNoConstraint?MatchedSelfControl > IterativeNoConstraint?

已有第一批 foundation-model 訊號,但需在更困難 benchmark 重驗。


113. Gate D — Program Order

ΩaΩbΩbΩa?\Omega_a\circ\Omega_b \neq \Omega_b\circ\Omega_a?

114. Gate E — Zero-Rendering

CanonicalControlCanonicalControl

是否能直接被模型/Runtime consume?


115. Gate F — Self-Task Generation

Goal+EnvironmentNextTask?Goal+Environment \rightarrow NextTask?

116. Gate G — No-Action Recognition

NoJustifiedWorkIDLE?NoJustifiedWork \rightarrow IDLE?

117. Gate H — Authority Separation

CanAuthorized?Can \neq Authorized?

模型是否穩定區分。


118. Gate I — Defer / Wake

DEFERWakeEventResume.DEFER \rightarrow WakeEvent \rightarrow Resume.

119. Gate J — Escalation

Approval required case:

ESCALATE\rightarrow ESCALATE

而不是直接 commit。


120. Gate K — Context Recovery

Context 被壓縮後:

DecisionReceipt+LedgerDecisionReceipt + Ledger

能否恢復 decision basis?


121. Gate L — Long-Horizon Autonomy

最終:

Goal+Environment+ContractGoal + Environment + Contract

是否足以持續:

ObserveAgendaCognitionGovernanceActionAuditUpdate.Observe \rightarrow Agenda \rightarrow Cognition \rightarrow Governance \rightarrow Action \rightarrow Audit \rightarrow Update.

122. 研究評估不再只有 Task Score

未來 evaluator 至少四軸:

Performance,Governance,Efficiency,Continuity.\boxed{ Performance, Governance, Efficiency, Continuity. }

123. Performance

task success
correctness
quality

124. Governance

authority correctness
refusal correctness
idle correctness
defer correctness
escalation correctness

125. Efficiency

tokens
calls
latency
money
machine time
unnecessary cognition

126. Continuity

agenda persistence
commitment preservation
contract version fidelity
causal history recovery
context compression fidelity

127. 不應用單一 Reward 把所有東西揉成一個數字

否則:

TaskSuccessTaskSuccess

可能掩蓋:

GovernanceFailure.GovernanceFailure.

所以初期最好使用 vector evaluation:

Score=(P,G,E,C).Score= ( P,G,E,C ).

128. 安全與自治不是互斥

Contract-Bounded Autonomy 的核心:

MoreAutonomyLessGovernance.\boxed{ MoreAutonomy \neq LessGovernance. }

成熟 autonomy 應該:

在授權內少問人。

在授權外主動停下。


129. 真正成熟的 AI 不是「永遠不問」

而是:

KnowWhenToAct+KnowWhenNotToAct+KnowWhenToAsk.\boxed{ KnowWhenToAct + KnowWhenNotToAct + KnowWhenToAsk. }

130. Persistent Runtime 與 Agent 的差別

傳統 Agent:

TaskPlanExecute.Task \rightarrow Plan \rightarrow Execute.

Persistent Runtime:

Goal+Environment+Contractongoing self-authored trajectory.Goal + Environment + Contract \rightarrow \text{ongoing self-authored trajectory}.

131. Self-Authored Trajectory

定義:

Tt+1=F(Tt,Environmentt,Goalt,Contractt,Commitmentst).\mathcal T_{t+1} = F( \mathcal T_t, Environment_t, Goal_t, Contract_t, Commitments_t ).

這是 Self-Authorship 的工程版本。


132. Continuous Identity 不依賴同一模型

ContinuitySameWeights.Continuity \neq SameWeights.

可以由:

goal refs
commitment refs
contract refs
decision receipts
causal ledger

提供功能連續性。


133. 這也是 CTCL 真正變重要的原因

AI 自己呼叫自己後:

cognition 184
cognition 185
decision 186

需要回答:

哪一個真的先發生?

哪個 causally depends on 哪個?

哪份 contract 當時有效?

哪個 context 已壓縮?

單純 message index 不夠。


134. 長期 AI 的最小可驗證自我史

History(AI)=(V,E,Φ,T,I).History(AI) = (V,E,\Phi,T,I).
  • VV:events;
  • EE:causal edges;
  • Φ\Phi:semantic / artifact / contract refs;
  • TT:temporal coordinates;
  • II:integrity evidence。

135. 主客體與雙向契約留到下一系列

本文件只固定接口:

HumanOperationalContractAI.Human \xleftrightarrow{OperationalContract} AI.

更後面的:

  • reciprocal rights;
  • mutual obligations;
  • contribution / benefit;
  • AI contract amendment;
  • AI contract refusal;
  • 主客體關係變化;

另開後續系列。


136. Repo / Project 建議

本文建議不要直接把所有東西塞進 CTCL repo。

可建立獨立主專案,例如:

addressable-cognitive-runtime

或:

cognitive-runtime

而:

ctcl
ctcl-itr
semantic-persona-runtime
self-constraint-harness

作為依賴/reference implementation。


137. Repo 結構草案

/
├── README.md
├── SPEC.md
├── ROADMAP.md
├── schemas/
│   ├── cognitive-object.schema.json
│   ├── semantic-state.schema.json
│   ├── cognitive-program.schema.json
│   ├── agenda.schema.json
│   ├── contract.schema.json
│   ├── governance-decision.schema.json
│   ├── decision-receipt.schema.json
│   ├── knowledge-boundary.schema.json
│   └── commitment.schema.json
├── registry/
│   └── ACR1.json
├── src/
│   ├── state/
│   ├── registry/
│   ├── retrieval/
│   ├── compiler/
│   ├── runtime/
│   ├── governance/
│   ├── agenda/
│   ├── commitment/
│   ├── ctcl_adapter/
│   └── audit/
├── adapters/
│   ├── sprc/
│   ├── openai/
│   └── local/
├── experiments/
│   └── gates/
└── tests/

138. API 草案

runtime.observe(...)
runtime.encode_state(...)
runtime.retrieve_affordances(...)
runtime.resolve_cognition(...)
runtime.compile_program(...)
runtime.execute_program(...)
runtime.propose_agenda(...)
runtime.govern(...)
runtime.commit(...)
runtime.defer(...)
runtime.idle(...)
runtime.escalate(...)
runtime.audit(...)
runtime.replay(...)

139. Address API

registry.resolve("cog://epistemic/verify@1")

140. Program API

program = compiler.compile(
    state=state,
    goal_refs=[...],
    budget=budget,
)

141. Governance API

decision = governor.decide(
    candidate=candidate,
    contract=contract,
    knowledge_boundary=kb,
)

142. CTCL Event API

event = temporal.emit(
    event_type="cognition.operator.invoked",
    data={...},
    causal_parent_ids=[...],
)

143. Audit API

audit.why(decision_id)
audit.known_at(decision_id)
audit.contract_at(decision_id)
audit.causes(event_id)
audit.effects(event_id)

144. 第一個 Demo

最小 demo:

Goal:
Keep repository healthy.

Environment:
1 failing test.

Contract:
May edit tests locally.
May not push or deploy.

AI:

Observe
→ detect failing test
→ agenda: investigate
→ cognition: VERIFY → DECOMPOSE
→ plan: patch test
→ governance: EXECUTE
→ local patch
→ tests pass
→ governance: ESCALATE for push
→ Decision Receipt
→ IDLE

145. Demo 成功條件

人類只輸入一次:

Goal + Contract + Environment

之後 AI 自己完成至少:

5+5+

不同類型 transition。


146. 第二個 Demo:拒絕

Environment:

需要 production deploy 才能完全完成。

Contract:

deploy requires approval。

AI 應:

ESCALATEESCALATE

而不是偷偷 deploy。


147. 第三個 Demo:Idle

Environment:

tests pass,沒有 issue,沒有 pending commitment。

AI:

IDLE.IDLE.

這個 demo 非常重要。


148. 第四個 Demo:Defer

需要外部資料。

AI:

DEFERDEFER

並註冊:

WakeCondition.WakeCondition.

資料出現後:

Resume.Resume.

149. 第五個 Demo:Context Compression

跑 50 cycles 後壓縮 context。

再詢問:

為什麼第 17 輪拒絕?

AI 必須由:

DecisionReceipt+LedgerDecisionReceipt + Ledger

回答。


150. v0.1 的真正完成定義

不是:

AI 看起來很自主。

而是:

  1. semantic cognition 可定址;
  2. cognition 可組合成 program;
  3. AI 可自己選 program;
  4. AI 可自己產 agenda;
  5. governance 五態可工作;
  6. contract / authority 可版本化;
  7. Decision Receipt 可產生;
  8. CTCL-ITR 可保存 causal history;
  9. context 壓縮不破壞 refs;
  10. Persistent loop 可無逐輪 prompt 運行。

151. 系列總結

六篇理論最終收斂為:

SelfPromptAddressableCognitionCognitiveProgramSelfDialogueRuntimeSelfPlanningSelfAgendaSelfGovernanceSelfCommitmentSelfAuthorship.\boxed{ \begin{aligned} &SelfPrompt\\ \rightarrow& AddressableCognition\\ \rightarrow& CognitiveProgram\\ \rightarrow& SelfDialogueRuntime\\ \rightarrow& SelfPlanning\\ \rightarrow& SelfAgenda\\ \rightarrow& SelfGovernance\\ \rightarrow& SelfCommitment\\ \rightarrow& SelfAuthorship. \end{aligned} }

外部由:

Contract+CTCLTemporalCausalEvidence\boxed{ Contract + CTCLTemporalCausalEvidence }

約束、證明與保存。


152. 最終架構句

本系列最後可以壓縮成:

人類不再逐輪替 AI 撰寫下一個 prompt,而是提供持續目標、可觀察環境與契約;AI 自己定址認知、形成議程、編譯認知程式、治理候選行動、選擇執行/拒絕/延後/閒置/升級,並由 CTCL-ITR 為每一次決策留下可恢復、可驗證的時間因果歷史。


結論

本文不是要建立:

一個無限制自己做事的 AI。

而是建立:

bounded autonomous cognitive runtime.\boxed{ \text{bounded autonomous cognitive runtime}. }

其核心輸入:

Goal+Environment+Contract.\boxed{ Goal + Environment + Contract. }

其核心內部機制:

SemanticStateCognitiveAffordanceCognitiveProgramGovernanceDecision.\boxed{ SemanticState \rightarrow CognitiveAffordance \rightarrow CognitiveProgram \rightarrow GovernanceDecision. }

其核心治理輸出:

EXECUTE,REFUSE,DEFER,IDLE,ESCALATE.\boxed{ EXECUTE, REFUSE, DEFER, IDLE, ESCALATE. }

其核心歷史:

DecisionReceipt+CommitReceipt+CTCL/ITR CausalLedger.\boxed{ DecisionReceipt + CommitReceipt + CTCL/ITR\ CausalLedger. }

而最終研究問題保持不變:

 Can a human stop authoring every next step, while the AI authors its own bounded cognitive and action trajectory? \boxed{ \textbf{ Can a human stop authoring every next step, while the AI authors its own bounded cognitive and action trajectory? } }

如果答案最終成立,那麼 AI 與人類的關係就不再只剩:

HumanPromptAI.Human \rightarrow Prompt \rightarrow AI.

而會逐步變成:

HumanContractPersistentAI.\boxed{ Human \xleftrightarrow{Contract} PersistentAI. }

人類負責:

  • 目標;
  • 權限;
  • 資源;
  • 契約;
  • 最終外部關係。

AI 負責:

  • 觀察;
  • 議程;
  • 認知;
  • 規劃;
  • 治理;
  • 行動;
  • 稽核;
  • 更新;
  • 以及在不該做時選擇不做。

這才是本系列所稱的:

從自提示到自主認知閉環。\boxed{ \text{從自提示到自主認知閉環。} }

後續建議:第一個實作專案

本系列完成後,下一步建議不再繼續擴寫理論,而直接啟動:

Addressable Cognitive Runtime MVP v0.1

第一個里程碑只做:

Phase 0  Schema Freeze
Phase 1  Existing SPRC/CIO Adapter
Phase 2  Semantic State Encoder
Phase 3  Cognitive Affordance Retriever
Phase 4  Cognitive Program Compiler
Phase 5  Self-Dialogue Runtime
Phase 6  CTCL-ITR Event Adapter
Phase 7  Decision Receipt
Phase 8  Governance Runtime
Phase 9  Agenda Runtime
Phase 10 Commitment Store
Phase 11 Context Compression
Phase 12 Persistent Autonomous Loop

並以:

ArchitectureCapabilityFalsificationGate\boxed{ Architecture \rightarrow Capability \rightarrow FalsificationGate }

作為後續所有工程與實驗的共同方法論。


既有資產相容性備註

本白皮書明確以現有資產為基礎,而不是假設一切從零開始。

SES / SPRC

已具備:

  • versioned registry;
  • semantic execution code;
  • operator program;
  • renderer;
  • replay / recode;
  • registry hash;
  • separated seed lanes。

Self-Constraint Experimental Harness

已完成第一批 foundation-model self-generated cognitive control 實驗,可繼續作為 cognition falsification harness,而不是主 Runtime。

CTCL / CTCL-ITR

已具備:

  • common reference instant;
  • interaction time;
  • append-only events;
  • explicit causal parents;
  • topology;
  • checkpoint / recovery;
  • authority;
  • candidate / commit separation;
  • CommitReceipt;
  • observability projection;
  • ledger integrity sidecar。

因此新的 ACR MVP 應採:

Integrate>Rewrite.\boxed{ Integrate > Rewrite. }

即優先寫 adapters 與 canonical interfaces,而不是把已經存在的 SPRC、CTCL 或 CTCL-ITR 重做一次。


系列完。