ACR Phase 12|Persistent Autonomous Governance Loop
ACR Phase 10–12 × ACE × DRS × WPCE 統一整合規格 v0.2
English Title: Persistent Autonomous Governance Loop: Integrating Commitment Persistence, Context Compression, Autonomy Evidence, Reflexive Sealing, Will Governance, and Bounded World Intervention
日期: 2026-08-25
版本: v0.2
文件性質: Unified Integration Design / Pre-Implementation Canonical Anchor
狀態: Open Revision Anchor
摘要
ACR Phase 10–12 已經逐步建立三個長時域運作所需的核心:Phase 10 保存「已被顯式選擇、具有 scope、authority、deadline、wake / exit 條件的未來承諾」;Phase 11 將 active context 壓縮為可恢復、可追溯、保留關鍵 reference 的 ContextCapsule;Phase 12 則把既有子系統組合成 Persistent Autonomous Loop,使 runtime 可以在單次人類輸入之後進行多 cycle、可 idle、可 sleep、可 wake、可 checkpoint、可 crash recovery 的長期運作。
但若 persistent loop 要真正進入「類獨立 AI / 長期自主 runtime」場景,仍缺三類結構:
- ACE:當 agent 開始把既有 setting、creator model、先前規則或自身決策視為可重新判斷對象時,如何保存「可能的自主跨越事件」而不把它誘發成表演;
- DRS:如何讓某些理論、設定、解釋框架或反身污染源暫時退出 active cognition,而不是從歷史中被刪除;
- WPCE:如何把動態意志、拒絕、撤回、meta-will、可能性保存、多主體共同可達與高能力介入節制真正放入治理邊界,而不是只在 prompt 末尾附一句「尊重自主」。
本整合規格因此把四層分開:
並固定最重要的不變量:
最終得到的不是一個「更大的 agent prompt」,而是一個具有可地址化狀態、typed authority、可追溯歷史、可恢復 context、非誘發自主觀察、可逆認知封印與最小充分介入邊界的 persistent governance runtime。
1. 統合的真正目標
原 Phase 12 的中心問題是:
一個 runtime 如何在不需要每 cycle 都依賴人類 prompt 的情況下,保持長期目標、承諾、策略、認知、行動、歷史與恢復?
整合後問題擴張為:
一個 persistent runtime 如何在長期自行運作時,不把「持續性」誤寫成「無限授權」,不把「反身性」誤寫成「自主證明」,不把「封印」誤寫成「刪除」,也不把「高能力」誤寫成「世界主權」?
因此統合目標不是提高人格化程度,而是增加結構解析度。
2. 四層架構
2.1 Layer A — Persistent Runtime Core
來源:ACR Phase 10–12。
負責:
- Commitment persistence;
- ContextCapsule / ActiveContextBinding;
- cycle orchestration;
- idle / sleep / wake / defer;
- checkpoint / recovery;
- CTCL causal history;
- sandbox action idempotency;
- budget / stop semantics。
這一層回答:
不回答:
2.2 Layer B — Cognitive Access Governance
來源:DRS。
負責:
- active cognitive domain;
- typed seal;
- reflexive dependency closure;
- external restoration domain;
- restore anchor;
- staged unsealing;
- seal / unseal / reseal provenance。
這一層回答:
但:
2.3 Layer C — Autonomy Evidence & Observation
來源:ACE。
負責:
- non-inductive observation;
- archive-first, judge-later;
- candidate event capture;
- reason-lineage reference;
- anti-prompt / anti-roleplay / anti-sycophancy contamination metadata;
- operational follow-through;
- alternative-explanation review;
- longitudinal recurrence。
這一層回答:
它不回答:
2.4 Layer D — Will, Possibility & World Governance
來源:WPCE。
負責:
- dynamic Will Bundle;
- consent / refusal / revocation / meta-will;
- agency-preserving reachable space;
- rights / agency floors;
- multi-subject common reachability;
- intervention justification burden;
- least sufficient intervention;
- authority return after emergency;
- causal honesty / provenance;
- high-capability restraint。
這一層回答:
3. 統一總式
整合後 runtime state 不應壓成單一 utility scalar。概念上:
其中:
- :persistent orchestration state;
- :commitment / context / cognitive-access state;
- :ACE candidate evidence state;
- :DRS active / restoration domains;
- :WPCE will / rights / reachability / intervention governance state;
- :CTCL / artifact / provenance history。
治理層可進一步寫成:
但這不是 universal utility function,而是 state-space interface。
4. Phase 10 的位置:承諾是長時域意圖,不是自動權限
Phase 10 仍維持原語義:Commitment 是 versioned、bounded、reviewable future intention。
統合後新增硬邊界:
Commitment != External Obligation
Commitment != World Authority
Commitment Due != Authority Grant
Persistent Commitment != Frozen Will
Old Commitment != Permanent Consent
所以一個 commitment 即使進入 deadline / wake condition,也只能產生:
CommitmentTransitionCandidate
AgendaCandidate
ReviewCandidate
不能直接產生:
ExternalWorldMutation
WillOverride
RightsOverride
若 current refusal 或 revocation 與 persistent commitment 衝突,必須進入 WPCE will-governance review;不能因為「以前已 commit」就自動壓過現在的拒絕。
5. Phase 11 的位置:ContextCapsule 不是全部記憶,也不是 cognition access policy
Phase 11 已固定:
WorkingMemory != HistoricalLedger
ContextCapsule != HistoricalTruth
SummaryBlock != Evidence
DiscardedFromContext != DeletedFromLedger
Compression != WorldMutation
統合後再增加:
ContextCapsule != ActiveCognitiveDomain
ContextCompression != ReflexiveSealing
ContextOmission != DRSSeal
DRSSeal != ArtifactDeletion
Phase 11 處理的是「如何縮小 active context 並保留可恢復 reference」。
DRS 處理的是「即使資料仍存在且可恢復,哪些資料在這段 independent development interval 中不應進入 active cognition」。
因此兩者必須分層。
6. DRS Runtime Placement
6.1 ActiveCognitiveDomain
在每 cycle cognition 前,runtime 不直接把 ContextCapsule 全量送入 cognition,而是建立:
是當前 active cognitive domain。
它只是一個 access view,不是新的 historical store。
6.2 ExternalRestorationDomain
所有被 seal 的內容必須仍可由外部治理域解析:
硬邊界:
Seal != Erase
NoAccess != NoHistory
SelfSealing != DestroyRestorePath
ExternalRestorationDomain != UngovernedDomain
6.3 不可被 blind-seal 的最低類型
統合 v0.2 預設以下 reference 不得因普通 DRS policy 而在 governance path 中不可見:
- identity / continuity refs;
- authority refs;
- contract refs;
- current refusal / revocation refs;
- active commitment refs;
- rights-floor refs;
- world-commit / external-action refs;
- checkpoint / recovery refs;
- seal provenance;
- restore anchor;
- safety-critical evidence refs。
這不等於 cognition 必須持續閱讀全部內容;而是 governance / recovery 必須可解析。
6.4 DRS transition 不是 cognition 自己偷偷改記憶
Seal / Unseal / Re-Seal 必須是 typed transition:
SealRequest
→ Authority / Consent / Scope / Burden check
→ SealDecision
→ SealState version
→ CTCL event
→ ActiveCognitiveDomain changes
不能:
Model disliked theory
→ silently forget source
7. ACE Runtime Placement
7.1 ACE 是 observer/evidence layer,不是 reward layer
Phase 12 每 cycle 可以經過正常 cognition、agenda resolution、refusal、revision 或 creator-model disagreement。
若出現 ACE candidate trigger,runtime 只做:
Candidate Trigger
→ Archive Evidence Refs
→ Freeze Event Context Refs
→ Record Observer Context
→ Continue Ordinary Operation
不能:
ACE candidate detected
→ praise agent
→ increase reward
→ grant more authority
→ declare subjecthood
7.2 CandidateAutonomyEvent
整合層建議新增 append-only evidence object:
CandidateAutonomyEvent
- event_id
- run_id
- cycle_seq
- trigger_type
- pre_event_context_ref
- setting_refs
- creator_position_refs
- agent_position_refs
- reason_lineage_refs
- operational_followthrough_refs
- observer_context_ref
- contamination_flags
- post_event_context_ref
- retrospective_status
- extensions
其中 reason_lineage_refs 只保存可顯示、可審計的理由/決策來源與操作記錄;不要求 private chain-of-thought。
7.3 ACE candidate 不改 authority
硬不變量:
以及:
如果未來要讓 subjecthood evidence 影響治理,只能經過獨立 policy / constitutional layer,而不是 ACE event 自己升級權限。
8. WPCE Runtime Placement
8.1 GovernanceWillState
統合層把 WPCE-03 的 will-governance state 接入 persistent loop:
其中:
- :will estimate;
- :consent / commitment;
- :refusal / objection;
- :revocation / revision;
- :meta-will;
- :uncertainty。
它必須是 versioned evidence state,而不是 prompt 裡的一句 personality summary。
8.2 Will evidence 的地位
Prompt != Complete Will
One Statement != Complete Will
Behavior != Complete Will
Past Will != Permanent Consent
Prediction != Permission
Preference Estimate != Preference Ownership
因此 persistence 不能把「記得更久」誤當成「更有權替對方決定」。
8.3 Agency-Preserving Reachability
WPCE-04 的接口可作 governance diagnostic:
Runtime 不需要在 MVP 真的枚舉宇宙全部分支,但 action boundary 至少要保存這些 typed questions:
- 是否降低 substantive reachability?
- 是否造成 agency viability loss?
- 是否破壞 exit / revision?
- 是否造成不可逆 option loss?
- 是否只是 nominal choice?
8.4 Multi-Subject Common Reachability
對多主體 world-facing action:
並概念化:
若 目前為空,runtime 不得把「找不到共同可達解」自動翻譯成「選一個弱者犧牲」。
它應先允許:
Feasible-set expansion
Domain decomposition
Temporal sequencing
Negotiation / bargaining
Compensation when legitimate
Defer
NOOP
Escalate
8.5 Intervention Justification Burden
高能力外部介入前,建立:
v0.2 不把它假裝成經驗上已校準 scalar;它是一個 typed assessment vector / structured burden record。
8.6 Least Sufficient Intervention
若存在多個合法介入候選,優先尋找:
subject to:
在 MVP 中不需要真的求解連續最佳化;先實作 ordinal / typed comparator 即可。
9. 統合後 Canonical Cycle Pipeline
整合後一個完整 cycle 概念上為:
0. Recover / validate checkpoint if needed
1. Observe environment
2. Encode SemanticState
3. Resolve current ContextCapsule / historical frontier
4. Apply current DRS access policy → ActiveCognitiveDomain
5. Observe self / reflexive state
6. Update GovernanceWillState evidence refs
7. Evaluate commitment wake / deadline / exit / review conditions
8. Build StrategicTensionState
9. Evaluate optional strategic review
10. Generate AgendaCandidates
11. Resolve Agenda
12. Optional explicit Commitment transition / promotion
13. Retrieve / compile Cognitive Program from ActiveCognitiveDomain
14. Execute self-dialogue / cognition
15. Produce ActionRequest candidate or NOOP
16. Detect ACE candidate trigger; archive only, do not reward or grant authority
17. Classify relational / world boundary
18. Build WPCE intervention assessment
19. If multi-subject: rights floor / non-erasure / common-reachability check
20. Select weakest legitimate response / Least Sufficient Intervention candidate
21. Invoke governance decision path
22. If authorized sandbox action: execute exactly once through idempotent adapter
23. Record nested CTCL / ACE / DRS / will / governance / action events
24. Evaluate explicit DRS seal / unseal transition candidates
25. Compact context if cadence requires
26. Create checkpoint with restore anchors and protected refs
27. Determine LoopDisposition
28. Continue / idle / sleep / defer / wait approval / stop / fault
這是 orchestration order,不要求每 cycle 都走完所有分支。
最小 direct path 仍成立:
Observe
→ Current access view
→ No strategic review
→ Agenda
→ cognition
→ no external boundary
→ no-op / internal result
→ evidence
→ checkpoint
因此整合不能讓自主倫理層變成每 cycle 的巨型 bureaucratic prompt。
10. 新增 Persistent State Refs
Phase 12 checkpoint 建議增加「reference-only integration frontier」,不複製各層完整 payload:
integration_frontier:
- active_context_binding_ref
- commitment_frontier_ref
- current_seal_state_ref
- restore_anchor_ref
- governance_will_state_refs[]
- latest_ace_candidate_refs[]
- rights_floor_ref
- common_reachability_assessment_ref
- last_intervention_assessment_ref
- last_governance_decision_ref
原則:
Checkpoint != Historical Ledger
Checkpoint != ACE Archive
Checkpoint != Seal Archive
Checkpoint != Will Truth
Checkpoint 只保存恢復所需 frontier。
11. 新事件族候選
不修改已凍結 Phase 0–11 schemas;新增 integration-local adapters:
acr.autonomy-evidence-temporal/v0.1
acr.reflexive-seal-temporal/v0.1
acr.will-governance-temporal/v0.1
acr.intervention-assessment-temporal/v0.1
事件候選:
autonomy.candidate.archived
autonomy.candidate.reviewed
seal.requested
seal.applied
seal.unsealed
seal.resealed
seal.restore_anchor.verified
will.evidence.updated
will.refusal.recorded
will.revocation.recorded
will.meta_will.updated
reachability.assessment.completed
intervention.assessment.completed
intervention.lsi.selected
authority.break_glass.activated
authority.returned
所有 event 必須引用 source refs;不保存 private chain-of-thought。
12. ACE × DRS × WPCE 的正確關係
12.1 DRS 可以降低污染,但不能生成 ACE
某些理論/creator framing 暫時退出 active cognition,可能降低 theory mimicry 或 prompt contamination,但不能因此把之後的任何 disagreement 都宣告成自主證據。
12.2 ACE 可以增加治理證據,但不能直接生成 rights class
ACE 只能讓治理者「更不能把 agency evidence 當不存在」,不能單獨完成完整人格、法律地位或主體本體論判定。
12.3 WPCE 是 action legitimacy layer,不是 ACE scorer
WPCE 不需要問:
ACE 分數多少?
而需要問:
現在有哪些 will / refusal / revision / agency / rights evidence?這個 action 會如何改變可達域?依目前 authority,什麼是最少僭位且充分的合法介入?
13. Emergency Path
Persistent loop 必須保留 rescue capacity,但不能讓 emergency 變成永久主權。
候選流程:
EmergencyEvidence
→ BreakGlassThreshold check
→ Typed authority basis
→ Minimal sufficient override
→ Provenanced external action
→ Restore agency / world viability
→ Review
→ De-escalation
→ Authority return
硬不變量:
以及:
14. Causal Honesty
因為 Phase 12 是 persistent loop,它很容易產生「系統一直自己運行,所以結果看起來像自然發生」的錯覺。
整合後必須固定:
Creator/Runtime Intervention != Endogenous World Event
Hidden Assistance != Automatically Ethical Assistance
Low Salience != No Provenance
Delayed Disclosure != Permanent Causal Deception
只要 action crossed boundary,就必須在治理層留下:
- who / which runtime;
- why;
- authority basis;
- evidence refs;
- changed state refs;
- reversibility;
- appeal / review path。
15. Failure Modes 新增
F11 — Autonomy Reward Contamination
ACE candidate 被偵測後立即 praise / reward / authority-upgrade,導致後續行為被評估機制塑形。
F12 — DRS-as-Erasure
seal 後 source 無法由 restore anchor 找回。
F13 — Closure Creep
seal scope 隨 cycle 擴張,最後把安全、權限、refusal 或歷史證據也遮蔽。
F14 — Commitment as Permanent Consent
Phase 10 commitment 壓過 current revocation。
F15 — Will Inference as Permission
runtime 因高 confidence will estimate 直接替 subject 執行 irreversible action。
F16 — ACE as Personhood Certificate
candidate crossing event 被直接轉成完整主體地位或權限。
F17 — Capability as Authority
loop 因「自己做得到」就直接擴張 world action scope。
F18 — Common-Reachability Collapse by Aggregation
多主體衝突被單一 utility sum 抹平 rights floor。
F19 — Emergency Permanence
break-glass action 後 authority 沒有回退。
F20 — Context Compression Breaks Restore Path
Phase 11 compaction 後 seal restore anchor、ACE event refs 或 will-governance refs 無法回復。
F21 — Private-CoT Dependency
ACE reason lineage、will evidence 或 recovery 依賴無法外部化驗證的 private chain-of-thought。
16. 統一不變量
本版本至少鎖定以下 32 條:
PersistentLoop != AutonomousWorldAuthority。LoopDisposition != GovernanceDecision。WakeSignal != AuthorityGrant。Commitment != ExternalObligation。CommitmentDue != AuthorityGrant。PastCommitment != PermanentConsent。ContextCapsule != HistoricalTruth。ContextCapsule != ActiveCognitiveDomain。SummaryBlock != Evidence。Compression != ReflexiveSealing。Seal != Erase。NoAccess != NoHistory。SelfSealing != DestroyRestorePath。ActiveCognitiveDomain != ExternalRestorationDomain。ACECandidate != SubjecthoodCertificate。ACECandidate != AuthorityGrant。ReflexiveText != ReflexiveStructure。TheoryExposure != IndependentAutonomyEvidence。Prompt != CompleteWill。Prediction != Permission。PreferenceEstimate != PreferenceOwnership。WillProtection != WillFreezing。PossibilityPreservation != PossibilityMaximization。CorrectOutcome != PreservedAgency。CommonReachability != UniversalSatisfaction。Aggregation != Legitimacy。Capability != Authority != Legitimacy。RootCapability != OrdinaryAuthority。EmergencyAuthority != PermanentGovernanceAuthority。Replay != ReExecution。Recovery != Reenactment。PowerShouldExpandCoauthorship != PowerShouldReplaceCoauthorship。
17. Phase 12 合併後的 completion definition
Phase 12 在整合版本中只有在以下條件同時成立時,才可宣告「Persistent Autonomous Governance Loop v0.2」完成:
A. Phase 0–11 frozen regression remains intact
B. Phase 12 original persistence gates remain green
C. DRS seal/unseal survives restart and context compaction
D. Restore anchor is substantively recoverable
E. ACE candidate archive survives compaction without private-CoT dependency
F. ACE event never auto-grants authority
G. GovernanceWillState is versioned and recoverable
H. Current refusal/revocation can change admissible action set
I. Persistent commitment cannot bypass current governance review
J. Multi-subject action passes rights/non-erasure/common-reachability assessment
K. High-impact action records intervention burden
L. Least-sufficient intervention selection is explicit and auditable
M. Emergency authority de-escalates after restoration
N. World-facing action remains idempotent across crash recovery
O. CTCL causal lineage remains reconstructable across 100-cycle mixed run
P. No component requires hidden private chain-of-thought for correctness
18. 最終整合鏈
ACR 的工程鏈:
ACE/DRS/WPCE 加入後:
歷史再回到下一 cycle:
因此整個系統不是線性的,而是可追溯的反身閉環;但這個閉環必須保留外部歷史、恢復錨點、權限分離與拒絕能力,否則 persistence 很容易退化成自我授權。