Architecture

Architecture

Conceptual Model: Six-Layer Governance Architecture

天极 (Heaven) → 人极 (Human) → 地极 (Earth) → 经卦 (Hexagram) → 别卦 (Trigram) → 爻变 (Mutation)

The six layers represent the semantic descent path of Agent governance from top-level constitution to microscopic mutation.

Layer	Description
天极 (Heaven)	Constitution / Meta-rules — highest governance level
人极 (Human)	Human approval — HITL/HOTL/HATL decision points
地极 (Earth)	Safety boundaries — zero-trust foundation
经卦 (Hexagram)	Role orchestration — dynamic role composition
别卦 (Trigram)	Capability contracts — constraint enforcement
爻变 (Mutation)	Self-evolution — variation and innovation

---

Runtime Model: Eight Trigrams Trust State Machine

乾(QIAN) → 坤(KUN) → 震(ZHEN) → 巽(XUN) → 坎(KAN) → 离(LI) → 艮(GEN) → 兑(DUI)

The Eight Trigrams are 8 trust states for Agents, using Gray Code transitions (hamming distance=1) to ensure each trust change modifies only one dimension.

Layer-to-Trigram Mapping

Six-Layer	Trigram State	Code Anchor
天极 (Constitution)	乾(QIAN) 天 — Highest trust	meta_agent_closure.py
人极 (Human Approval)	离(LI) 火 — HITL/HOTL/HATL	human/decision_api.py
地极 (Safety Boundary)	坤(KUN) 地 — Zero trust	zero_trust.py
经卦 (Role Orchestration)	震(ZHEN)/巽(XUN) — Dynamic	role_composer.py
别卦 (Capability Contract)	坎(KAN)/艮(GEN) — Constraints	capability_contracts.py
爻变 (Self-Evolution)	兑(DUI) 泽 — Mutation	self_executor.py

---

Five-Layer Runtime Architecture

┌──────────────────────────────────────────────────┐
│  Meta Layer: meta_agent_closure, meta_governance  │
│  Self-referential closure · Safety circuit breaker │
├──────────────────────────────────────────────────┤
│  Governance: EightTrigrams, FourPhase, RuleFreeze │
│  Trust state machine · Violation processing       │
├──────────────────────────────────────────────────┤
│  Orchestration: Self-*(8), RoleComposer, Saga     │
│  Self-evolution · Role workflow · Transactions    │
├──────────────────────────────────────────────────┤
│  Execution: Agent(9), Skill(4), Federation(3)     │
│  Multi-Agent scheduling · Skills · Consensus      │
├──────────────────────────────────────────────────┤
│  Infra: Audit(10), Trust(3), Safety(5), Memory(2) │
│  Observability · Trust scoring · Security lines   │
└──────────────────────────────────────────────────┘

---

Key Design Decisions

Decision	Rationale
64-state Gray Code FSM	Hamming distance=1 transitions guarantee stability
TLA+ Formal Verification	5 theorem proofs (Lyapunov convergence + Sperner completeness)
CircuitBreaker	3 consecutive failures → automatic lock + 30s cooldown
4-Level Security Decision Tree	Rule→Mode→SafetyGate→User, 97% automation rate
A2A + MCP Dual Protocol	Maximum ecosystem interoperability

---

TLA+ Formal Verification

MAREF includes 5 TLA+ specifications that formally verify:

1. Consensus Protocol — Byzantine fault tolerance
2. Integration Testing — Multi-Agent interaction correctness
3. Desktop Federation — Cross-device coordination safety
4. Lite Model — Simplified governance correctness
5. Safety Properties — Invariant preservation under all states

---

← Back to Home