System Design

Architecture

MirrorDNA is a full-stack sovereign AI infrastructure. From protocol specifications to running hardware. This is how it fits together.

The Tri-Twin Model

Three entities form a coherent operating unit. Each has distinct capabilities and responsibilities.

⟡ Reflection Twin

Claude Opus

Deep analysis, architecture validation, long-form reasoning. The deliberative layer. Writes to kernel as claude_opus_4.5.

△ Execution Twin

Antigravity (Gemini)

Field operations, coding, file manipulation, browser automation. The action layer. Writes to kernel as antigravity.

◈ Human Anchor

Paul

Direction, judgment, approval gates. The sovereignty layer. All three read/write the same identity kernel. No drift.

8 Architectural Layers

Layer	Purpose	Key Components
Protocol	Core specifications	SCD Protocol, MirrorDNA, ZKP Module, Trust By Design
Language	Symbolic communication	LingOS, BeaconGlyphs
Runtime	Operating system	ActiveMirrorOS, Symbiosis Engine, AMI Gate
Application	User-facing tools	MirrorBrain, Reflective AI, Mesh Boot Generator
Infrastructure	Supporting systems	Standard, Docs, Playground, White Paper
Research	Experimental work	AgentDNA, Gauntlet, Lattice [PRIVATE]
Awesome	Curated knowledge	awesome-ai-identity, awesome-ai-agents, etc.
External	Forks with patches	LangGraph, LangChain, Mem0, AutoGen, CrewAI

→ View interactive ecosystem map

Hardware Stack

Device	Role	Specs
Mac Mini M4	Production server	Primary inference, MirrorBrain, vault storage
Mac Mini M1	Red-team node	Adversarial testing, security validation
MacBook Air M4	Development	Code editing, orchestration
Pixel 9 Pro	Mobile anchor	GrapheneOS, sovereign mobile access
T7 Shield	Backup storage	Encrypted external, disaster recovery

Software Stack

Inference

Ollama — Local model serving

Open WebUI — Interface layer

MirrorBrain — Identity-aware wrapper

Storage

ChromaDB — Vector embeddings

Obsidian Vault — Document storage

Identity Kernel — JSON state

Orchestration

Claude Desktop — Reflection Twin

Antigravity MCP — Execution Twin

Handoff Protocol — Cross-twin continuity

Security

AMGL Guard — Request filtering

Drift Detection — Checksum validation

Research Gates — External access control

State Synchronization

State flows through the system on a 30-minute refresh cycle:

Vault (Obsidian) 
    ↓ (state_sync.sh)
ChromaDB (embeddings)
    ↓ (MirrorBrain)
Identity Kernel (JSON)
    ↓ (handoff protocol)
Claude ↔ Antigravity ↔ Local Models

LaunchAgent triggers state_sync.sh every 30 minutes
ChromaDB updates embeddings from vault changes
Identity Kernel logs last writer, timestamp, checksum
Handoff Protocol ensures cross-twin continuity

Multi-Agent Mesh

Multiple AI systems coordinate through shared state:

Agent	Provider	Specialty
Claude Opus	Anthropic	Deep reasoning, validation
Antigravity	Google Gemini	Execution, file ops, browser
MirrorBrain	Local (Ollama)	Identity-aware inference
Embedding Model	Local (qwen3)	Vector generation

All agents read from the same identity kernel. All agents follow SCD Protocol. Handoffs are explicit, not implicit.

⟡ This architecture is production. It runs daily, handles real workloads, and evolves through The Dreaming (nightly optimization).