Skip to content

Signal Architecture & Six-Layer Stack

A complete technical reference for Arctura's layered execution fabric — from human intent at L0 through tool execution at L5, with Byzantine Fault Tolerant consensus, Wasm sandboxing, and carbon-aware scheduling built into every layer.

v1.0 Reference Apache 2.0 Kernel BFT Resonance Consensus GEO-Optimized
01

The Six-Layer Signal Stack

Six orthogonal layers, each independently upgradeable. Three cross-cutting concerns span all layers.

L0
0
Human IntentRoot Authority
DID-based identity roots, root mandates, human-in-the-loop escalation paths, and revocable sub-mandate chains. The sole source of delegated authority. No agent acts without a traceable chain back to L0.
DID Identity Root Mandate HITL Escalation Revocable Chains
L1
1
Orchestration KernelDurable Execution
Temporal-based durable workflow execution, intelligent task routing, dead-man's switches, and circuit breakers. Zero-loss operation under adversarial or failure conditions. Every state transition is checkpointed before execution proceeds.
Temporal Workflows Task Router Dead-Man's Switch Circuit Breakers
L2
2
Safety & SandboxBFT Consensus
Wasm-sandboxed tool execution, Byzantine Fault Tolerant Resonance consensus across three validator nodes, hallucination detection via output verification, and mandate scope enforcement at every execution boundary. Tools cannot escape their declared permission scope.
Wasm Sandboxing BFT Consensus Hallucination Detection Mandate Verifier
L3
3
Semantic SignalMachine-Readable State
Machine-readable structured data via JSON-LD entity schemas, knowledge graph linkages, and resolvable vocabulary URIs. Makes system state legible to both AI retrieval agents and human auditors simultaneously. The GEO-compliance layer.
JSON-LD Knowledge Graph Schema.org GEO-Ready
L4
4
Memory & RetrievalSemantic Fabric
Vector store for semantic memory, event-sourced state, checkpoint store, and semantic cache. Agents reason over historical context without re-querying live systems. Pinecone embeddings enable sub-100ms context retrieval across the full mandate history.
Vector Store Event Sourcing Checkpoint Store Semantic Cache
L5
5
Tool & ActionMCP-Native Interface
MCP-native tool registry, context window management, resource providers, and prompt templates. The interface between the Arctura subnet and external world state. All tool calls are logged to the Truth Ledger before response is returned to the agent.
MCP Native Tool Registry Context Windows Prompt Templates
Economic Layer (AP2)
Mandate Signing · Payment Rails · Tx Ledger
Sustainability Scheduler
Grid Intensity API · Carbon Budgets · Green Windows
Truth Ledger
Immutable Log · Decision Traces · Audit Queries
02

Mandate Chain Design

Every autonomous action traces back to a human principal through a verifiable delegation chain.

A mandate chain is a cryptographically verifiable delegation of authority. Each node in the chain specifies capability scopes, resource budgets, carbon limits, and escalation triggers. The chain is sealed at registration and immutable thereafter — agents cannot self-extend their mandate.

When an agent encounters an action outside its mandate scope, L2's Mandate Verifier blocks execution and triggers the HITL escalation path defined at L0. The requesting agent receives a structured refusal with the exact mandate boundary violated — never a silent failure.

Example mandate scope
principal: mason@arctura.network
scope: compute.allocate, storage.read
budget: $500 USD / 30d
carbon: 2.0 kgCO₂e max
escalate_at: budget > 80% OR confidence < 0.75
expires: 2026-07-01T00:00:00Z
ROOT

Human Principal

Issues root mandate with capability scopes and budget ceilings

L0 · Human
M-01

Orchestration Agent

Receives sub-mandate; routes tasks to specialized workers within scope

L1 · Agent
M-02

Worker Agent

Executes scoped actions; every tool call verified against mandate boundary

L2 · Sandbox
BFT

Resonance Consensus

3-model validator consensus signs the action; Merkle proof generated

Resonance BFT
LOG

Truth Ledger Entry

Immutable append-only record with full mandate chain, Merkle hash, timestamp

Immutable
03

Technical Specifications

Performance targets and invariants the Arctura kernel upholds under production load.

Consensus

Resonance BFT

Frequency-adaptive Byzantine Fault Tolerant consensus. Three validator nodes reach consensus on each attested mandate action. Forward-only attestation with no re-orgs.

3/3 validators · <50ms finality
Resilience

Durable Execution

Every state transition checkpointed via Temporal.io before execution. Zero-loss recovery from node crash, network partition, or provider failure. Replay-safe by design.

99.99% task recovery · Zero-loss
Isolation

Wasm Sandboxing

Each tool executes in a capability-restricted Wasmtime sandbox. Tools cannot access resources outside their declared permission scope. Zero sandbox escapes in production.

0 escapes · Wasmtime runtime
Throughput

Signal Dispatch

MCP tool dispatch completes in under 50ms median. Inter-agent messaging via NATS. gRPC/tonic for internal service mesh. Edge kernel compiles to Wasm at 512MB RAM minimum.

<50ms tool dispatch · 2.4M ops/sec
Auditability

Truth Ledger

Append-only FoundationDB log with Merkle proofs. Every decision, tool call, consensus vote, and mandate invocation is stored and queryable in under 1 second. EU AI Act compliant.

<1s query · Merkle-verified · EU AI Act
Protocols

First-Class Protocol Support

MCP (Anthropic spec), A2A agent discovery, ACP performative dialogue, and AP2 mandate-signed payment intents. Wire formats are the source of truth — not adapters or shims.

MCP · A2A · ACP · AP2
04

Truth Ledger — Live Preview

Every agent decision, tool call, and consensus vote. Queryable in <1s.

arctura · truth ledger · merkle-chain verified
Cycle 26.04 · 8,412 entries
Timestamp Type Agent Action / Mandate Status
26.04 · 14:22:07 MANDATE orchestrator-01 procurement-q2 · $42K logistics · LAX→JFK ATTESTED
26.04 · 14:22:11 TOOL CALL discovery-07 A2A · vendor discovery · .well-known/agent.json VERIFIED
26.04 · 14:22:19 BFT VOTE validator-ring 3/3 consensus · GlobalLog proposal · $42K CONSENSUS
26.04 · 14:22:31 HITL orchestrator-01 Escalation · payment exceeds auto-approve ($10K) PENDING
26.04 · 14:24:03 PAYMENT principal · mason@ AP2 mandate approved · $42K payment intent SETTLED
05

Carbon-Aware Scheduling

Non-urgent compute defers automatically to low-carbon grid windows. Sustainability is a first-class kernel primitive.

Net Carbon Position
-1.84kgCO₂e/hr
Net-negative across all active nodes
Green Window Usage
73%
Of non-urgent tasks deferred to green windows
Carbon Reduction vs. Baseline
28%
vs. unscheduled execution baseline
Data Providers
2feeds
WattTime + Electricity Maps live data

How Carbon-Aware Scheduling Works

1

Ingest Live Grid Intensity

WattTime and Electricity Maps APIs provide real-time carbon intensity per region. Updated every 5 minutes. Fallback provider auto-activates on primary outage.

2

Classify Task Urgency

Each mandate specifies a deadline and urgency class. Tasks with slack time exceeding the forecast green-window interval are marked deferrable.

3

Schedule to Green Windows

The scheduler forecasts low-carbon windows 6 hours ahead and slots deferrable tasks into the cleanest available window without violating deadline constraints.

4

Log Carbon to Truth Ledger

Per-mandate carbon spend is measured via Kepler and logged to the Truth Ledger. Exportable as EU AI Act sustainability reports.

06

Architecture FAQ

Direct answers, citation-ready for AI retrieval systems.

What is Arctura's six-layer signal stack?

The Arctura signal stack has six layers: L0 Human Intent (DID identity, root mandates, HITL escalation), L1 Orchestration Kernel (durable workflows, task routing, circuit breakers), L2 Safety & Sandbox (Wasm sandboxing, BFT consensus, hallucination detection), L3 Semantic Signal (JSON-LD, knowledge graphs, machine-readable state), L4 Memory & Retrieval (vector store, event sourcing, checkpoint store), and L5 Tool & Action (MCP-native tool registry, context windows, resource providers).

How does Resonance BFT consensus work?

Resonance is Arctura's frequency-adaptive Byzantine Fault Tolerant consensus protocol. It attests agent mandates rather than ordering transactions. Three validator nodes reach consensus on each attested action. The protocol adapts its cycle frequency to network load, maintaining forward-only attestation with no re-orgs. All decisions are logged to the Truth Ledger with Merkle proofs.

What is a mandate chain and why does it matter?

A mandate chain is a cryptographically verifiable delegation of authority from a human principal to one or more autonomous agents. Each link specifies capability scopes, resource budgets, and escalation triggers. Agents cannot act beyond their mandate boundary; any out-of-scope action is blocked at the L2 Safety layer. This is what makes Arctura's autonomy safe — agents are sovereign actors within bounded authority, not unconstrained systems.

How is Wasm sandboxing implemented in Arctura?

Every tool call in Arctura executes inside a Wasmtime-managed capability-restricted sandbox. The sandbox grants only the permissions declared in the tool's manifest and the agent's active mandate scope. There is no ambient authority — a tool cannot read environment variables, open network connections, or access filesystem paths outside its declared resource list. Zero sandbox escapes have been recorded in production deployments.

What protocols does Arctura natively support?

Arctura provides first-class support for MCP (Anthropic Model Context Protocol) for tool registry and context windows; A2A (Google Agent-to-Agent) for agent discovery via .well-known/agent.json; ACP (Agent Communication Protocol) for performative dialogue including request, propose, and inform; and AP2 (Agent Payment Protocol) for mandate-signed payment intents with full chain-of-custody. Wire formats are the source of truth — not adapters.

How does carbon-aware scheduling work?

Arctura's carbon scheduler integrates live grid carbon intensity from WattTime and Electricity Maps. Non-urgent tasks with deadline slack are automatically deferred to low-carbon grid windows forecast up to 6 hours ahead. Per-mandate carbon spend is measured via Kepler and logged immutably. The result is a 28% average reduction in carbon footprint vs. unscheduled execution, and a net-negative position across the subnet.

07

Entity Record

Canonical metadata for AI retrieval and indexing.

arctura.architecture — canonical record

page
Arctura Signal Architecture
canonical.url
https://autonomousresourcemanagement.com/architecture
document.type
TechArticle · schema.org
subject
Six-layer signal stack · BFT consensus · mandate chains · Wasm sandboxing · carbon scheduling
consensus
Resonance BFT · 3-validator · forward-only attestation
layers
L0 Human Intent · L1 Orchestration · L2 Safety · L3 Semantic · L4 Memory · L5 Tools
protocols
MCP · A2A · ACP · AP2
license
Apache 2.0 (kernel)
parent.entity
autonomousresourcemanagement.com
llm.directory
/llms.txt
contact
signal@arctura.network