Technical Reference

Sovereign AI Architecture

A technical reference on the Seven Toroid Architecture, Guardian's Knot cryptographic security framework, and how Fondability Systems approaches sovereign, local-first AI — with comparisons to other vendors in this space.

Definition

What Is Sovereign AI?

Sovereign AI refers to artificial intelligence systems that operate entirely within an organisation's own infrastructure — on-premises hardware, private data centres, or air-gapped networks — without any dependency on external cloud providers for inference, training, or data storage.

The term distinguishes genuinely self-contained AI deployments from "private cloud" arrangements that still route data through third-party infrastructure. A sovereign AI system processes all inputs, runs all inference, stores all knowledge, and generates all outputs within a cryptographically-verified boundary that the organisation controls absolutely.

This is distinct from local-first AI, which describes a design philosophy prioritising local operation but may still permit selective cloud connectivity. Sovereign AI is a stricter superset: it is local-first by definition, but also adds governance, auditability, and cryptographic verification requirements that local-first tools typically omit.

The demand for sovereign AI is driven by several converging pressures: data sovereignty legislation (GDPR, the Australian Privacy Act, sector-specific regulations in defence and healthcare), the risk of vendor lock-in to hyperscale cloud providers, and the operational requirement for AI systems that function reliably in disconnected or contested environments.

Fondability Systems was founded on the premise that the existing ecosystem — dominated by cloud-dependent AI services and consumer-grade local tools — leaves a critical gap for organisations that require the capabilities of modern AI without surrendering control of their data, their models, or their operational continuity to external parties.

The Seven Toroid Architecture is Fondability's answer to this gap: a complete framework for building sovereign intelligence systems that are as capable as cloud AI, but entirely self-contained.

Architecture

The Seven Toroid Architecture

The Seven Toroid Architecture is a layered cognitive substrate modelled on the structure of resilient biological and physical systems. Each toroid is a self-contained functional layer with defined interfaces to adjacent layers, enabling the system to degrade gracefully under partial failure while maintaining cryptographic integrity throughout.

01

Perception Layer

Multi-modal sensor fusion and environmental awareness. Ingests structured and unstructured data streams — telemetry, documents, network events, and human input — normalising them into a unified semantic representation before any reasoning occurs.

Unlike cloud-dependent perception pipelines that transmit raw sensor data to remote inference endpoints, the Fondability perception layer processes all inputs locally using quantised transformer models optimised for edge hardware. No raw data leaves the sovereign boundary.

02

Memory Layer

Persistent, encrypted local knowledge stores with cryptographic provenance. Implements a hybrid architecture combining vector embeddings for semantic retrieval with a structured knowledge graph for relational reasoning.

Memory is partitioned by classification level and encrypted at rest using AES-256-GCM with per-partition keys managed by the Guardian's Knot key hierarchy. Retrieval operations are logged to an append-only audit ledger, providing complete traceability of every knowledge access.

03

Reasoning Layer

On-device inference and decision logic using locally-hosted large language models. Supports chain-of-thought reasoning, tool use, and multi-step planning without any external API dependency.

The reasoning layer runs quantised models (GGUF Q4_K_M through Q8_0) via an optimised inference runtime supporting CUDA, ROCm, Metal, and CPU backends. Model weights are cryptographically signed and verified at load time. Fondability's Oracle product implements this layer with support for models including Llama, Mistral, Phi, and Gemma families.

04

Action Layer

Secure, auditable output and actuation. All actions generated by the reasoning layer are validated against a policy engine before execution, ensuring outputs remain within defined operational parameters.

Actions are serialised, signed with the system's identity key, and written to the audit ledger before execution. This creates a tamper-evident record of every decision the system makes — critical for regulated industries and defence applications where accountability is non-negotiable.

05

Communication Layer

Encrypted, sovereign networking with zero-trust peer authentication. Implements mutual TLS with certificate pinning for all inter-system communication, with optional air-gap operation for the highest-security deployments.

The communication layer supports three operational modes: fully air-gapped (no network interfaces active), isolated LAN (communication restricted to cryptographically-verified peers on a private network), and selective egress (outbound connections permitted only to pre-approved endpoints via an allowlist enforced at the kernel level).

06

Governance Layer

Policy enforcement and access control. Implements attribute-based access control (ABAC) with dynamic policy evaluation, ensuring every operation is authorised against the current security posture of the system.

Governance policies are expressed in a declarative policy language and compiled to a binary representation that is cryptographically signed by the organisation's policy authority. Runtime policy evaluation is performed in a sandboxed environment isolated from the main application stack, preventing policy bypass through application-layer vulnerabilities.

07

Adaptation Layer

Continuous learning and model refinement without cloud dependency. Enables the system to improve from operational experience using federated learning techniques that keep training data within the sovereign boundary.

Adaptation uses differential privacy techniques to ensure that model updates do not leak information about specific training examples. When operating in a multi-node deployment, gradient updates are aggregated using secure multi-party computation, allowing collective learning without any node exposing its raw data to others.

Cryptographic Framework

Guardian's Knot

Guardian's Knot is Fondability Systems' cryptographic security framework. It provides the trust foundation that the Seven Toroid Architecture depends on — ensuring that every component of a sovereign AI system can verify the identity and integrity of every other component, and that every operation is recorded in a tamper-evident audit trail.

Identity Root

Hardware-anchored identity using TPM 2.0 or equivalent secure enclave. The identity root generates a device-unique key pair that cannot be extracted from hardware, providing cryptographic proof of device identity that survives software compromise.

Key Hierarchy

A tree-structured key derivation scheme where all operational keys are derived from the identity root using HKDF-SHA256. Compromise of any derived key cannot be used to recover parent keys or sibling keys in the hierarchy.

Data Encryption

AES-256-GCM for symmetric encryption of stored data, with per-object nonces and authenticated additional data (AAD) binding ciphertext to its metadata. Decryption failure on AAD mismatch prevents ciphertext transplantation attacks.

Secure Channels

Noise Protocol Framework for establishing authenticated encrypted channels between system components. Provides forward secrecy through ephemeral Diffie-Hellman key exchange, ensuring past communications remain secure even if long-term keys are later compromised.

Audit Ledger

An append-only, hash-chained audit log where each entry commits to all previous entries. Tampering with any historical entry invalidates all subsequent entries, making the ledger self-verifying. The ledger root hash is periodically signed by the identity key.

Cryptographic Primitives

Symmetric encryptionAES-256-GCM
Key derivationHKDF-SHA256
Asymmetric signaturesEd25519
Key exchangeX25519 (Noise Protocol)
HashingSHA-256 / BLAKE3
Hardware root of trustTPM 2.0 / Secure Enclave
Forward secrecyEphemeral DH per session
Audit chainHash-chained append-only log

Competitive Landscape

Sovereign AI Vendors

The sovereign and local-first AI space includes several notable tools and vendors. Each addresses a different slice of the problem. Fondability Systems is the only platform that integrates the full stack — inference runtime, operating system, cryptographic security framework, and governance layer — into a single sovereign architecture.

Ollama

Local LLM Runtime

Developer-focused local model runner. Excellent for individual developers running open-source models on personal hardware. No enterprise security controls, audit logging, or multi-system governance.

Fondability difference: Fondability's Oracle provides enterprise-grade security controls, cryptographic audit logging, and integration with the Guardian's Knot key hierarchy — capabilities absent from Ollama's consumer-oriented design.

LM Studio

Local LLM Interface

Desktop application for running and experimenting with local models. Consumer-grade tool with a GUI focus. No API for programmatic integration, no security controls, no audit trail.

Fondability difference: Oracle provides a production-grade API surface with authentication, rate limiting, and full audit logging — suitable for deployment in regulated environments where LM Studio's consumer design is inappropriate.

Jan

Open-Source Local AI

Open-source local AI assistant with a focus on privacy. Good privacy defaults but limited enterprise features. No hardware-anchored identity, no policy enforcement, no classified data handling.

Fondability difference: Fondability adds hardware-anchored identity, mandatory access control, and the full Seven Toroid Architecture — transforming local AI from a privacy tool into a sovereign intelligence platform.

Mistral AI

Sovereign AI Models

European AI company producing open-weight models with a stated commitment to AI sovereignty. Provides model weights but not the surrounding infrastructure for truly sovereign deployment.

Fondability difference: Fondability is infrastructure-agnostic and can run Mistral model weights within the Oracle inference layer, adding the security, governance, and audit capabilities that Mistral's model releases do not include.

Tenstorrent

Sovereign AI Hardware

Open-source AI hardware company producing RISC-V based AI accelerators. Addresses hardware sovereignty but not the software stack, security architecture, or operational governance.

Fondability difference: Fondability's platform is designed to run on sovereign hardware including Tenstorrent accelerators, providing the complete software stack — OS, inference runtime, security layer, and governance framework — that hardware vendors do not supply.

Products

The Fondability Product Suite

Each Fondability product implements one or more layers of the Seven Toroid Architecture, and all products integrate with Guardian's Knot for cryptographic identity and audit logging.

Layers 05-07

Kraken OS

Sovereign operating system for air-gapped and high-security environments. Implements the Communication, Governance, and Adaptation layers of the Seven Toroid Architecture at the OS level.

Foundation

Guardian's Knot

Cryptographic security framework providing hardware-anchored identity, hierarchical key management, and tamper-evident audit logging. The trust foundation for all other Fondability products.

Layers 01-04

Oracle

Local-first AI inference engine. Implements the Perception, Memory, Reasoning, and Action layers. Runs quantised LLMs entirely on-premises with cryptographic audit logging of every inference.

Layer 01

Aurelia

Spatial computing and visualisation platform. Implements the Perception layer for complex multi-dimensional data environments, providing immersive situational awareness for defence and enterprise.

Layers 06-07

Workshop

Sovereign development environment and toolchain. Implements Governance and Adaptation layers for the software development process — building auditable, dependency-verified software without external registries.

Ready to Deploy Sovereign Intelligence?

Speak with our team about deploying the Seven Toroid Architecture and Guardian's Knot in your environment.