KERNEL
Technical foundation for Covenant's sovereign AI infrastructure
AlfredOS, the sovereign AI operating system
AlfredOS is a deployable AI operating system for running private models and agentic workflows on infrastructure the user/organization controls. A deployment couples two systems, Lattice (the execution engine) and NOD3 (the graph-based interface and memory), running on the Conduit infrastructure framework. Execution is typed and compiled rather than driven by free-form tool calls, memory is an explicit graph rather than a context window that fills and degrades, and each tenant holds the keys to its own environment. Every customer runs one or more Lattice deployments with a NOD3 deployment per user.
Alfred Core
The modular components that make up AlfredOS.
→ Conduit
AI infrastructure framework. Type-safe, compute-provider-agnostic building blocks for deploying and composing open-source models. The foundation the stack runs on.
GitHub ↗→ Lattice
Policy and action execution engine. Typed, signed, composable actions; plan-before-execute; an async runtime with no tool calls.
Lattice SDK on GitHub, coming soon
→ NOD3
Graph-based interface, tightly coupled to Lattice. Conversation and memory as an explicit, branchable graph, durable state the user owns, not a context window that closes.
Security & Encryption
Security and encryption research and development.
→ Model Encryption Protocol (MEP)
Cryptographic inference with structure-preserving transforms. Encrypt model weights and inputs; execute on standard cloud GPUs without exposing IP or data.
→ Lattice Security Model
Signed action registration and Lattice-verified invocation, establishing a clean trust boundary between Lattice Core and the action surface.
Branching Innovations
Novel innovations developed in research that contribute to the performance of Alfred.
→ Model Data Language (MDL)
Type-safe model programming. Turns language models into structured, validated interfaces with automatic schema compilation.
→ LM Lite
Multi-model batching runtime. Efficient execution routing and replica management; multiple models per GPU at a fraction of vLLM's overhead.