Technical Whitepaper v1.1 | SigmaNex Protocol

ABSTRACT

The centralization of Large Language Model (LLM) inference introduces a critical systemic risk to global knowledge availability. Dependency on cloud APIs creates a "kill-switch" for intelligence during infrastructure failures or geopolitical censorship events.

SigmaNex proposes a hardware-agnostic, portable runtime environment capable of serving quantized LLMs on consumer-grade hardware without internet connectivity. This paper outlines the "Zero-Trace" architecture, memory management strategies for low-resource inference, and the ethical imperative for decentralized intelligence reserves.

1. Preface: The Enclosure of Commons

In the 21st century, cognitive labor is being rapidly outsourced to algorithmic systems. However, unlike the engines of the Industrial Revolution which could be owned by factories, the engines of the Cognitive Revolution are renting their intelligence from a handful of centralized providers.

This centralization poses an existential threat to autonomy. If the "oracle" resides in a server farm in Northern Virginia, what happens when the fiber optic cables are severed? What happens when the Terms of Service change to exclude medical advice during a pandemic?

SigmaNex starts from a different axiom: Intelligence is a utility that must survive the collapse of the grid.

2. Technical Architecture

The SigmaNex protocol is designed to operate on "scavenged hardware"—standard x86_64 laptops and desktops found in typical residential or office environments. We reject the assumption of H100 clusters.

2.1 The Inference Kernel

The core engine utilizes a custom build of llama.cpp, optimized for CPU inference using AVX2 and AVX-512 instruction sets. By implementing generalized matrix multiplication (GEMM) directly on the CPU, we achieve acceptable token generation rates (5-12 t/s) on purely commodity silicon.

To achieve this, we rely on the GGUF file format, which allows for aggressive quantization (reducing 16-bit floats to 4-bit integers) with negligible perplexity loss.

[ USB STORAGE ] --> [ RAM DISK (TmpFS) ]
|
+--> [ BOOTLOADER ] --> [ KERNEL 6.8 ] --> [ LLAMA RUNTIME ]
|
+--> (AVX2 INSTRUCTIONS)

            // SigmaNex Memory Allocation Strategy (Pseudocode)
            void allocate_tensors(Model model) {
            // Detect hardware capabilities
            bool has_avx2 = check_cpu_flag("avx2");
            bool has_cuda = check_gpu_cuda();

            if (has_cuda && vram > model.size) {
            // Offload 100% layers to GPU
            offload_layers(model.all_layers);
            } else {
            // Fallback to CPU Zero-Copy mmap
            // Critical for survival on non-gaming laptops
            mmap_load(model.path);
            enable_numa_optimization();
            }
            }
        

2.2 The Persistence Layer

The OS is stripped of all telemetry. It boots directly into a read-only SquashFS filesystem. Modifications to the system state are written to a RAM overlay (OverlayFS), meaning they vanish instantly upon reboot.

For user data that must survive (e.g., medical logs, map annotations), we utilize a strictly isolated LUKS-encrypted partition on the USB drive itself. The decryption key is never stored on the host machine.

3. Security Architecture: Zero-Trace

In reduced-trust environments, the physical capture of hardware is a legitimate threat model. SigmaNex adheres to the "Zero-Trace" doctrine:

No Registry Writes: The bootloader does not touch the host Windows Registry.
No Pagefile Usage: We disable swap/paging to prevent RAM contents (including decrypted model weights and chat context) from being written to the host HDD.
Air-Gap Default: The networking stack is disabled by default at the kernel level. Checksum verification of models is done offline via pre-calculated hashes.

4. Knowledge Assurance (RAG)

A disconnected AI cannot "Google" the answer. To prevent hallucinations—which can be fatal in medical contexts—we implement a **Static RAG (Retrieval Augmented Generation)** system.

The USB image ships with a vector database (ChromaDB) containing embeddings of verified, public domain survival literature:

"Where There Is No Doctor" (Hesperian Health Guides)
"US Army Survival Manual FM 3-05.70"
"Standard First Aid & CPR" (Red Cross)

When a user asks a medical question, the system intercepts the query, retrieves the relevant text chunk from these trusted documents, and injects it into the context window with a system prompt: "Answer solely based on the provided context."

5. Governance & Roadmap

SigmaNex is an initiative led by Lead Engineer Martin Adrian Sebastian. We follow a "Supporter-First" development model: The codebase is open and accessible to verified supporters throughout the SDLC (Software Development Life Cycle), ensuring transparency and community auditability before the final hardened release.

Roadmap 2026-2027

Q4 2026: Alpha Release (USB Image v0.1)
Q2 2027: Mesh Networking (LoRa) for text-based off-grid consensus.
Q4 2027: Hardware "Referral Architecture" (Solar + Compute cases).

CITATION:
Sebastian, M. A. (2026). The SigmaNex Protocol: Sovereign Intelligence Architecture. Draft v1.1. https://sigmanex.net/whitepaper

SigmaNex Protocol.