Sovereign Neural Box

Sovereign Neural Box is a small, bare-metal x86-64 inference appliance. It boots as a UEFI application, loads a packed ternary model, and serves fixed-size input frames over a UART stream. The runtime combines two integer-only model paths: a feed-forward MLP and recurrent causal linear attention.

Highlights

• Dual-architecture inference: 64→32→16 ternary MLP and 16-dimensional causal linear attention with a persistent 16×16 recurrent state.
• No-heap inference core: src/main.rs is #![no_std]; the core does not enable Rust alloc or use Vec/heap allocation. Model, DMA, frame, and state storage use fixed-size buffers. UEFI file I/O writes directly into the preallocated DMA-aligned shard buffer.
• UEFI x86-64 target: built for x86_64-unknown-uefi with the repository's nightly toolchain.
• User-friendly model updates: the GPT appliance image has a FAT32 EFI System Partition and a FAT32 NEURAL_DATA volume. Replace NEURAL_DATA:\weights.bin to load a different model at the next boot.
• Fallback behavior: UEFI SimpleFileSystem volumes are searched before ExitBootServices; invalid or missing files fall back to legacy raw-NVMe shard lookup, then to a safe built-in identity model.
• UART streaming: COM2 accepts 64 signed-byte inputs and returns 16 little-endian i32 outputs. The standalone NR control marker resets recurrent attention state.

Repository layout

• src/ — UEFI entry point, model kernels, shard parsing, UART, NVMe, and shared-memory support.
• tools/package_image.py — GPT/FAT32 appliance image builder; tools/package_image.ps1 is its PowerShell wrapper.
• tools/payload_builder/ — host-side Rust NEUR shard generator (mlp or attention).
• tools/test_dual_volume.ps1 — QEMU test for FAT-based model loading and UART streaming.
• tools/test_attention_sequence.ps1 — QEMU test for recurrent attention accumulation and reset.
• ml/ — optional model training and export utilities.
• DEPLOYMENT.md — detailed flashing, model-update, and server deployment guidance.

Model format

A NEUR shard begins with a 16-byte header: ASCII magic NEUR, little-endian version and input dimension, a model-type byte, little-endian output dimension, and a final hidden/attention dimension byte. Ternary weights use two bits per weight: 00 is zero, 01 is +1, and 11 is −1.

For attention, the recurrent state is a row-major 16×16 matrix of i32 values. It accumulates key/value outer products across frames and is reset by the UART control marker.

Build

Install the Rust nightly toolchain and the UEFI target listed in rust-toolchain.toml, then build the release EFI application:

cargo +nightly build --target x86_64-unknown-uefi --release

The core is no_std and uses fixed storage for model weights, I/O frames, and attention state. File-system protocol metadata may be managed internally by UEFI firmware; no Rust heap allocator is enabled by this crate.

Create the dual-volume image

Build the EFI binary first. The packager uses dist/production_shard.bin when present, or creates a small valid default MLP shard if it is absent.

python tools/package_image.py

This creates dist/neural_box_appliance.img, a GPT disk image with a protective MBR:

1. ESP: FAT32, contains \EFI\BOOT\BOOTX64.EFI and \STARTUP.NSH.
2. NEURAL_DATA: FAT32, contains the default model as \weights.bin.

For an Attention shard, generate it with the host builder and pass it as the packager's shard input (the default packaging path is dist/production_shard.bin):

cargo run --manifest-path tools/payload_builder/Cargo.toml --target x86_64-pc-windows-msvc --release -- --model attention --output dist/production_shard.bin
python tools/package_image.py

The image can also be prepared with tools/package_image.ps1. See DEPLOYMENT.md before writing an image to physical media.

UEFI model ingestion

Before leaving Boot Services, the application scans UEFI SimpleFileSystem handles using a fixed caller-owned handle array and searches for \weights.bin or \NEURAL_WEIGHTS\weights.bin. A valid shard is read directly into the 4 KiB DMA-aligned buffer and dispatched by its model-type field. If no file is found or parsing fails, the application attempts the legacy raw-NVMe locations; if that also fails, it runs the safe built-in fallback model.

To update a deployed appliance, mount the NEURAL_DATA FAT32 volume on a desktop OS and replace its root weights.bin with a valid NEUR shard. No EFI partition modification is needed.

UART protocol

• Input: NB followed by exactly 64 raw signed i8 bytes.
• Output: NR, one dimension byte, then that many little-endian i32 values (16 for both supported models).
• Attention reset: send standalone NR with no following payload. This is a control event and does not produce an output frame.

The UEFI streaming loop has a finite frame limit and timeout intended for appliance/QEMU verification. COM1 carries text diagnostics; COM2 carries the binary protocol.

QEMU verification

With QEMU installed and assets/OVMF.fd available:

cargo +nightly build --target x86_64-unknown-uefi --release
python tools/package_image.py
.\tools\test_dual_volume.ps1

The dual-volume test boots the GPT image with the disk attached as NVMe, sends eight input frames, verifies 16-element responses, and asserts that COM1 reports a FAT-loaded shard and zero dropped frames. To verify attention state and reset:

.\tools\test_attention_sequence.ps1

License

Licensed under either of:

• MIT License
• Apache License, Version 2.0

at your option.