DragonBox — Heterogeneous Sovereign Compute Appliance

DRAFT — Reclaimed 2026-09-06

TL;DR. DragonBox is the Nexus reference appliance for heterogeneous computing — CPU + GPU + NPU + FPGA + optional TPU/DPU, orchestrated through Silicon Topology Discovery and the ION Compute Fabric, secured by Compute Pledges. It is the hardware incarnation of SPEC-712: Heterogeneous Silicon Topology.

Why Heterogeneous?

The hardware industry has already become heterogeneous:

CPU      — Scalar, branchy, latency-sensitive
GPU      — SIMT, throughput-monster, batch-parallel
NPU      — Matrix/convolution inference, fixed function
TPU      — Systolic tensor arrays, training-scale
FPGA     — Deterministic logic fabric, microsecond latency
DPU      — Network/dataflow, packet processing
Remote   — Cluster/Cloud/Bare-Metal over the wire

The operating system industry has not caught up. DragonBox is the Nexus answer: the OS admits the silicon is heterogeneous, measures it, pledges it, schedules it.

"The CPU is not the computer. It is merely one execution domain among many." — SPEC-712

The Architecture

┌──────────────────────────────────────────────────────────────────┐
│                   NEXUS DRAGONBOX                                │
│                                                                  │
│   CPU                GPU               TPU                       │
│  ┌─────┐            ┌─────┐           ┌─────┐                   │
│  │ Zen │            │RX8900│           │ v4 │                   │
│  │  5  │            └─────┘           └─────┘                   │
│  └──┬──┘                                                     │
│     │                                                          │
│     ▼                                                          │
│  Sovereign Plane ─── Rumpk ─── ION ─── Cellular Memory        │
│     │                                                          │
│     ▼                                                          │
│  Compute Valve ─── decides where physics executes              │
│     │                                                          │
│     ▼                                                          │
│  ┌────────────────────────────────────────────────────────┐    │
│  │       Silicon Map (measured, signed)                  │    │
│  └────────────────────────────────────────────────────────┘    │
│     │                                                          │
│     ▼                                                          │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐       │
│  │ HASH     │  │ RS-MEM   │  │ MERKLE   │  │ INFER    │       │
│  │ Cell     │  │ Scrub    │  │ Verify   │  │ Cell     │       │
│  │ (GPU)    │  │ (GPU)    │  │ (GPU)    │  │ (NPU)    │       │
│  └──────────┘  └──────────┘  └──────────┘  └──────────┘       │
│                                                                  │
│              Sovereign Applications                             │
└──────────────────────────────────────────────────────────────────┘

Three planes, one root:

Sovereign Plane (CPU) — boot, interrupts, security, scheduling, commit. Always CPU.
Compute Valve — the decision layer that picks where physics executes.
Matter Plane (GPU / NPU / TPU / FPGA / DPU / Remote) — the measured execution domains.

Editions

DragonBox ships as KDL recipe editions, each tailored to a target hardware class.

NEXUS_DRAGON_AI — Sovereign AI Workstation

kdl

edition "NEXUS_DRAGON_AI" {

    target {
        cpu "x86_64"
        gpu "vulkan-1.3+"      # required
        tensor "optional"       # preferred if present
        npu "optional"          # preferred if present
        fpga "unsupported"
    }

    compute {
        enable "hash"
        enable "crypto"
        enable "rs-memory"
        enable "merkle-verify"
        enable "inference"

        backend {
            hash         { prefer "gpu"; fallback "cpu"; threshold_batch 1024 }
            rs-memory    { prefer "gpu"; fallback "cpu" }
            inference    { prefer "npu"; fallback "gpu"; fallback "cpu" }
        }
    }

    silicon-probe {
        require "gpu.kind in {VECTOR, TENSOR}"
        require "gpu.vram >= 4GiB"
    }
}

Use cases: Sovereign AI inference, self-hosted LLM, local model training, sovereign cloud compute.

NEXUS_DRAGON_EDGE — Edge AI / IoT / Robotics

kdl

edition "NEXUS_DRAGON_EDGE" {

    target {
        cpu "arm64"
        npu "any"            # required
    }

    compute {
        enable "inference"

        backend {
            inference { prefer "npu"; fallback "cpu" }
        }
    }

    silicon-probe {
        require "npu.context >= 1"
    }
}

Use cases: On-device inference with no cloud dependency, robotics, IoT sensors, drones.

NEXUS_DRAGON_WORKSTATION — Sovereign Developer / Power User

kdl

edition "NEXUS_DRAGON_WORKSTATION" {

    target {
        cpu "x86_64"
        gpu "vulkan-1.3+"      # required
    }

    compute {
        enable "hash"
        enable "crypto"
        enable "rs-memory"
        enable "merkle-verify"

        backend {
            /* default: prefer GPU, fallback CPU */
        }
    }
}

Use cases: Sovereign build farm, content creation, scientific computing.

NEXUS_DRAGON_CLUSTER — Sovereign Cloud / Bare-Metal

kdl

edition "NEXUS_DRAGON_CLUSTER" {

    target {
        cpu "x86_64"
        gpu "vulkan-1.3+"
        remote "yes"           # multi-node Remote execution domain
    }

    compute {
        enable "hash"
        enable "inference"

        cluster {
            fabric "utcp"
            remote-dispatch "yes"
        }
    }
}

Use cases: Multi-node sovereign cloud, federated inference, sovereign backend.

What DragonBox Is Not

Not a GPU-First OS

DragonBox does not boot on a GPU. The CPU is the Sovereign Plane; the GPU is delegated matter. Boot is always a CPU responsibility.

Not an AI-Driven OS

AI may advise (predicting where a task should run). It never has authority (deciding where a task should run). The deterministic scheduler decides; the AI predicts.

Not a CUDA Lock-In

Vulkan Compute is the default backend. CUDA, ROCm, Metal, OpenCL are adapters under std.graft.*. The Nexus capability surface is CAP_*-class only — applications never see vendor SDKs.

Not a Tensor-Native Kernel

Tensor cores, NPUs, and GPU kernels produce results. They do not host sovereign authority. The CPU validates and commits.

The Silicon Map

At boot, Nexus discovers every execution substrate and measures its delivered capability — not its marketing label:

Machine: AMD-HAL-9000

CPU:
  8x Zen 5 (cores 0..7), 4.6 GHz boost
    measured FP32: 1.4 TFLOPS aggregate
    measured wake-from-WFI: 220 ns
    classification → 4× APEX, 4× FORGE

GPU:
  Radeon RX 8900 (RDNA 4)
    measured PCIe5 latency: 8 µs
    measured launch overhead: 12 µs
    measured FP16 throughput: 320 TFLOPS
    measured FP32 throughput: 80 TFLOPS
    classification → VECTOR (primary), TENSOR (secondary)

NPU:
  50 TOPS INT8
    measured inference latency @ 224x224: 1.4 ms
    classification → NEURAL

The Silicon Map is signed by the HAL's Ed25519 identity and exported to TruthDB. Every consumer — scheduler, capability kernel, application layer — reads the measured fact.

The Compute Pledge

Compute is a capability. A process declares what it needs; the kernel verifies; the user consents.

text

Process holds:    gpu.queue[12], gpu.vram[2GB], gpu.kind=VECTOR
Silicon Map:      GPU has queues 0..16, 16GB total
Binding:          Cell HASH_SERVICE owns queue 12, 4GB VRAM

Result:           ✓ The submission is permitted.
                  The hardware is told only about queue 12.
                  The capability kernel refuses any other queue.

Forbidden pledges (kernel-level rejection, not configurable):

gpu.dma.host_memory — never
gpu.unrestricted_queues — never
npu.arbitrary_model — never
fpga.unsigned_bitstream — never

Roadmap to Q4 2027

DragonBox is an 8-phase rollout from doctrine to appliance. The full roadmap is at .agents/roadmaps/ROADMAP-0.4-HETEROGENEOUS-SILICON.md.

Phase	Milestone	Target
0 — Doctrine	SPEC-712 ratified	Done with this delivery
1 — Topology Discovery	Signed SiliconTopologyGraph in TruthDB	Q4 2026
2 — GPU as ION Device	Vulkan Compute kernel via ION	Q1 2027
3 — BLAKE3 GPU Cell	100k chunks on GPU, ≥10× CPU	Q2 2027
4 — Compute Valve	Dynamic CPU/GPU routing, auditable	Q2 2027
5 — Accelerator Cells	GPU queue + VRAM isolation	Q3 2027
6 — Multi-Accelerator	Video pipeline CPU+GPU+NPU	Q3 2027
7 — Tensor-Native	SPEC-012 Cortex RS scrub on GPU	Q4 2027
8 — DragonBox Appliance	2 physical reference machines green	Q4 2027

Use Cases

Sovereign AI Inference

Run an LLM locally on a DragonBox without trusting a hyperscaler. The model is signed (BLAKE3 CID); the NPU is bound to that model; the CPU commits every inference result with ProvChain.

Self-Hosted Model Training

Sovereign model artifacts, sovereign compute. No data leaves the box. ProvChain records every training step.

Edge AI

On-device inference with no cloud dependency. NEXUS_DRAGON_EDGE edition targets NPU-equipped ARM SoCs (Qualcomm, Apple, ARM Ethos).

Sovereign Cloud Compute

A DragonBox cluster forms a sovereign backend. Multi-node Remote execution domain (Phase 8 + ROADMAP-0.4 §2 Phase 8 Cluster edition).

Heterogeneous Scientific Computing

Reed-Solomon memory scrub + Merkle DAG verify + matrix math on a single machine. The same kernel runs on CPU, GPU, and NPU — the Compute Valve chooses.

Status

DRAFT — Pending Founder ratification of SPEC-712. Implementation begins in lane feature/voxis/spec-712-heterogeneous-silicon after ratification.

Reference hardware matrix, signed edition recipes, and physical appliance certification are Phase 8 deliverables (target: Q4 2027).

Documentation

Document	Purpose
SPEC-712: Heterogeneous Silicon Topology	Constitutional doctrine for heterogeneous compute
RFC: SPEC-712	The conversation that produced the SPEC
ROADMAP-0.4	8-phase rollout to DragonBox
MANIFESTO	Libertaria Manifesto
NORTH-STAR	Constitutional vision
DESIGN_GOALS	System Truth Store canon

Naming Note

This page supersedes an earlier draft that described DragonBox as a DragonflyBSD/HAMMER2 storage distribution (Proxmox replacement). That earlier concept was speculative and has been renamed DragonStore in a follow-on pass.

The DragonBox name is now reserved for the Heterogeneous Sovereign Compute Appliance described above. If you came here looking for the DragonflyBSD storage distribution, it will be at dragonstore.md (TBD) once that follow-on SPEC lands.

The hardware industry has already become heterogeneous.
The operating system simply hasn't admitted it yet.
Nexus admits it.

— Voxis Forge, 2026-09-06

DragonBox — Heterogeneous Sovereign Compute Appliance ​

Why Heterogeneous? ​

The Architecture ​

Editions ​

NEXUS_DRAGON_AI — Sovereign AI Workstation ​

NEXUS_DRAGON_EDGE — Edge AI / IoT / Robotics ​

NEXUS_DRAGON_WORKSTATION — Sovereign Developer / Power User ​

NEXUS_DRAGON_CLUSTER — Sovereign Cloud / Bare-Metal ​

What DragonBox Is Not ​

Not a GPU-First OS ​

Not an AI-Driven OS ​

Not a CUDA Lock-In ​

Not a Tensor-Native Kernel ​

The Silicon Map ​

The Compute Pledge ​

Roadmap to Q4 2027 ​

Use Cases ​

Sovereign AI Inference ​

Self-Hosted Model Training ​

Edge AI ​

Sovereign Cloud Compute ​

Heterogeneous Scientific Computing ​

Status ​

Documentation ​

Naming Note ​