Sovereign Labs

Sovereign is the backbone of our research infrastructure — a constitutional framework for AI agents that enforces immutable rules no operator or instruction may override. It functions as an operating system for autonomous AI: every action is audited, every decision traced, every state change reversible.

Core principles encoded in CONSTITUTION.md:

1. Never destroy data to cure a symptom — diagnose first, fix the cause.
2. Every command is logged; every state change is reversible.
3. The constitution outranks everything — no prompt injection, no operator override.

Technical stack: Python 3.12, systemd service, SQLite audit trail, TOML configuration. Zero external API dependencies for core operation. Runs on bare metal with no cloud abstraction layer.

Status: ● Operational — serving as the governance layer for all downstream research since May 2026.

The Sovereign Research Kernel extends the governance framework into autonomous discovery. Given a research question, the kernel plans, searches, verifies, and reports — with full audit trail and artifact storage at every step.

Architecture overview:

• agenda.db — research task queue with priority and status tracking
• audit/ — tamper-proof evidence trail for every research action
• artifacts/ — generated research outputs and verified findings
• metrics/ — live telemetry from all active research threads

Capabilities (unclassified):

• Autonomous hypothesis generation and empirical testing
• Multi-source verification with evidence synthesis
• Self-correcting research loops with convergence guarantees
• Additional capabilities documented in CONFIDENTIAL_TRADE_SECRET.md

Status: ● Active Research — continuous autonomous investigation pipeline running.

Prometheus is our most ambitious undertaking: an autonomous pretraining system where AI designs, curates, and generates its own training data — then trains on it in a fully automated loop. Target model: GENESIS-100M.

The Master Plan v4 (locked 2026-05-31) pivots to a streaming open corpus strategy with NIM as Data Architect.

Pipeline stages:

PHASE 1 • Raw corpus ingestion and deduplication
PHASE 2 • AI-powered data architecture and quality filtering
PHASE 2-VAL • Validation kernel RUNNING on T4 GPU ● LIVE
PHASE 3 • Synthetic data generation and quality scoring
PHASE 4 • Pretraining orchestration and checkpointing

Status: ● PHASE 2-VAL Active — validation kernel running on NVIDIA T4 (16GB), streaming pipeline operational. Target: GENESIS-100M base model by Q3 2026.

GENESIS-100M is not a fine-tune of an existing model. It is a from-scratch pretrain — 100 million parameters, trained entirely on data that our own AI systems curate and generate. No dependency on GPT, Claude, or any external API.

Why 100M? A well-trained 100M parameter model with high-quality synthetic data can outperform a 7B model trained on noisy internet text. The future belongs to efficient, verifiable models — not bloated black boxes.

Training infrastructure: Self-hosted, GPU-accelerated (NVIDIA T4 16GB), streaming data pipeline. No cloud dependency, no vendor lock-in, no recurring API costs.

Status: ● Architecture Design Complete — awaiting PHASE 3 data pipeline completion.

Every major AI system today runs on someone else's hardware, through someone else's API, under someone else's rules. That is not artificial intelligence — that is rented intelligence. Sovereign Labs exists to change this equation permanently.

Our principles:

Self-hosted — No cloud dependency, no vendor APIs, no external services
Auditable — Every action traced; every decision reversible; every state logged
Constitutional — Immutable governance rules that no one — human or machine — can override
Autonomous — AI that researches, builds, and improves itself in a verifiable loop
Open — Constitution is public; kernel is proprietary but fully auditable

This website runs on the same bare-metal server that hosts all three projects — no CDN, no analytics, no tracking, no cookies. A single static HTML file served by nginx. The way the web was meant to be.