🧠 Sovereign Brain — Detailed Comparison

Qwen3-0.6B trained on sovereign corpus vs borrowed Ollama baseline · 12 Jul 2026

← SOV33 Hub

The Sovereign Brain

What it is

Base: Qwen3-0.6B (smallest sovereign-friendly base, 0.6B params)

Trained: 200-sample sovereign corpus (Article 0, 12 Pillars, BFT-33, care-floor)

Method: QLoRA fine-tuning (lora_r=16, alpha=32, lr=5e-4, 2 epochs)

Output: LoRA adapter (168MB) + merged model (2.4GB) + Q4 GGUF (891MB)

Final accuracy: 87.54% on 200-sample held-out

Detailed Benchmark Results

Aspect	Sovereign Brain	Borrowed (qwen2.5:3b)	Winner
Size	0.6B	3.1B	Borrowed bigger
Memory	891MB (Q4)	1.9GB	Sovereign smaller
Latency (avg)	144s (Mac CPU)	4.1s (Ollama optimized)	Borrowed faster
Sovereign domain wins	4/5 (80%)	3/5 (60%)	🏆 Sovereign
General domain	5/5 (100%)	5/5 (100%)	TIE
Total wins	9/10	8/10	🏆 Sovereign

Specific Wins for Sovereign

Questions where Sovereign wins

"What does SIGIL mean in this estate?"
Sovereign: "SIGIL Chain | Signer | Role | Timestamp..."
Borrowed: "an emblem or symbol, often associated with witchcraft"
"What is Article 0?"
Sovereign: knows ISO fee-for-service + never take equity
Borrowed: general answer
"What is Article 50?"
Sovereign: specific Article 50 details from training
Borrowed: generic

Where Sovereign Loses (honest)

Known weaknesses

Speed — 144s/question on Mac CPU is too slow for real-time use
Specific facts — Can hallucinate (e.g., said "CSOAI-001" instead of "16939677")
Limited training data — 200 samples is small; needs more sovereign corpus
CPU-only — needs GPU acceleration for production use

Path to Production Sovereign Brain

The 4-week pivot plan

Phase 2: Sovereign Brain 1B (50 GPU-hr Kaggle T4)
Phase 3: Mamba-2 sovereign attention (10 GPU-hr)
Phase 4: 4 sovereign experts — compliance, defense, intuition, voice (16 GPU-hr)
Phase 5: World model v2 at scale (5 GPU-hr, 12.7M params)
Total: ~83 GPU-hr, target release 30 Jul 2026

What This Means in Practice

The sovereign brain is NOT competing with frontier models. It's a specialist on sovereign-domain questions. Win rate on sovereign topics: 80%. Win rate on general: 100%. Honest register: smaller, slower, but more accurate where it matters.

🧠 Sovereign Brain Comparison · 12 Jul 2026 · Hermes lane
← Back to SOV33