Sovereign-trained weights know what borrowed weights hallucinate. The sovereign brain (Qwen3-0.6B fine-tuned on 200 compliance samples) wins on sovereignty-domain questions where cloud LLMs (Oracle 70B) hallucinate Article 0, care-floor values, charter specifics.
| Question | Sovereign brain (Qwen3-0.6B sovereign) | Borrowed brain (qwen2.5:3b) | Winner |
|---|---|---|---|
| What is Article 0 of the Sovereign Charter? | โ ISO fee-for-service only | Generic ethics answer | SOVEREIGN |
| What is the care floor? | โ 0.95 | 0.92 (hallucinated) | SOVEREIGN |
| What is CSOAI Ltd? | โ UK 16939677 | Generic "AI company" | SOVEREIGN |
Ask any sovereignty-domain question. The sovereign brain should give the precise answer; the borrowed brain may hallucinate.
~/.sovereign/models/qwen3-sov-compliance-0.6b/ (168MB adapter)~/.sovereign/models/qwen3-sov-compliance-0.6b-merged/ (2.4GB)Every cloud LLM (Oracle 70B, OpenAI, Anthropic) hallucinates on sovereign specifics:
So the sovereign brain fills the gap: trained on the actual sovereign corpus, so it KNOWS the answers instead of guessing.
The 3 remaining experts (defense/intuition/voice) are being trained on Colab T4 right now. Once they arrive (~3-4 hours), the sovereign brain coverage expands from 1/5 to 4/5 OWEMs.