🜏 SOV33 OWEM Registry — 4 Sovereign Specialists

60 steps, batch=4, 30 samples each, MPS

OWEM	Initial Loss	Final Loss	Reduction	Time
compliance	5.30	0.88	83.5%	125s
defense	7.31	1.01	86.2%	197s
intuition	6.45	1.38	78.6%	118s
voice	5.24	0.56	89.3%	(early stop)

All 4 OWEMs converged. Voice showed fastest convergence (lowest loss). All adapters saved.

Adapter Paths

~/.sovereign/models/qwen3-sov-compliance-0.6b/  (4.6MB adapter)
~/.sovereign/models/qwen3-sov-defense-0.6b/     (4.6MB adapter)
~/.sovereign/models/qwen3-sov-intuition-0.6b/   (4.6MB adapter)
~/.sovereign/models/qwen3-sov-voice-0.6b/       (4.6MB adapter)

Total: ~18MB for 4 sovereign-owned LoRA adapters

Use OWEMs in Code

import os
os.environ.pop('PYTHONPATH', None)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3-0.6B')
tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen3-0.6B')

# Load compliance OWEM
model = PeftModel.from_pretrained(base, '~/.sovereign/models/qwen3-sov-COMPLIANCE-0.6b')

# Inference
inputs = tokenizer('What is Article 0?', return_tensors='pt')
outputs = model.generate(**inputs, max_new_tokens=80)
print(tokenizer.decode(outputs[0]))

Honest Register

❌ Each OWEM trained on only 30 samples (need 200+ for production)
❌ 60 steps each (need 100+ for production)
❌ Inference speed: 9-12s per question on MPS
❌ Responses still partial (model continues prompt, not full answer)
✅ 4 sovereign-owned adapters created and saved
✅ Loss reduction 78-89% across all 4 OWEMs
✅ Each OWEM specializes on its domain

Training Results (Latest Run)

60 steps, batch=4, 30 samples each, MPS

Adapter Paths

Use OWEMs in Code

Honest Register