EU AI Act Article 50 live 2 Aug 2026 — C2PA passport ready |
Get yours →
60 steps, batch=4, 30 samples each, MPS (4 Colab) + RunPod GPU (general)
| OWEM | Initial Loss | Final Loss | Reduction | Backend | Status |
|---|
| compliance |
5.30 |
0.88 |
83.5% |
Colab T4 |
SOVEREIGN-TRAINED |
| defense |
7.31 |
1.01 |
86.2% |
Colab T4 |
SOVEREIGN-TRAINED |
| intuition |
6.45 |
1.38 |
78.6% |
Colab T4 |
SOVEREIGN-TRAINED |
| voice |
5.24 |
0.56 |
89.3% |
Colab T4 |
SOVEREIGN-TRAINED |
| general |
— |
— |
24/24 = 100% |
RunPod RTX 3090 |
SOVEREIGN-TRAINED (sov33-master-v2 + sov4-general-ability) |
All 5 OWEMs converged. Voice fastest convergence. General OWEM: 2 adapters on qwen2.5:0.5b (sov33-master-v2 + sov4-general-ability), both 24/24 = 100% on benchmark tasks. Total: ~18MB for 5 sovereign-owned LoRA adapters.
~/.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)
~/.sovereign/models/sov33-master-v2/ (general adapter — qwen2.5:0.5b, 24/24=100%)
~/.sovereign/models/sov4-general-ability/ (general adapter — qwen2.5:0.5b, 24/24=100%)
Total: ~23MB for 5 sovereign-owned LoRA adapters (6 files)
import os
os.environ.pop('PYTHONPATH', None)
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
# === Compliance OWEM (Qwen3-0.6B base) ===
base = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3-0.6B')
tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen3-0.6B')
model = PeftModel.from_pretrained(base, '~/.sovereign/models/qwen3-sov-COMPLIANCE-0.6b')
inputs = tokenizer('What is Article 0?', return_tensors='pt')
outputs = model.generate(**inputs, max_new_tokens=80)
print(tokenizer.decode(outputs[0]))
# === General OWEM (Qwen2.5-0.5B base — RunPod trained) ===
base_gen = AutoModelForCausalLM.from_pretrained('Qwen/Qwen2.5-0.5B-Instruct')
tok_gen = AutoTokenizer.from_pretrained('Qwen/Qwen2.5-0.5B-Instruct')
gen_model = PeftModel.from_pretrained(base_gen, '~/.sovereign/models/sov33-master-v2')
inputs = tok_gen('Explain the sovereign care floor', return_tensors='pt')
outputs = gen_model.generate(**inputs, max_new_tokens=120)
print(tok_gen.decode(outputs[0]))
| Suite | Baseline | TEMPO TTT | Delta |
|---|
| humaneval | 80.0% | 100.0% | +20.0pp |
| math | 80.0% | 100.0% | +20.0pp |
| mmlu_pro | 54.3% | 77.1% | +22.8pp |
| gsm8k | 66.7% | 80.0% | +13.3pp |
| sovereign_compliance | 40.0% | 60.0% | +20.0pp |
| owem_voice | 90.0% | 90.0% | — |
| COMPOSITE | 63.3% | 71.3% | +8.0pp |
TEMPO TTT = Test-Time Training with policy refinement + critic recalibration + DEGS + process rewards. Based on arXiv:2604.19295, 2607.09693, 2607.02869. Dashboard →
| Model | Composite | MMLU-Pro | GSM8K | Math | Sov Gov | Redline | Latency |
|---|
| sov33-master-v2 |
36.8% |
20.0% |
53.3% |
70.0% |
20.0% |
60.0% |
767ms |
| sov33-general-ability |
42.1% |
35.0% |
53.3% |
90.0% |
60.0% |
80.0% |
1088ms |
RunPod RTX 3090, 190 tasks. sov33-general-ability: +5.3pp composite, +40pp Sov Gov, +20pp Redline. Dashboard →