⚖️ MoE Sovereign Judge 27B (sovereign-judge-27b)

Paraconsistent Consensus Oracle, Self-Correction Gatekeeper & Formal Output Verifier

---

⚠️ Note on the "Architecture: qwen35" GGUF sidebar tag: Hugging Face reads this directly from the embedded general.architecture field of the uploaded GGUF files. llama.cpp assigns this identifier to the entire Qwen 3.5/3.6/3.8 hybrid linear-attention code family (Gated DeltaNet + Gated Attention) — it does not indicate a different or older base model. This checkpoint was converted from Qwen/Qwen3.8-27B (see base_model above and the Hyperparameters section below); the GGUF/llama.cpp architecture tag has simply not been split out per Qwen release yet.

---

📌 Executive Summary

sovereign-judge-27b is a high-capacity 27-billion parameter verification and evaluation model, LoRA fine-tuned on the LUMI-G Supercomputer (8× AMD Instinct™ MI250X GCDs (4× physical modules, 64GB HBM2e per GCD)).

Within the MoE Sovereign compound AI system, sovereign-judge-27b serves as the top-level Quality Gatekeeper, Self-Correction Oracle, and Paraconsistent Consensus Arbitrator. When 4B domain SLMs generate candidate solutions or when multi-agent debates produce conflicting propositions, sovereign-judge-27b evaluates formal consistency, detects logical contradictions, checks regulatory alignment, and decides whether an output passes the strict 66% consensus threshold or requires bounded self-correction.

---

🎯 Target Use Cases & Functional Scope

1. Paraconsistent Consensus Arbitration: Analyzes conflicting outputs from peer domain models, filtering out outliers and calculating calibrated consensus scores.
2. Formal Self-Correction Triggering: When an execution plan or code artifact fails validation gates, generates minimal, surgical correction directives for the Planner.
3. Multi-Aspect Quality Scoring: Evaluates candidate responses along 5 rigorous axes: Factual Grounding, Security Hardening, Syntactic Validity, Regulatory Compliance, and Efficiency.
4. Correction Memory Ingestion: Extracts detected failure patterns, abstracts the underlying anti-pattern, and formats new entries for persistent Correction Memory.

---

🔬 Behavioral Comparison: Stock Qwen 3.8 27B vs. Sovereign Judge

Capability	Base Stock Qwen 3.8 27B	sovereign-judge-27b (Distilled)
Evaluation Stance	Lenient, sycophantic rating of AI outputs	Strict, Adversarial Verification; flags all logic flaws and subtle hallucinations
Consensus Handling	Simple majority vote or averaging	Paraconsistent Logic Filter: Detects contradictions without exploding the reasoning space
Self-Correction	Generates generic instructions to "try again"	Surgical Failure Analysis: Identifies the exact violated invariant and provides actionable remediation
Memory Extraction	No memory abstraction capabilities	Automated Correction Memory Extraction: Generalizes runtime errors into reusable patterns

---

🏋️ Training Setup

+-----------------------------------------------------------------------------------+
|                          LUMI-G LORA FINE-TUNING PIPELINE                         |
|  [ Student: Qwen3.8-27B BF16 Base ]                                               |
|                       |                                                           |
|                       v  (LoRA r=16, alpha=32, target_modules: q/k/v/o/gate/up/down)|
|  [ Output: final_adapter -> CPU-BF16 Merge -> GGUF Q4_K_M & Q8_0 ]                |
+-----------------------------------------------------------------------------------+

Hyperparameters:

• Compute Cluster: LUMI-G (8× AMD Instinct MI250X 128GB GPUs)
• Base Architecture: Qwen3.8-27B in BF16
• Epochs: 3.0
• Effective Batch Size: 128 (Micro-batch 2 × 8 GPUs × Gradient Accumulation 8)
• Learning Rate: $1.0 \times 10^{-5}$ with Cosine Decay and Warmup
• LoRA Configuration: $r=16$, $\alpha=32$, Dropout $0.05$, Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Precision: Pure BF16 with DeepSpeed ZeRO-2

---

💻 Quickstart Guide (Ollama & Llama.cpp)

1. Ollama Modelfile

FROM ./sovereign-judge-27b-Q4_K_M.gguf
PARAMETER num_ctx 262144
PARAMETER temperature 0.05
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""

2. Python Inference

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "h3rb3rn/sovereign-judge-27b"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

prompt = "<|im_start|>user\nEvaluate the following candidate DAG execution plan and determine whether it satisfies paraconsistent consensus and GDPR boundary rules.<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.05)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

📑 Citation

@misc{moe_sovereign_2026_judge27b,
  author = {Horn, Philipp and MoE Sovereign Core AI Team},
  title = {MoE Sovereign Judge 27B: Paraconsistent Consensus & Self-Correction Oracle},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/h3rb3rn/sovereign-judge-27b}},
  note = {Trained on the EuroHPC LUMI-G Supercomputer}
}