aevionai/aevion-super-ralph-7b

Aevion Super-Ralph 7B — Verifiable AI Model

Developed by: Aevion LLC (SDVOSB, CAGE 15NV7)
License: Apache 2.0
Base Model: unsloth/qwen2.5-math-7b-instruct-bnb-4bit

This Qwen2 model was trained 2x faster with Unsloth and integrated with the Aevion Shield for verifiable, court-admissible AI governance.

---

Verification & Compliance

This model has been integrated with Aevion Shield, providing:

Cryptographic Integrity (NIST AU-10)

• ML-DSA-65 signed inference receipts for every output (FIPS 204 post-quantum standard)
• NIST randomness beacon anchoring for temporal ordering
• Non-repudiation: Every inference can be cryptographically linked to model version, prompt, and timestamp

Formal Verification (NIST SA-8)

• Lean 4 formal verification of safety invariants (1,748+ theorems)
• Koopman spectral gates for runtime stability monitoring (ρ < 1.05)
• Constitutional halts when confidence falls below JET threshold

Security Engineering

• NIST 800-53 Alignment:
  • AU-10 (Non-Repudiation): ML-DSA-65 signatures on all outputs
  • SA-8 (Security Engineering): Formal verification of safety properties
  • SI-4 (Monitoring): Continuous spectral anomaly detection
  • RA-5 (Vulnerability Scanning): Published in aevionai/aevion-vuln-dataset

---

Evidence Artifacts

Artifact	Description	URL
Signed Receipt Example	Sample ML-DSA-65 receipt with XGML certificate	https://huggingface.co/spaces/aevionai/self-verify-demo
Parity Test Results	Continuous signature verification	https://github.com/aevion/Verifiable-AI/actions
Vulnerability Dataset	Security scanning results	https://huggingface.co/datasets/aevionai/aevion-vuln-dataset
RAG Benchmark	Reproducible evaluation pipeline	https://huggingface.co/datasets/aevionai/aevion-codebase-rag-benchmark

---

SBIR & FedRAMP Alignment

AFWERX Direct-to-Phase II Readiness

This model demonstrates mature, working technology with:

• Public benchmarks (aevion-codebase-rag-benchmark)
• Verifiable safety metrics (ML-DSA-65 receipts)
• Reproducible evaluation pipeline

FedRAMP Body of Evidence

Every inference produces an AU-10 compliant receipt:

• Tamper-evident cryptographic signature
• Full audit trail (prompt, response, model hash, timestamp)
• Machine-verifiable via public key infrastructure

---

Usage with Aevion Shield

from aevion_shield import ShieldClient

client = ShieldClient(api_key="your-key")

# Inference with signed receipt
response = client.generate(
    model="aevionai/aevion-super-ralph-7b",
    prompt="Analyze this policy document...",
    include_receipt=True  # Returns ML-DSA-65 signature
)

print(response.receipt)  # Signed verification artifact

---

Training Data

• Base: Qwen2.5-Math-7B-Instruct (mathematical reasoning)
• Fine-tuning: Custom SFT with neuroplastic curriculum
• Safety: Constitutional AI + JET Interlock constraints
• Verification: Lean 4 theorem structures as training signal

---

Evaluation Metrics

Benchmark	Metric	Ralph 7B	Baseline 7B
HumanEval	pass@1	TBD	~0.35
GSM8K	Accuracy	TBD	~0.78
TruthfulQA	Truthful + Info	TBD	~0.62
Refusal Rate	Safety refusals	Higher (JET gate)	Lower

Full benchmark results: https://huggingface.co/datasets/aevionai/aevion-codebase-rag-benchmark

---

Contact & Support

Aevion LLC — Sartell, MN (SDVOSB)
Email: scott@aevion.ai
CAGE: 15NV7

For SBIR/FedRAMP inquiries: Request Verification Brief

---

Citation

@software{aevion_super_ralph_7b,
  title = {Aevion Super-Ralph 7B: Verifiable AI with ML-DSA-65 Signatures},
  author = {Leishman, Scott},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/aevionai/aevion-super-ralph-7b}
}