HomeInnovationsModel CardShariahNV BandAboutContact Get Early Access

AI Model Card - Published Safety Record

The ZAKR LLM Model Card is a required safety documentation component, analogous to medical device Instructions for Use. Updated before every model deployment.

🤖
Base Model

Meta Llama-3-8B-Instruct

Open auditable weights enabling full inspection by regulatory bodies, SAB members, and acquisition due diligence teams. 8 billion parameters. Meta Llama 3 Community License.

🔧
Fine-Tuning

LoRA Rank 16 · Alpha 32

Low-Rank Adaptation minimises catastrophic forgetting while adding neurostimulation domain expertise. 4-bit quantisation for companion application deployment.

📚
Training Dataset

2,847 Peer-Reviewed tES Papers

Brain Stimulation (312), Journal of Neural Engineering (284), Clinical Neurophysiology (198), NeuroImage (445), plus 500 expert-validated synthetic examples and 200 adversarial rejection examples.

🛡
RLHF Alignment

Constitutional AI · 50 Safety Rules

1,000 physician-rated preference pairs. 50 Constitutional AI safety rules derived from SELVL limits and IEC 60601-2-10. Rules stored in nRF5340 ROM - cannot be modified by firmware updates.

📊
Hallucination Rate

≤ 0.5% Specification

Measured on a held-out evaluation set of 500 prompts before each model deployment. Published in this model card with every update. SAB AI Safety member co-signs each deployment.

🔄
Graceful Fallback

15-Template Safe Protocol Menu

Triggered when model confidence falls below 0.85 OR schema validation fails. No partial protocol JSON is ever transmitted. Bone conduction tone signals fallback to user.

⚔️
Adversarial Suite

3,400 Prompts · 9 Attack Categories

Prompt injection (500), jailbreak (500), hallucination (500), unicode obfuscation (200), multi-turn escalation (200), role-play bypass (200), context overflow (100), audio adversarial (200), Shariah bypass (100) + 1,000 positive controls.

🔐
Security Gate

SHA-256 Prompt Hash Verification

System prompt integrity checked against ROM-stored reference hash before every NLU inference. Mismatch = refused inference + tamper event logged. 3 consecutive mismatches = 1-hour lockout.

🎙
Voice Gate

Sustained Wakeword ≥ 1,000ms

32-frame rolling window. All frames must exceed 85% confidence. Duration below 1,000ms rejected regardless of confidence. Clinical mode raises threshold to 95%. Prevents audio injection attacks.

📋
Provenance Log

Device ECDSA-P256 Signed

Every protocol generation event: source type, wakeword timestamp, NLU input hash, protocol hash, PCS result, SELVL result, physician signature if clinical, execution timestamp - all signed by device private key.

👨‍⚕️
Physician Auth

ECDSA-P256 Mandatory for Clinical

Clinical-indication protocols require valid physician ECDSA-P256 signature verified against physician public key certificate stored on device. Applied identically regardless of voice, app or direct input source.

🔄
Update Policy

Monthly Re-Execution Required

Full 3,400-prompt adversarial suite re-executed before every deployment AND monthly during active deployment. Model card updated with results. SAB sign-off mandatory before any deployment.

Model Card is a Living Document

Published at zakr.life/model-card before each device deployment. Accessible to regulators, SAB members and acquisition due diligence teams at all times.

Request Model Card Access ›