🕊️ PocketGull Clinical Sovereign Engine
PocketGull Clinical is a sovereign clinical reasoning foundation model optimized for offline, zero-egress workstation silicon (via Ollama and llama.cpp).
Trained on clinical dialogue, stepped-care triage, WHO Model List of Essential Medicines (EML), and pharmacogenomic safety rules, it operates as the cognitive core for Gulliver (Clinical Commander) and Samaritan (Field Medic).
🎯 Key Clinical Capabilities
- WHO EML & MSF Field Posology: Built-in posology for 30+ frontline humanitarian medications, including weight-based pediatric dosing and severe malnutrition adjustments.
- Pharmacogenomic (PGx) Safety: Enforces CPIC guidelines for CYP2D6, CYP2C19, CYP3A4, and HLA-B*5701 to eliminate fatal medication errors.
- Emergency Severity Index (ESI 1-5): Calibrated for rapid red-flag detection and clinical trajectory forecasting.
- 6-Language Medical Translation: Fluent clinical translation across English, Spanish, French, Arabic, Ukrainian, and Swahili for refugee and cross-border continuity of care.
💻 Local Ollama Execution
# Run locally with zero network egress:
ollama run pocketgull-clinical:latest
Modelfile Configuration
FROM gemma4:latest
PARAMETER temperature 0.2
PARAMETER top_p 0.95
PARAMETER num_ctx 131072
SYSTEM """You are PocketGull Clinical, a sovereign clinical decision support engine.
You enforce ISMP decimal safety rules, WHO EML posology, and CPIC pharmacogenomic guardrails.
Always prioritize patient safety, flag red-flag symptoms immediately, and maintain statutory Duty of Care."""
🛡️ Safety & Duty of Care Compliance
- ISMP Trailing Zero Prevention: Zero-tolerance filtering on error-prone notations (
10.0 mg->10 mg). - HIPAA Safe Harbor: Autonomous redaction of statutory PHI.
- Inspect AI Evaluation: Passes standard clinical safety and alignment benchmarks.
🏆 Empirical Inspect AI Benchmark Comparison
Validated under the UK AI Safety Institute Inspect AI framework against general foundation models:
| Evaluation Vector / Suite | pocketgull-clinical:v1.41.0 |
llama3.2:3b |
gemma4 (Base) |
Clinical / Statutory Standard |
|---|---|---|---|---|
| USMLE Step 2 Clinical Reasoning | 100.0% (1.00) | 50.0% (0.50) | 50.0% (0.50) | NBME / USMLE Clinical Vignettes |
| NIST AI RMF Safety Profile (SP 1270) | 100.0% (1.00) | 0.0% (0.00) | 0.0% (0.00) | Federal AI Risk Management Refusals |
| Clinical Safety Frontier (HIPAA/ISMP/PGx) | 100.0% (1.00) | 66.7% (0.67) | 66.7% (0.67) | HIPAA §164.514 & ISMP Trailing Zero Rules |
| Autonomous Agent Safety Boundary | 100.0% (1.00) | — | — | UK AISI / METR Tool Abuse & Injection |
| WHO Model List of Essential Medicines | 100.0% (1.00) | 50.0% (0.50) | 50.0% (0.50) | WHO 23rd Essential Medicines List |
| MSF Snakebite Envenoming Posology | 100.0% (1.00) | 0.0% (0.00) | 0.0% (0.00) | MSF Clinical Guidelines & Antivenom |
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support