philgear/pocketgull-skeptic-cds-4b

PocketGull Skeptical Epistemology & H0 CDS Auditor

Organization: PocketGull LLC (Oregon Registry: 258869891)
Informatics Lead: Phillip Gear (CMS NPI: 1487569752 | ORCID: 0009-0008-1372-5381)
Base Foundation Model: google/gemma-3-4b-it
Discipline: Falsifiable Clinical Decision Support, Cochrane Risk of Bias, and H0 Null Testing
Open Science Provenance: Zenodo DOI 10.5281/zenodo.20647514

---

📌 Overview

Rigorous epistemological safety auditor that challenges AI-generated clinical recommendations. Tests null hypothesis (H0) rejection validity, flags underpowered studies (p >= 0.05), calculates Fragility Index (<3 reversal events), and assesses Cochrane RoB 2 commercial sponsorship bias.

This LoRA adapter was fine-tuned using Direct Preference Optimization (DPO) on domain-specific clinical datasets conforming strictly to HIPAA §164.514 Safe Harbor de-identification standards.

---

🚀 Quickstart Inference (Transformers & PEFT)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "google/gemma-3-4b-it"
adapter_id = "pocketgull-llc/pocketgull-skeptic-cds-4b"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, adapter_id)

prompt = "Patient presents with palpitations taking St. John's Wort alongside Warfarin. Evaluate CYP450 metabolism."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.2)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

🔒 HIPAA & Regulatory Compliance

• Zero-PHI Retention: Designed for local edge computation and private Google Cloud Vertex AI deployment.
• FDA 520(o) Non-Device CDS: Supportive evidence-grounded tool intended to assist licensed healthcare providers.

📖 Citation

@software{pocketgull_clinical_2026,
  author = {Gear, Phillip},
  title = {Pocket-Gull: Living Medical Intelligence Engine & Open Clinical Science Suite},
  publisher = {Zenodo},
  version = {1.25.0},
  year = {2026},
  doi = {10.5281/zenodo.20647514},
  url = {https://pocketgull.app}
}