stem-bio-ai / docs /PREPRINT_DRAFT.md
Codex
sync hf space snapshot
6a1cba7
|
Raw
History Blame
3.09 kB

Preprint Draft: STEM-BIO-AI Methodology

Title: Deterministic Evidence-Surface Auditing for Bio-Medical AI Repositories

Authors: Yun, Kwansub (flamehaven01)


Abstract

The proliferation of large language models (LLMs) in bio-informatics has led to a surge in "Bio-AI Slop" — repositories that utilize clinical marketing language without underlying technical rigor, data provenance, or safety safeguards. We present STEM-BIO-AI, a deterministic framework that audits the "evidence-surface" of a repository without relying on LLM inference. By mapping 40+ observable signals across documentation, source code (AST), and configuration, we define a tiered scoring system (T0-T4) that serves as a preliminary structural alignment signal for regulatory frameworks (EU AI Act, FDA SaMD). We demonstrate the efficacy of this approach by identifying critical failure modes in existing Bio-AI projects, including mock-data fallbacks and insecure infrastructure mounts.


5. Regulatory Alignment

Mapping STEM-BIO-AI scores to EU AI Act High-Risk AI requirements and FDA SaMD pillars. We emphasize that this tool measures the structural readiness for accountability rather than empirical clinical performance. A T4 score indicates that the repository contains the necessary infrastructure and governance artifacts required for a formal regulatory audit.

  • The Problem: The "Black Box" of AI in medicine is not just the model, but the entire repository lifecycle.
  • The Gap: Existing tools focus on code security (SAST) or model performance (benchmarks), but lack an integrated view of biological responsibility.
  • The Solution: A local-first, zero-LLM scanner that makes the audit trail 100% traceable.

2. Methodology: The 4 Stages of Evidence

  • Stage 1: Declarative Surface. Analysis of README and docs for domain vocabulary and clinical boundaries.
  • Stage 2R: Cross-Surface Consistency. Detecting "Stale" documentation or contradictions between claims and config.
  • Stage 3: Verifiable Integrity. CI/CD, domain tests, data provenance (IRB), and bias measurement.
  • Stage 4: Replication Liquidity. Containers, lockfiles, and artifact references.

3. Implementation: Deterministic Diagnostics

  • SMILES-DECEPT: Detecting chemistry slop via stack-based grammar validation.
  • MOUNT-AUDIT: Identifying insecure container configurations for clinical data.
  • RUN-TRACE: Taint-tracking of biological tool subprocesses.
  • SILENT-MOCK: Detecting library fallbacks to simulated data.

4. Evaluation and Benchmarking

  • Analysis of 50+ Bio-AI repositories from GitHub.
  • Case studies: Biomni (Subprocess/Mock risk), BioClaw (Mount risk).
  • Correlation with human expert auditor scores (Ω >= 0.95).

5. Regulatory Alignment

Mapping STEM-BIO-AI scores to EU AI Act High-Risk AI requirements and FDA SaMD pillars.


6. Conclusion

STEM-BIO-AI provides a scalable, private, and verifiable first line of defense for the institutional adoption of Bio-Medical AI.