A Deterministic Classification Core with an Agentic Interpretation Layer: An Architecture for Auditable, Human-Gated Clinical Variant Analysis

2026/07/04 by Vladimir Mitev · 1 voice
Biochemistry, Genetics and Molecular Biology · Computer Science · #Genomics and Rare Diseases #Machine Learning in Healthcare #Multimodal Machine Learning Applications

paper · doi:10.5281/zenodo.21189571

openalex publication_date 2026/07/04 · openalex created_date 2026/07/05 · openalex updated_date 2026/07/05

Abstract

Agentic AI in care delivery is increasingly framed as autonomy with minimal oversight, and recent work places language models inside ACMG/AMP variant classification itself. We argue the opposite for auditable deployment: keep the decision core deterministic and free of model output, confining agents to an interpretation layer acting on already-classified results. Our central, falsifiable claim is that the load-bearing human gate sits upstream, at evidence admissibility, not at output release.

Discussions

• Everyone's racing to put LLMs inside clinical variant classification. We went the other way. Our new preprint: keep the decision core deterministic and AI-free, and confine agents to interpretation. [bsky, 3 points, 0 comments]

Related

• Agentic AI framework for End-to-End Medical Data Inference
• Artificial Intelligence-Driven Clinical Decision Support Systems
• Agentic Loafing: An AI Decision Delegation Risk
• Auditing for Human Expertise
• MedMMV: A Controllable Multimodal Multi-Agent Framework for Reliable and Verifiable Clinical Reasoning