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

Key Points

• The study aims to explore a deterministic approach to AI in clinical variant analysis, emphasizing human oversight for decision-making.
• Propose an architecture that separates deterministic decision-making from interpretation by AI agents.
• Discuss the role of human oversight at the evidence admissibility stage rather than at the output stage.
• Analyze the implications of an agentic AI model in variant classification frameworks.
• The proposed architecture ensures that classifications are based on deterministic outputs, maintaining integrity in clinical applications.
• Human judgment significantly influences admissibility of evidence in variant classification, enhancing accountability.
• Findings suggest that autonomy in AI should not replace human decision-making responsibilities.