Evidence Admissibility Is the Load-Bearing Human Gate in Auditable Clinical Variant Classification
Vladimir Mitev
- Published
- 14 July 2026
- Record type
- Preprint
- Version
- Version 2.0
- Article DOI
- 10.5281/zenodo.21346784
Abstract
Abstract
Clinical AI oversight is often concentrated at final output. In variant classification, an autonomous component can affect the decision earlier by changing the reference data and evidence snapshot consumed by the classifier.
This Perspective places the load-bearing human transition at evidence admissibility (G1). An authorized person promotes a provenanced, version-locked evidence snapshot before it enters a deterministic, model-output-free ACMG/AMP classification core. Qualified sign-out (G4) remains necessary when the result is released.
The paper proposes six evaluation metrics covering unauthorized evidence admission, attribution of classification changes, model-to-core boundary violations, exact reconstruction, narrative contradiction and prioritization omission. These are proposed tests, not reported validation results.
Evidence boundary
Evidence boundary
The framework addresses auditability and governance. It does not establish classification correctness, clinical validation, comparative effectiveness or fitness for purpose. Complete separation also increases the workload of human literature curation.
01
Where the human gate sits
G1 governs which evidence snapshot may enter the classifier. The classification core then produces a frozen class and criterion trace without model output. Language-model agents remain downstream, where they can support interpretation without changing the computed class.
G4 governs qualified release. The two gates protect different transitions: evidence admission before computation and sign-out after computation.
02
How the proposal can be tested
The proposed metrics ask whether evidence changes were authorized, whether classification drift can be attributed to a recorded change, whether model-generated material crossed into the class-determining path and whether a result can be reconstructed from its manifest.
The revised record reports no new dataset or completed validation study. It defines the architecture, the gates and the measurements needed to test them.