Evidence-Binding Failure in AI-Assisted Drug Discovery Pipelines: The Admissibility Gap Between Plausibility and Decision-Grade Support

Abstract

Preprint. This paper characterizes an “Admissibility Gap” failure regime in AI-assisted drug discovery and development diligence workflows, where acceptance-weighted evaluation and verification scarcity lead to audit-shaped but non-admissible outputs. It introduces “Evidence-Binding Failure” as the core mechanism: decision-relevant claims that are not bound to verifiable, resolvable evidence objects under a declared access policy. The paper discusses mitigations including evidence objects, time-gating, fail-loud handling, and auditable run records, and relates these to a companion reliability contract framework (VECTR).

Ähnliche Arbeiten