Evidence Before Retrieval: A Data-Readiness Architecture for Governed HIV Information Retrieval

Scientific retrieval systems may return relevant information without establishing whether a claim is sufficiently supported, context-complete, current, or appropriate for the intended use. We introduce Evidence Before Retrieval, a data-readiness architecture that preserves source observations, reconstructs context, and assigns evidence states before claims enter a retrievable knowledge layer. HIVGraph operationalizes this architecture across HIV guidelines, drug-interaction resources, and terminology and regulatory sources. In diagnostic analyses, agreement with expert-adjudicated labels increased from 91/200 to 192/200 relationships over three refinement rounds, and the same LLM answered 98/100 benchmark questions correctly with graph-supplied evidence versus 64/100 without retrieval. Together, these initial analyses characterize data organization, recurring failure modes, and retrieval behavior, providing a foundation for further validation of the system’s potential to improve evidence grounding in clinical decision support.