Source-Constrained Claim Verification and Risk-Adaptive Retrieval for Evidence-Grounded Healthcare Search: Development and Benchmark Study (Preprint)

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

BACKGROUND Access to reliable healthcare information depends on finding relevant sources and accurately representing what they establish. Agentic search systems can retrieve information from provider registries, biomedical literature and interoperable health records, but a valid citation does not necessarily support the claim attached to it. Healthcare information seekers increasingly use conversational AI to find providers and understand biomedical evidence. Retrieval-augmented systems can connect generated answers to external sources, yet the presence of a citation does not establish that it supports the statement beside it. This distinction matters when a system combines provider registries, research articles and health data records, because each source supports different kinds of claims. OBJECTIVE Evaluating claim-level evidence verification and evidence-sufficiency decisions separately and together, while measuring unsupported claims, answer completeness and appropriate abstention. METHODS The Source-Constrained Claim Graph (SCCG) and Risk-Adaptive Retrieval and Abstention Controller (RARC) is implemented and tested and their effectiveness is reported. RESULTS The primary outcome depends on rate of unsupported factual claims; secondary outcomes includes claim-to-source accuracy, appropriate abstention, answer completeness and expert-rated usefulness. The existing platform passed 11 of 11 acceptance cases with Gemini and 10 of 11 with Gemma; the remaining Gemma case ended in an interoperability workflow timeout. Both paths passed all applicable deterministic citation, retrieval and unsupported-request checks. Median response times were 10.60 and 167.41 seconds, respectively, under different runtime conditions. CONCLUSIONS This study examined an open-source healthcare search platform using 11 acceptance cases covering provider discovery, biomedical research, health information, interoperability and unsupported requests. The Gemini path passed all 11 cases, while the Gemma path passed 10; its remaining interoperability case ended in a runtime timeout. Both paths completed the applicable citation, retrieval and rejection checks. These findings show that the existing platform followed its expected workflows in this small benchmark, but they do not establish whether every answer claim was supported by its cited evidence. These results describe the existing platform; the proposed evidence-sufficiency and claim-verification components have not yet been evaluated.