An Auditable AI Pipeline for Long-Tail Healthcare Resource Discovery and Directory Maintenance

Community healthcare directories are fragmented across heterogeneous web sources and become unreliable as organizational information changes. This study presents an auditable pipeline that combines deterministic query-space construction, language-model extraction, entity resolution, evidence verification, and human-mediated directory maintenance. The Agentic Research System uses a four-axis Query Expansion Matrix to diversify searches across service, geographic, attribute, and source dimensions. Pathways extends this discovery workflow with event-sourced provenance and a Find–Verify–Ask–Close maintenance model. We evaluate request interpretation on a 200-query benchmark containing canonical and paraphrased requests, and evaluate missing-data restoration on a sealed benchmark of 500 organizational fields. Across five extraction conditions, micro-F1 increased from 0.8916 for basic rules to 1.0000 for semantic matching combined with Gemini on the development fixture. In the restoration benchmark, 142 fields were exact, 284 were supported but non-exact, 63 were weakly supported, and 11 were unresolved after manual review. The results characterize Pathways as an assistive and auditable prototype rather than an autonomous source of truth, and identify the need for held-out language benchmarks, independent evidence validation, and live clinical evaluation.