Queryable Evidence from Scientific Papers at Scale: A Statistical Graph and API for Health Research

Quantitative evidence in the health literature is locked within discrete papers: researchers extract effect sizes and group comparisons one at a time. We demonstrate AnonGraph a deployed, statistical graph whose atomic unit is a statistical finding --- an effect size or group comparison with its confidence interval or $p$-value --- extracted from scientific literature via PubMed, Europe PMC, and OpenAlex by a GenAI pipeline, grounded to Wikidata, and served through a public REST API and versioned dataset deposits. It is publicly accessible for research and non-commercial use. Generative AI is used upstream to robustly compile unstructured scientific text into structured, provenance-linked evidence, rather than serving as the end-user conversational interface. We present the deployed system, three worked demonstrations on diabetic retinopathy (evidence landscaping, literature synthesis, and hypothesis generation), and an account of limitations observed in closed-beta deployment, including a quantified extraction-accuracy issue and its mitigation.