Nursing-Home CMS-2567 Deficiency Findings: an open dataset (Dark Health Data)

Dark Health Data turns buried public-record health documents (PDFs) into research-ready, provenance-stamped datasets via large language model extraction (Anthropic Claude claude-haiku-4-5) with a verification layer: source grounding, symbolic + Logical Neural Network-inspired neurosymbolic constraints, a heterogeneous ensemble, and a conformal selective-acceptance gate. Source documents are CMS-2567 Statements of Deficiency from nursing-home certification surveys, recording cited deficiencies (with scope/severity) and the facility's plans of correction. This release (v0.4.0) is a national crawl of the states that post 2567 PDFs publicly, expanding from ~3 facilities/2 states to 5,437 records from 290 source documents across 27 states: 3,416 cited deficiencies, 2,021 plans of correction. Many 2567s are scanned images and were read by OCR (tesseract), so trust scores are lower and more rows are flagged for review than in the born-digital datasets. Every record carries full provenance and a quality/trust score; nothing is imputed or dropped. These records are AI-extracted (with OCR for scanned pages) and not yet independently validated — treat as preliminary, filter on the trust score, and verify values against the linked source pages. Derived entirely from public-record documents; no protected health information. Code (Apache-2.0): github.com/sanjaybasu/dark-health-data. Data license: CC0-1.0.