Medical benchmark.

Rasterized docs ship as page-image PDFs (no text layer). They test the OCR / layout path, not text-layer extraction — treated as a sibling-class to the scanned-doc set, not equivalent to born-digital business PDFs.

• Focused customer benchmark for preventive-screening chart uploads, not the public all-provider benchmark.
• 30 public synthetic, de-identified, or sample PDFs across labs, mammography, imaging, colorectal screening, scanned forms, and fax-like documents.
• Only extract hosted, Reducto, and Gemini Structured Output were evaluated because those were the providers relevant to this customer comparison.
• Gemini used gemini-2.5-flash via Vertex AI with structured JSON output and page-aware [y_min, x_min, y_max, x_max] bounding boxes normalized to 0..1000.
• Each provider ran once per document (repeats=1). Latency is wall-clock from the benchmark runner.
• Consensus F1 is provider agreement against a pseudo-reference across the three successful providers, NOT human-labeled accuracy.
• BBox metrics are source-grounding agreement in PDF point space, not human-labeled layout accuracy.