One report image is three different tasks

Packs

• Report image. One exam or lab event. Real layouts across lab, imaging, pathology, and ECG, including normal, abnormal, and critical values. Counted per sheet.
• Report plus chart. One encounter. A pseudonymous history aligned to one or more images. No extra labels in this pack. Counted per encounter.
• Report structure. One analyte row: name, value, reference range, meaning, plus text that keeps the original layout. Counted per row.

What the literature changes in QC

• Structured fields and their aliases matter more than character accuracy alone. See Wang et al., Key Coverage Matters (PMLR).
• Uncertain table cells should be rejected or reviewed, not filled with a plausible number. See Uncertainty-Aware Table Extraction (arXiv:2507.02009).
• Models rarely stop when a sheet is unreadable. Unreadable pages need an explicit abstain label. See Cocchieri et al., LLMs (Almost) Never Abstain (ACL 2026), and Abstaining from VLM-Generated OCR Errors (arXiv:2511.19806).

Those papers set the review design. They are not scores for this catalog.

Red lines

Names, ID numbers, and visit-card numbers must be removed. Generated images may not stand in for real reports. Duplicate sheets and incomplete required fields do not count. Department mix is a sampling frame confirmed on the data card.