GRITWORKS OCR & EXTRACTION - REVIEW COPY

One supplier / buyer workflow: purchase order, goods and services invoices, return credit and account statement. $3,558.80 invoiced - $103.68 credit = $3,455.12 due.

10 input PDFs: five clean + five negative. Annotated PDFs and the remaining variants are in the separate complete archive.

FORMATS
Each input PDF has .json expected answers, .boxes.json field boxes in PDF points
(top-left origin), and _bbox.json normalized vertices and 96-DPI pixel boxes.
The full archive includes clean annotated PDFs; click a rectangle for key/value.
Boxes cover selected populated fields, not every label or static word.
Clean/variant bodies are rasterized. Negative PDFs contain no hidden text layer.
Expected values do not change when a field is obscured or clipped.
Geometric captures carry forward homographies, projected polygons and clipped
rectangles. Visibility means clipping, not an assertion of OCR legibility.
No OCR confidence is asserted. Word boxes are supplied for the Indian-language
source forms; other authored/variant files provide field boxes only.
Keep JSON and annotated PDFs out of extraction-model inputs.

17 STANDARD CAPTURES
blur, jpeg_artifacts, black_white_threshold, scan_border, scan_lines, bleedthrough, scanner_bw_dithered, old_fax_noise_threshold, old_scanner_rollers, grayscale, photocopy, low_ink, bad_photo, fax_low_contrast, coffee_stain, scanner_color, fax_blur_noise

PROVENANCE
Industrial PO from Docs for website.zip; goods/services invoice and credit memo layouts from the supplied approved OCR download; account ledger adapted to this connected case. Layouts reconstructed with new synthetic business data.

All files are review samples. Source files and the website remain unchanged.
