Willow scan cleanup: actual files, 2026-10-12

The input and clean reference are self-authored fictional English worksheets. Blue simulated answers use a handwriting-style font; red correction marks are drawn; the page-2 shadow is synthetic. This is not a real-handwriting or real-photo accuracy benchmark.

The native workspace processed two full pages in one handwriting-removal batch, then one selected residual on page 2. The whole-page output deliberately retains the actual residuals. The selection output was downloaded after accepting its preview. No other paid cleanup generation was used. Quoted maxima were 20 credits for the batch and 10 for the selection; actual account billing was not independently reconciled.

The native PNG ZIPs can be compared directly: page 1 is byte-identical; page 2 changes only inside the recorded small bounding rectangle. PDF exports contain JPEG page images and no text layer, so a PNG and PDF image need not have identical pixels. All exported PDFs have two pages.

Only page 2 was then switched to shadow removal and black/white. Their JPEGs and the final two-page PDF are included. Page 1 retains its batch result, including observed residue. Undo and redo were used on page 1 and restored that result. No OCR, Word export, phone camera or book-curve operation was tested.

The baseline PDF is a native original-color export before paid cleanup. The clean reference is authored ground truth for printed layout, not generated by the service. All eight native files preserve their downloaded bytes. This outer ZIP was assembled locally; the two inner image ZIPs are actual native downloads. See checks.json for details and limits.
