Document Rescue Atlas / public evidence
Five public documents expose five different failure modes: handwriting, dense tables, faded notebooks, structured forms and right-to-left manuscript text. Inspect the source, the recovery preview and the uncertainty before trusting any output.
Inspectable cases
These are visual recovery previews, not accuracy certificates. Each case records what makes the source difficult, what becomes reviewable and what still needs a human decision.
Case 01 / handwriting and form structure
Handwriting crosses printed guide lines and varies in spacing, stroke width and character shape. A useful recovery must preserve the form's field context instead of returning one undifferentiated block of text.
Source: NIST Special Database 19 sample image. Public benchmark source.
Case 02 / handwriting inside a dense table
The text is only half the problem. A name attached to the wrong row or column becomes a confident-looking factual error. Recovery therefore needs row identity, column meaning and visible review flags.
Case 03 / faded notebook and reading order
Low contrast, historical handwriting and irregular line flow make plain OCR brittle. The useful outcome is a navigable note with uncertain passages exposed, not silently normalized prose.
Source: Library of Congress, Alexander Graham Bell papers. See the source page for rights information.
Case 04 / printed form and field semantics
A clean-looking form can still fail when labels, checkboxes and values lose their relationships. Recovery quality depends on preserving field meaning and warning reviewers where a mark or value is ambiguous.
Source: Internal Revenue Service public Form W-4 PDF. U.S. federal government document.
Case 05 / right-to-left historical script
Historical glyph forms, connected script and right-to-left reading order create a different problem from Latin print. A responsible recovery keeps directionality and uncertainty visible for a qualified reviewer.
OCR failure library
A PDF can look readable to a person while its text layer is empty, scrambled or detached from the page structure.
Search and copy return nothing because every page is only an image.
Columns, notes and headers are extracted in the wrong sequence.
Correct words land under the wrong field or belong to the wrong record.
Names, dates and numbers contain plausible but consequential substitutions.
Faded ink disappears while paper texture becomes noise.
Printed labels survive while handwritten answers are dropped.
Right-to-left lines or multilingual fragments are reordered incorrectly.
The output looks fluent but hides unresolved passages and missing provenance.
Next decision
The free browser check identifies format, size and basic recovery signals without uploading the document. If the file is accepted, the same flow can continue to the one-time $9 recovery.