Sheaf
Scanned and photographed documents
Real document work arrives as a scan of a print of a fax, often photographed on somebody's desk.
The ordinary case
Most document tooling treats a clean digital PDF as the normal input and everything else as degraded. That is backwards for operational work. The clean PDF is the lucky case. The normal case is a scan with a slight rotation, a staple shadow down one edge, and a signature page photographed in portrait on a phone.
Sheaf is built for the normal case. A page with no text layer is not routed to a person, flagged as low quality, or read with a reduced set of capabilities. It is read.
A text layer says what a document is, not what it says
This is the least obvious finding in the whole pipeline, and it is measured rather than assumed. When a form is filled in, the printed chrome — the field labels, the form number, the footer — often lives in the text layer while the filled values do not. The text layer is therefore excellent evidence of which form this is and unreliable evidence of what was entered on it.
Treating those two questions as one is how systems end up confidently reading a blank form and reporting the labels as values.
Use the text layer to identify. Read the page to extract. Where the two disagree about whether a field was filled, the page wins.
Reading a page with nothing to read
Where there is no usable text layer, the page is rendered and read optically, and the resulting geometry becomes the same page geometry every citation is bound to. From that point on a scanned page behaves exactly like a digital one: values carry a file, a page and a span, and can be highlighted rather than described.
A tick is content
Checkboxes are where transcription and meaning part company. A ticked box and an empty box are a few pixels apart and mean opposite things, and plenty of pipelines drop the distinction entirely because it is not text. Sheaf reads checkbox state as content, because on a great many operational forms the ticks are the answer and the prose is packaging.
What still gives it trouble
Handwriting in free-text fields remains the hardest input in document work, and confident handwriting recognition is where we would rather abstain than guess. Severe skew, heavy shadow across a value, and pages photographed at an angle that hides part of the sheet will also produce an abstention rather than an answer. In each case the outcome is a flagged gap you can see, not a number you cannot check.