A 104-page packet.
Every edge found.

Boundary Engine / 104 pp. → 33 documents

01 / INPUT

Bring the mess.

Do not prep. Do not sort. Do not clean up. Just point Sheaf at the raw intake and let the engine handle the reality of operational documents.

Scans

Phone Photos

Merged Packets

Application Dumps

Wrong-order Pages

Junk Pages

Dense Tables

Signatures

02 / PROCESS

Finding the edges.

One file arrives holding thirty documents and no table of contents. Sheaf decides where each one starts and ends, and covers every page while doing it.

Read

Every page, first.

The whole packet goes through text and layout analysis before anything is cut, so scans and photographs count the same as clean digital pages.

Signal

Find the seams.

Looks for the marks a new document leaves behind: a fresh header, a restarted footer, a "page 1 of 6" counter, a change of form entirely.

Cut

Commit the boundary.

Places every cut so the runs tile the packet end to end. No page is left sitting between two documents, and none is claimed by both.

Name

Label the run.

Each separated document is typed from its own content, so a run named scan_0042.pdf still comes out as a title commitment.

Technical detail How splitting works, and what it costs to lose a page Read the spec →

03 / OUTPUT

Structured payload.

The file doesn't need to arrive perfect, but the data leaves that way. Ready for your existing infrastructure.

One packet in, a list of real documents out — each with the page range it occupies and the type it was recognized as.

The ranges tile the original file end to end, so nothing you were sent goes missing on the way through.

{
  "packet_id": "pkt_8849-2A",
  "pages_in": 104,
  "documents": [
    { "type": "loan_application", "pages": "1-4" },
    { "type": "pay_stub", "pages": "5-6" }
  ],
  "documents_out": 32,
  "pages_orphaned": 0
}

How splitting works

A hundred-page file arrives holding thirty documents and no table of contents. Splitting decides where each one starts and ends — and, just as importantly, refuses to lose a page while doing it.

The ranges tile the file end to end.
A page cannot fall between two documents.

Complete coverage is the constraint, not the ambition.

The page ranges have to tile the file end to end. Not mostly cover it — tile it. Every page belongs to exactly one document, so there is no quiet gap for a page to disappear into and no overlap that files the same page twice under two different names. A split that leaves three pages unaccounted for has not done ninety-seven percent of the job; it has produced a file you cannot trust the count of.

An incomplete answer is completed, not rejected.

If the pass comes back with a gap in it, that is not a broken answer. It is an answer with a gap. Throwing the whole thing away and starting again spends an entire file to recover one range, and tends to fail the same way twice. The gaps are closed deterministically instead, so a partial result still turns into a usable split rather than an error message.

Rules bought almost nothing, and that is measured.

An earlier version of this stage had a signal layer in front of it: hand-written rules for fresh headers, restarted footers, form numbers, the usual tells. Across a benchmark of fourteen files, roughly forty lines of those rules improved exactly one of them.

Every remaining miss had the same shape — a document whose own footer, Page 3 of 6, was the only evidence it hadn’t finished yet. That is not a rule you can write in advance. It is a thing you have to read.

Over-splitting and over-merging fail differently.

Over-split

One document arrives as several fragments. Irritating, visible, and quick to correct — the pieces are all in front of you.

Over-merge

Two documents arrive as one. The second is now invisible: nothing is missing from any count, so nothing looks wrong, and it stays wrong until someone opens the pages.

They are not equally bad, and the way the question is framed changes which one you get more of. Both are benchmarked, on clean exports and on scans, because the two file kinds fail in opposite directions.

Re-cutting doesn’t cost you your work.

A document’s identity is the pages it occupies, not an id handed to it. When a later run draws the boundaries differently, an existing approval carries over if the new range sits entirely inside the one that was approved, and holds back when it only partly overlaps — rather than inheriting a decision nobody actually made about those pages.

Split and merge are mirrors of one another. A proposed split has to account for every page of the document it came from, or approving it would silently drop whatever was left over.

A 104-page packet.Every edge found.