P→MPDFMDSystem documentation
Format referenceOpen converter ↗

Level 3 · Method index

Last modified: (dates/navigation repair). Content current as of: (last retained content commit; content verification date unavailable). Original content last edited: . Current claims have not been freshly reverified.

How recognition decisions are made

Each method combines several observations, tests relevant look-alikes, and records the basis of its decision. The pages below begin with the reader’s question rather than an implementation module.

Questions

Choose one problem

What has recognition actually got wrong?

Exact source images, observed errors, expected results, failure causes, and the current repair plan.

→ (outdated · September 2026)

Which lines belong together?

Physical lines, columns, paragraph boundaries, reading order, and cross-page continuation.

→

Is this a numbered paragraph or a list?

Printed labels, embedded lists, nesting, restarts, continuation prose, and quotations.

→

Is this text bold, italic, underlined, colored, or highlighted?

Native metadata, scan measurements, relative style, line geometry, scripts, and exact font identity.

→

Is this a footnote, footer, or page number?

Calls and bodies, recurrence, page zones, concurrent numbering, and page relationships.

→

Is this a table, signature, stamp, or unknown region?

Cell graphs, borderless alignment, multi-page continuation, meaningful graphics, and retained unknowns.

→

Common contract

Requirements shared by every method

  1. Identify the source observations used.
  2. Attach confidence only to the field being decided.
  3. Test positive examples and similar-looking negative examples.
  4. Preserve competing interpretations when evidence is insufficient.
  5. Represent every meaningful source region, suppress it with a reason, or mark it unknown.
  6. Avoid rules tied to fixture wording, filenames, exact coordinates, fixed page counts, or known answers.