Research notes: preserving layout when transforming documents

One of the questions guiding Lúmen's research sounds simple: after translating or anonymizing a document, does it still look like the same document?

Why it matters

For whoever uses the result, layout is information. The position of a signature, the alignment of a table, the hierarchy of a contract — all of it carries meaning. Returning plain running text forces someone to rebuild the document by hand, which cancels out much of the automation.

What makes the problem hard

• Text changes length. A translated sentence can be 30% longer and no longer fit in the original space.
• Scanned documents have no structure. Before preserving the layout, you have to rebuild it from the image.
• Tables and forms break easily. Merged cells, borderless columns and handwritten fields confuse the reading.
• Anonymizing isn't just deleting. A badly placed redaction box can hide too much or too little.

How we're approaching it

The current line of investigation treats the document as two layers: content and geometry. We first rebuild the geometry — blocks, lines, tables, positions — and only then transform the content, respecting the boundaries of each block.

This is a research post, not a product post: it describes a problem under study, not an available feature. When there's a usable result, it will show up in Lúmen's changelog.