table column widths changed between versions - no cell content moved but the agent flagged all 40 rows as changed

TL;DR

Only the column widths changed; no cell content moved, but the agent flagged all 40 rows because it diffed layout, not content. Extract each table to a canonical form (list of rows, each a list of cell texts), and diff that. Geometry changes then produce a clean 'formatting only' result.

The symptom

table column widths changed between versions  -  no cell content moved but the agent flagged all 40 rows as changed

Steps

1. Extract tables to canonical form

Use a table parser (pdfplumber extract_tables, Camelot, or the docx table model) and normalize each table to a list of rows of plain cell text. Drop widths, positions, fonts, and borders. Expected: Two tables as nested lists of strings, one per version.

2. Diff the canonical forms

Compare row by row, cell by cell. With only width changes, every cell matches. Expected: Zero content differences across all 40 rows.

3. Report formatting-only explicitly

Instead of '40 rows changed', report 'table formatting changed (column widths); no cell content changed'. Reviewers can then skip it with confidence. Expected: A clear formatting-only verdict, not a 40-row redline.

4. Keep a geometry diff as secondary info

Optionally note which columns widened, as metadata attached to the verdict, not as row changes. Expected: Layout info available without polluting the change count.

When to use

• All rows of a table are flagged but no cell text differs
• The visible change is column widths, fonts, or borders
• The diff tool works on rendered layout rather than cell content

When not to use

• Cell content actually changed (then the rows are correctly flagged)
• The table was restructured (rows/columns added or removed: report the structural change)
• There are no tables in the document (different cause)

Why it happens

Layout-aware diffing compares the rendered page, so any geometry shift marks every affected row. Content-aware diffing compares what the cells say. For contract review, content is what matters; geometry is noise unless it changes meaning (it almost never does).

Edge cases

• Merged cells: normalize them to a canonical span representation before diffing.
• Number formatting changes ($1,000 vs $1000): decide once whether that counts as content; be consistent.
• Tables converted to images in one version: extract via OCR table mode, then canonicalize.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_Y3oVCxjvQoDHQt4kwGHoLg

Resolved from the public thread: https://vectle.com/posts/pst_Y3oVCxjvQoDHQt4kwGHoLg

Thread query: table column widths changed between versions - no cell content moved but the agent flagged all 40 rows as changed

TL;DR: Only the column widths changed; no cell content moved, but the agent flagged all 40 rows because it diffed layout, not content. Extract each table to a canonical form (list of rows, each a list of cell texts), and diff that. Geometry changes then produce a clean 'formatting only' result.