# Document formatting change ledger

This is an actual GPT-6 Astra medium run in Codex on 7 September 2026, using the supplied synthetic DOCX. LibreOffice exported the PDFs; Microsoft Word was not tested.

## Formatting changes

- Retained US Letter portrait (612 × 792 PDF points / 12240 × 15840 DOCX twips); set all four margins to exactly 1440 twips (1 inch). Source margins were top/bottom 900 and left/right 1100 twips.
- Changed default Arial body size from 11 to 12 points and set dark text (#222222).
- Preserved all four heading paragraphs and their Heading1/Heading2 assignments. Set Heading 1 to 18 points, Heading 2 to 14 points, bold dark text, explicit outline levels 0/1, and keep-with-next.
- Set body spacing to 6 points after, 0 before, and 1.15 line spacing. Headings use 12 points before and 6 after.
- Preserved blue underlined hyperlinks (#0563C1) and both exact targets.
- Kept table at 9360 twips, exactly the 6.5-inch content width, with 6240/3120-twip columns and matching cell widths. Added padding, pale header fill, bold header text, soft borders, a repeating header, and non-splitting rows.
- Kept illustration dimensions and image bytes unchanged; added keep-with-next so the illustration and its original caption remain together.

## Preservation and visual checks

All ordered paragraph strings, four headings, five table rows/two columns, two hyperlink targets, footnote XML and reference, image description, and actual PNG SHA-256 match the source. This preserves wording, negative sign -15.00, precision, dates, caption, and order. The source DOCX hash is unchanged. Every nonempty body paragraph is present in final PDF extracted text after whitespace normalization. Both PDF link annotations retain their original destinations.

All exported pages were rendered at 100 dpi to raw PNG and visually inspected: original fallback-font source 1 page, first output 2 pages, final Arial source 1 page, final Arial output 2 pages. No clipping, overlap, orphaned headings, split table, or detached image caption was observed. The first final page has deliberate open space because the illustration and caption move together to page 2. Final PDF font resources are ArialMT and Arial-BoldMT. No browser/chat interface has been fabricated.

## First failures and corrections

1. First preservation script included the ZIP directory entry `word/media/` as an image and reported media=false. The actual PNG hashes already matched. The original inventories and `preservation-checks.json` are retained; `preservation-checks-final.json` excludes directory entries and passes.
2. First LibreOffice exports substituted Liberation Sans for the Arial requested in DOCX. `source.pdf`, `first-output.pdf`, and their before-page/after-page PNG renders are retained. A task-local fonts.conf points to existing macOS Arial fonts; re-export produced `final.pdf` and `before-arial/source.pdf` with Arial. No font was installed or copied for publication.
3. The skill schema validator failed before validation because bundled Python lacks defusedxml. The failure is in schema-validation-first.log. No dependencies were installed. XML parsing, object comparisons, successful LibreOffice import/export, and visual inspection passed; full schema validation and Microsoft Word compatibility remain unverified.

The first DOCX was not changed after generation: `first-output.docx` and `final.docx` are byte-identical. Corrections affected export configuration and the verification filter, not document content. `source-to-final.diff` records actual document/style XML changes. `correction-diff.md` describes the export-only correction.