Word fields, tables of contents and cross-references: why plain-text round trips break a thesis

Word stores some content as instructions rather than ordinary characters. A field has a code and a displayed result: Word evaluates the code, then shows the result in the document. The result can look like text or a number even though its value is generated.

`{ TOC \o "1-3" }` says "collect every Heading 1 through 3, with its page number, and show them here". `{ REF _Ref12345 }` says "show the number of the thing bookmarked as _Ref12345". `{ SEQ Figure }` says "this is the next figure in sequence". You see Figure 3; the file contains an instruction that resolves to 3.

This is why figure numbers renumber themselves when you insert a figure in the middle, and why the contents page updates when you press F9. Nothing typed those numbers.

Fields are only one part of the structure. Heading levels are paragraph styles that feed the table of contents. A cross-reference usually combines a bookmark at the target with a REF field at the reference. Footnotes live in a separate document part, and citation managers may use Word fields, content controls or add-in metadata depending on the tool and platform. A workflow can preserve visible prose while damaging any of these relationships.

• Generated navigation: table of contents, table of figures and index entries.
• Numbering: SEQ fields for captions and references to numbered headings, equations, figures or tables.
• Document metadata: page counts, dates, filenames, document properties and linked values.
• Scholarly apparatus: footnote anchors, endnotes, bookmarks and citation-manager links that may not be ordinary text even when they look like it.

The distinction explains a common surprise: typing over the visible result may not edit the underlying instruction, while unlinking the field turns the current result into ordinary text. The document can look identical in both states until content moves or fields are refreshed.

When selected Word content is copied into a system that accepts only plain text, that system receives characters representing the current displayed results, not WordprocessingML field codes, bookmarks or styles. If the returned text replaces the original selection, the live instructions in that selection are gone. Copying does not damage the untouched original file; replacing structured content with the plain-text result does.

The damage is worse than it looks because the document still appears correct at the moment of paste. The numbers are all there. They are simply no longer connected to anything.

• The contents page freezes. It shows the headings and page numbers from before the rewrite. Edit anything and it silently goes stale, because there is no field left to update.
• Figure and table numbering stops being sequential. Insert a figure and everything after it keeps its old number. Delete one and there is a gap.
• Cross-references become wrong text. "As shown in Figure 3" is now the literal phrase, pointing at whatever is now third - or at nothing.
• Footnotes, comments and tracked changes may be omitted or flattened, while Zotero or EndNote citations can lose the links that allow the bibliography to be regenerated. The exact failure depends on what was selected and what the intermediate tool preserves.

That last one compounds with the previous article: a manuscript whose citation fields have been flattened cannot be reformatted by the reference manager at all, because the manager no longer knows which item each citation refers to.

Formatting creates a second class of loss. Plain text has no heading styles, section breaks, captions, table cell boundaries, equation objects, anchored images, headers or page-number settings. Pasting may apply the style of the destination paragraph, which makes a heading look roughly right while no longer being a Heading 2 that the table of contents can find.

Partial copying can be especially deceptive. You may leave the table of contents field intact but replace the headings it reads, or preserve a caption's visible number while deleting its SEQ field and bookmark. The next F9 update then exposes the damage somewhere other than the paragraph you edited.

The risk is replacement, not the clipboard by itself: a plain-text copy is fine for quoting or analysis, but it is not a structure-preserving representation to paste back over the source.

For a 1,500-word essay with no fields, none of this matters. For a thesis, it decides whether the work survives.

Any workflow whose only input and output are plain text lacks the information needed to reconstruct these Word objects reliably. The rewrite can be linguistically excellent and the structure can still be lost, because the loss occurred at serialisation: the tool never received the field code, bookmark, style or relationship to preserve.

This is the reason we process the file itself rather than asking for pasted text. Text eligible for rewriting is handled separately from structural objects, while the document remains the container. File-level processing still needs verification: after any tool returns a thesis, open the result in Word, update fields and inspect the relationships that matter to your document.

The longer the document, the less useful a visual spot check becomes. A broken bookmark on page 40 may not surface until a cross-reference on page 190 updates. Build a small acceptance checklist around object types - contents, one figure reference, one table reference, one footnote, one citation and one section break - rather than reading only the first few pages.

• Work on a copy. Keep the original DOCX and make a versioned processing copy before uploading, unlinking citations or accepting bulk replacements.
• Toggle field codes before you start. Alt+F9 in Windows Word or Option+F9 on Mac toggles field-code display; some Mac keyboards also require Fn. Record what fields and bookmarks the document actually contains.
• Update fields using the platform's selection shortcut. On Windows, Ctrl+A then F9 updates the main story; on Mac, Command+A then F9 (often Fn+F9) does the equivalent. Headers, footers and other text stories may need separate updating.
• Check a downloaded file the same way. After any tool touches your document, toggle codes and confirm the TOC and figure references are still instructions rather than text.
• Keep the citation plugin fields intact. In Zotero, "Unlink Citations" is irreversible for that copy - it converts every citation to plain text deliberately. A plain-text round trip does the same thing without asking.

Run a change test, not just an appearance test. In a disposable copy, insert a temporary caption before an existing figure, update fields and confirm later captions and cross-references move together. Add a temporary heading and refresh the contents. Undo or discard the test copy afterward. Live relationships reveal themselves by responding correctly to change.

Also inspect elements that field-code view does not cover: comments, tracked changes, footnote anchors, equations, floating images, section-specific headers and citation-manager refresh. Compare document statistics and page count for unexpected jumps, while remembering that benign reflow can change pagination.

A document that looks right before fields are updated has not yet passed the test. Refresh, exercise one live reference and keep the untouched original until submission is complete.