revision-analyser

Reads a Word document's drafting trajectory — not its final text. The lens-family member that surfaces how a document was written, not what it says.

Parses Word's embedded Track Changes (<w:ins>, <w:del>, <w:moveFrom>, <w:moveTo> elements in document.xml) to produce:

• timeline of insertions/deletions/moves with author and timestamp
• paste-burst detection (a single insertion ≥ N words ≈ paste from AI/elsewhere)
• per-author rollups (single-author vs collaborative editing)
• thrashing patterns (many small clustered edits)

The single strongest written-work process signal — but only if the student had Track Changes on. A .docx with no revisions is itself a signal: track_changes_enabled: false (often means the final version was pasted in). Explicit-only: .docx continues to auto-route to document-analyser for content; invoke revision-analyser deliberately for the revision history.

Scope: v1 reads Word .docx tracked changes only. Google Docs revision history (Drive API) and draft-sequence comparison are planned for v2.

Install

pip install revision-analyser

Use

Python:

from revision_analyser import RevisionAnalyser

result = RevisionAnalyser().analyse("essay.docx")
print(result.track_changes_enabled)        # True
print(result.total_insertions)             # 47
print(result.paste_burst_count)            # 2
print(result.authors[0].name)              # "Jane Student"
print(result.timeline_minutes)             # 184.5  (≈ 3 hours of editing)

CLI:

revision-analyser essay.docx
revision-analyser essay.docx --json
revision-analyser serve
revision-analyser manifest

HTTP (revision-analyser serve on port 8016):

curl -F file=@essay.docx http://localhost:8016/analyse

Signals

• track_changes_enabled — whether the doc carries any tracked revisions at all. Absence is itself a signal (final-paste workflow).
• Totals — insertion count, deletion count, move count.
• Word-level totals — insertion_word_count, deletion_word_count.
• Paste bursts — insertions whose single contiguous text is ≥ 25 words (tuneable). These are the strongest paste-from-elsewhere markers.
• Authors — list with per-author insertion/deletion totals. One author = self-edits; two = collaborative or feedback cycle.
• Timeline — timeline_first, timeline_last, timeline_minutes (span of activity).
• Events — a capped sample of the timeline for inspection.
• Flags — composite tags: paste_burst_present, single_author, no_revisions_recorded, short_timeline, multiple_authors.

The family

Triangulation matters here. A polished essay with no recorded revisions + a short total editing time (provenance-analyser) + a descriptive depth band (reflection-analyser) tells a much more focused story than any of the three signals alone.

Limits

• v1 reads tracked changes only. If the student never enabled Track Changes, the doc reports track_changes_enabled: false — the absence is informative but bears no detail about the actual editing process.
• Author names are whatever Word recorded (w:author attribute) — often an install-time username, sometimes "Author" if anonymous.
• Timestamps reflect when the tracked change was made per the authoring machine's clock; trustworthy on a single machine, less so across collaborators.
• Paste-burst detection is a heuristic — long literary quotes will also score. Surface the burst text in the response so callers can audit.

License

MIT