Mace-legal teaches Mike to read redlines

Most legal AI tools quietly flatten a document before the model sees it - accepting all changes, dropping comments, ignoring the red and blue strikethroughs that carry half the meaning in a negotiation. Mace-legal's fork pushes the other way. On the Word side, insertions, deletions and reviewer comments (with author attribution) are pulled out and handed to the model as inline markers it knows how to interpret. On the PDF side, a new helper reads the actual span colors - red for cuts, blue for adds, green for moves - to reconstruct redlines from the kind of marked-up PDFs that Litera or Workshare spit out.

The model is then explicitly taught what to do with all this: treat comments as marginalia, not body text, and decide whether the user wants the marked-up or clean version. The PDF path leans on PyMuPDF, a Python library for reading PDF internals, and the author notes the algorithm is ported from a separate diff-tooling product they maintain.