This task is optional, depending on how many RC rows remain unexplained after accounting for visible text diffs.

• Normalize out meaningless changes, eg. by stripping the Parsoid id attribute from all elements.
• Choose a library (native Python or an external command-line tool) which produces a diff between two HTML documents.
  • Ideally, this diff will itself be HTML so that it's possible to view the changes in the context of the document in a browser.
  • Ideally, invisible changes for example inside of data-mw attributes can be suppressed or listed separately as invisible_changes: true/false
• Test this algorithm thoroughly on a few actual old/new pairs from our sample—the real-world diffs are numerous and annoying. Iterate by normalizing out anything we don't care about.
• Automate some of the classification of these diffs if that makes sense: for example, img alt attributes.