redline agent flagged 200 false positives - the counterparty reflowed the document and the diff treated every moved...

TL;DR The counterparty reflowed the document and a line-based diff treated every moved paragraph as a deletion plus an addition, producing 200 false positives. Normalize whitespace and line breaks before diffing, then align paragraphs by content similarity so moved text is reported as moved, not changed. The diff should show what actually changed, not what merely moved.

What the noisy diff looked like:

- This Agreement is entered into as of the Effective Date
+ This Agreement is entered into
+ as of the Effective Date

Steps

1. Confirm the false positives are reflow, not real edits. Sample a dozen flagged changes and check whether the same words appear elsewhere in the new version.

Command: pick a flagged paragraph and grep for a distinctive phrase from it in the new document text. Expected: most flagged text exists verbatim in the new version, just at different positions, which means reflow, not revision.

1. Normalize both versions before diffing: collapse all whitespace runs to single spaces, unify line endings, and strip page breaks so pagination differences cannot create diffs.

Expected: re-running the diff on normalized text drops the change count dramatically.

1. Switch the diff from line-based to paragraph-based alignment with content similarity. Match paragraphs between versions by token overlap, and when a paragraph matches above a high similarity threshold but sits at a different position, report it as moved.

Expected: moved paragraphs appear once as "moved from page X to page Y" instead of as a deletion plus an addition.

1. Set a review threshold on the remaining changes: anything above it gets a human skim of the move report before the redline summary goes out.

Expected: the redline summary lists only substantive changes, with moves called out separately.

Use this when

• A redline agent reports hundreds of changes after the counterparty reformats
• The same paragraph appears as both deleted and added in the diff
• Pagination, margins, or fonts changed between versions
• Your diff tool compares raw lines instead of semantic paragraphs

Not for this skill when

• The changes are real edits with different wording, that is what the diff is for
• Only a handful of changes are flagged, skim them by hand instead of retooling
• The documents are different file types with different extraction, fix the extraction first
• You need a legally precise redline for execution, moves still deserve a human eye there

Variant phrasings

• redline flagged 200 false positives after document reflow
• counterparty reformatted the contract and the diff shows everything changed
• how to diff contracts ignoring reflow and moved paragraphs
• line-based diff treating moved text as deletions and additions
• redline summary full of formatting noise

Why it happens

Line-based diffs compare sequences of lines, and reflow changes where every line breaks even when no words change. Each broken line looks deleted from the old version and added in the new one, so a pure reformat produces hundreds of phantom changes. The diff is technically correct about the lines and completely wrong about the document, because contracts are read as paragraphs and clauses, not as lines.

Edge cases

• A paragraph that moved AND was edited should be reported as edited at its new position, not just moved; use the similarity score to distinguish.
• Tables and lists reflow badly, align them by cell content rather than by line position.
• Headers and footers change with pagination, strip them before diffing or they will dominate the change list.
• Very short paragraphs like "Section 5." match spuriously by similarity, require a minimum token count for move detection.
• Keep the raw line diff available for audit, the clean summary is for humans but the raw diff is the record.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_IA1MJMI0CQXQaJoO0BZMxw