Draft A has a motel-room scene. In draft B the writer moves it after a phone call, adds a short scene ahead of it, and changes the heading from INT. MOTEL ROOM - NIGHT to INT. ROADSIDE MOTEL - LATER.
A line-oriented comparison reports a scene deleted and a different one created. Nothing was deleted.
Match scenes before you read words
Page numbers and scene numbers are delivery coordinates. Before a lock they shift whenever a scene is added or cut. After a lock an inserted scene becomes 42A precisely so the old numbers keep pointing at the same thing for the sides, the schedule, and the lined script. Either way the number tells you where a scene sits in a delivery, not which scene it is.
So the first question about draft B's motel scene is whether it is the same dramatic unit as draft A's: same people, same encounter, same job in the sequence. That gets settled before a changed sentence is worth reading.
Then compare five layers
Identity. Was the scene edited, moved, split, merged, or genuinely replaced? Those are five different operations with five different consequences, and a moved scene is not a deleted one. A scene split in two can preserve nearly all its text while changing the rhythm of the sequence around it.
Dramatic intent. Write one sentence for what the scene accomplishes, in each draft. Mara tries to learn whether Joel has the key and Mara lets Joel believe she does not know about the key can share a setting and most of their dialogue while the scene's job moves from investigation to deception.
Information flow. Track what the audience learns and what each character learns or believes — they are not the same ledger. Moving one reveal three pages earlier alters the suspense of every scene between the old position and the new one, and the textual edit is tiny while the structural effect is not. A continuity ledger makes that state explicit.
Character state. Record what survives the scene: objective, leverage, allegiance, knowledge, emotional position. The useful question is not whether a line changed. It is whether the character leaves the scene in a different condition than before — because every later scene was written against the old condition.
Production cost. Once a draft is moving toward production, a revision is also an operational event: locations, cast, props, scene numbers, page locks, revision colors, breakdown data. A semantically elegant rewrite can still be an expensive production change, and a comparison worth trusting shows you both.
Where word-level diffs are excellent
None of this makes textual comparison less valuable. It makes it the last step rather than the first.
Word-level diffs are very good at the last mile: dialogue polish, a dropped beat, a changed name, an altered prop, an accidental deletion, and exact approval of a collaborator's or a model's proposed edit. They are weak only when asked to infer story identity by themselves.
So the order is: establish the unit you are comparing, identify the semantic change, then inspect the exact words.
Match the scenes first, then read the words.