The model has no memory between shots.

Every generation is independent. Shot four does not know what shot one looked like, and nothing in your prompt file makes it so — which means an attribute stated once and not repeated is re-rolled in every later shot. Describe the red waxed-cotton jacket in shot one and leave it out of shots two to six, and six different jackets come back. The document reads perfectly the whole time, because a human reader carries it forward without noticing they are doing it.

The engine is free and needs no account. Paste the cast and the shot list and it builds the restatement ledger: one row per locked attribute, one mark per shot, filled where the line states it and empty where it does not. The gaps are the finding. It also separates an attribute that was never restated from one restated as something else, distinguishes a paraphrase written from memory from a real contradiction, catches the two-shot that names two people and one jacket with no owner on it, and reports an hour that goes backwards. Every lane ships a worked example that costs nothing. Writing the report needs credits.