Decision. When the language model returns a wine that no geometric row pairs with, that wine is shown only if its name — whole, or either half — actually appears in the text the scanner read off the page. A nameless unpaired row is dropped, and logged.
Because. The refine step invents. Measured on eleven pages that are not wine lists at all, the geometric parse found 39 entries and the model refine turned those into 59. A reader that adds wines to a dessert menu adds wines to a real wine list too, and those are the ones you cannot spot by eye.
Instead of. Trusting the model because it is usually right, or dropping every unpaired row and losing the wines only the model could read. Grounding keeps the recall and spends it on text that exists.