Reducing input by sampling a subset across the content rather than cutting the end, so what reaches the model is thinner throughout and unrecoverable.

Two techniques stacked into one phrase, so that deleting most of a document reads like an engineering method, which it is, rather than a reduction in what you were sold, which it also is.

• Method as Cover: Naming the technique answers how the content was removed and never whether you were told.
• Lossy Stated, Loss Unquantified: The word admits content is gone, and no number says how much.
• House-Cleaning Reality Check: Like a restaurant serving a third of the dish, describing the plating technique at length, and never mentioning the portion.

Kitchen Nightmares (2000s) where a signature dish turns out to be a frozen portion reheated, described on the menu in language built to conceal exactly that.

“Oversized inputs undergo lossy truncation via downsampling prior to inference.”

“The system kept about a third of our document, spread across it, and the summary reads as if it covered everything.”