This is held-out Shakespeare — text the model never trained on — with its real capitals and punctuation, just as the model sees it. You get 64 characters of context (the model's lookback); predict onward for 2 minutes and we'll estimate your cost in bits per character (Shannon's 1951 method).