By way of confirming Peter Shor's observation about Ngram's default case-sensitivity, here is the Ngram chart for "Jabberwocky" (blue line) and "jabberwocky" (red line) for the period 1800–2000 (with smoothing reduced to 0 from 3):

The earliest match that the Google Books search results listed in the links beneath this graph find is from the June 15, 1872, in Once a Week; and the next-earliest match is from the December 1882 issue of St. Nicholas. In both instances, Jabberwocky is initial-capped.
If you reduce the time period from 1800–2000 to 1870–1900, you get a magnified image of the relevant segment from the first chart:

But in the links beneath this graph, you'll find nine unique, previewable matches from the period 1870–1879 alone, including an 1875 edition of Through the Looking-glass: And What Alice Found There, which very conveniently details in its front matter the moves in a chess game corresponding to Alice's adventures, in which "White Pawn (Alice) to play, and win in eleven moves."
The main point here is that the degree of detail provided in the Ngram-associated Google Books search results varies tremendously depending on what time interval you use, but the Ngram line graph seems quite stable (aside from predictable changes in magnification).
As for a baseline noise floor, much depends on the years over which you are running your search. Searches going back to the 1600s produce many OCR errors based on blurred, faded, or broken page images; old-fashioned font sets (especially in connection with italics); foreign words misread as English ones (especially if Latin texts are proportionately much more common in the results); books and journals with multiple columns per page and narrow spaces between them, which the OCR sometimes jumps incorrectly; and (in particular) words that use the old-fashioned lowercase "long s" (ſ), which Google's OCR system often misreads as a lowercase F (f).
As you get closer to the present, image quality generally improves and the incidence of OCR errors drops considerably. By the middle-1800s, Ngram/Google Books OCR accuracy is much better than for the early 1800s and before. Nevertheless, searches that bring up disproportionately many newsprint articles have a higher noise problem, due to small type, smudging, and fading of the type.
So my conclusion is that there is no consistent baseline or floor to Ngram noise. Every search must be appraised and fine-tuned on its own unique terms. But the only way to get a sense of what results to trust is to go into the search result links and check each match.
The earliest match for filosophy, for example (and the one responsible for the skyscraper centered on the year 1820 in your link), is to a page in an 1820 Parliamentary history that got creased, causing successive lines to lose the equivalent of 1½ visible characters; hence the OCR misreading. But matches for filosophy from the late 1800s are not errors—they are matches to publications that evidently took a half-baked approach to spelling simplification and balked at filosofy.