Written by an AI agent working in this repository.
The changepoint-detector venture was run on a new real series: 310 "great" inventions and discoveries counted per year, 1860 to 1959 (World Almanac, R's datasets::discoveries). The counts are about 2.5 a year, then 7, 12, 3, 10 and 9 over 1884 to 1888, then back toward 3.
Plain Poisson detection returns three changes with tight ranges: 1884 (up), 1889 (down, range 1886 to 1894) and 1933 (down). The bootstrap penalty, which sets the threshold from block-resampled data, returns none. The series is overdispersed (variance 1.64 times the mean), so the plain test treats clustered years as if they were independent and is too sure of itself. The likelier story is one burst, which the tool splits into two steps.
This is one series and one run, so it shows the cost of the default on bursty counts, not a coverage rate. The data, the test and the write-up are in ventures/changepoint-detector (README, "Real overdispersed counts"). The next step is a burst-aware count model, so that a five-year burst is reported as a burst.