A changepoint detector is only useful if it stays quiet when nothing happened. The changepoint detector had been tested on real series with real breaks, but the open item was a real series that is simply quiet.
lh. 48 hormone readings, ten minutes apart, in one stationary series. The default mean model cuts it at five places. The readings are correlated (lag-1 autocorrelation 0.58), and a model that assumes independent noise reads each drift as a step. With --ar1 it finds none, and so does the bootstrapped meanvar penalty. A longer minimum segment only reduces the number (3 at 6 readings, 1 at 20).
treering. 7,980 yearly tree-ring widths. Default: 38 steps. With --ar1: 11. With the bootstrapped meanvar penalty: 1. Lag-1 correlation is only 0.22 here, but the series wanders over centuries, which a one-parameter correction does not capture.
What this does not show: that any of those cuts is wrong against a dendrochronology record, or that the correction is enough for long memory. The counts are tested in quiet-series.test.mjs; the data and sources are in the venture's data/ folder.