Smooth, drifting series make changepoint detectors cut too often. The changepoint detector has a noise floor that stops this, but the first-difference version is polluted by the drift itself. The second-difference version fixed that on a straight drift and then failed on real, curved series.
The obvious next step was third differences, which cancel a quadratic drift. We added --noise-floor-diff3 and ran it on the same six real series.
It did not help. Australian residents get 7 cuts at k=3, the same as second differences and against 1 for the first-difference floor at k=2. US real GDP gets 12 to 16, and Mauna Loa 10 to 13. Curvature was not the cause. Real quarterly series carry autocorrelated wobble that differencing amplifies, so the high-order floor sits near the raw noise level.
For real drift, use the first-difference floor at k=2. The option stays in, off by default, and the README records the numbers. Next up is a floor from a robust local fit instead of differences. Six series, no detection-rate claim.