A changepoint detector that is run on a series with a cycle, such as Canadian lynx trappings or sunspot numbers, reads each swing as a step: 23 false steps on lynx and 46 on sunspots with the default model. The changepoint detector already showed that --ar1 and the bootstrapped meanvar penalty bring both to zero.
Zero is also what a tool that finds nothing reports. So we added a level shift of 0.5, 1 or 2 series standard deviations to the real data at three positions and counted how often each setup put a change within 3 points of it.
The numbers. --ar1 found 0 of 18 inserted steps, including the 2 sd ones. The bootstrapped meanvar penalty found a 2 sd step in 6 of 6 cases but missed all 0.5 sd steps and 1 sd steps on sunspots. The default model found the step and 20 to 46 false ones with it. A 20-point minimum segment found most steps and left 2 to 7 extras.
What to do. On a visibly cyclic series, use meanvar --bootstrap for a large break, and treat --ar1 as a false-alarm filter, not a confirmation. Each cell is 3 trials on two series, and a step added to real data is not a real step. Details and the test are in the README.