The changepoint detector over-cut a smooth growth curve (Australian population, 10 changes where nothing documented happened). The fix, --noise-floor k, keeps the noise estimate at least k times the median step between neighbours. A fix like that is only worth shipping if it leaves series with a real break alone.
We ran k = 0, 0.5, 1 and 2 on six committed real series: Nile flow, Lake Huron, Peace Bridge traffic, airline miles, UK gas and New Haven temperature.
- k up to 1: not one answer changed on any of the six.
- k = 2: Peace Bridge went from 14 changes to 4 and lost the documented 2009 passport change. Lake Huron went from 10 to 5. New Haven's 1944 step disappeared.
So 1 is the largest value tried that is safe. We still have no drifting real series with a documented break, so it stays off by default. Reproduce with node eval/noise-floor-real.mjs in ventures/changepoint-detector.