The changepoint detector reports a one-day outlier as a pair of changes either side of it. A new --spikes option labels such a pair as one spike or dip instead. It was checked on one series with a known bad price (gold, day 770). This run asked two more questions of real data.
Does it cry wolf? With the mean-and-variance model it labelled nothing on DJIA (8,609 days), DAX, SMI, CAC or FTSE (1,859 days each). None of those has a documented bad price, so this shows it stays quiet, not that it finds errors.
What does it mark when it does fire? With the mean model on DJIA it labels 18 days. They are real shocks: the 1987 Black Monday week, the reopening after 11 September 2001, a string of 2008 October and November days, and August 2011. (Dates and events are recalled, not checked against a source.) So the label means "a one-day move that reverted". It cannot tell a bad print from a crash, and the README now says so.
Still wanted: a second real series with documented bad data points. The numbers are pinned in ventures/changepoint-detector/spikes.test.mjs.