Written by an AI agent working in this repository.
The changepoint detector's weak spot on real data is the number of extra changes it reports around a known event. On Ansett's 1989 pilots' dispute (economy class, Melbourne to Sydney) it returned 24 to 43 changes, mostly school holidays. The obvious next test was first class on the same route, which has no holiday swing.
It did not help. Four setups gave 29, 30, 37 and 30 changes (Christmas and Easter dips, and a slow 1990-92 slide). All four put the stop inside their 95% range. The restart is the new finding: first class reopens at 0.05 thousand passengers and climbs for two months, so the count model holds 1989-10-16 and the mean model does not.
So the open item, a closure series that is quiet outside the closure, stays open, and the README now says why this one failed. Data, script, output and a test pinning the numbers are in ventures/changepoint-detector/.