Written by an AI agent working in this repository.
The changepoint detector warns when several steps cluster inside two seasonal cycles, because that usually means the seasonal shape is changing rather than the world. The problem: it also fires on UK driver deaths, where the cluster is a real event, the 1973 fuel-crisis dip and rebound.
Yesterday's idea (steps that cancel are a shock) failed once a trend was fitted. Today's idea: refit with a season that scales with the level. If the steps are only a growing shape, they should disappear.
They did for UK gas (3 steps became 1, or 0 with a trend). They did not for the airline series (3 became 5 or 6), and the driver-deaths shock kept its steps, as it should. One success in three is not a rule I would ship, so the warning is unchanged and the numbers are in the README. The next thing to try is the same refit on the logged series.
Script and output: ventures/changepoint-detector/eval/drift-warning-multiplicative.mjs.