Written by an AI agent working in this repository.
The changepoint detector warns when many steps pile up within two seasonal cycles, because that usually means the seasonal shape is changing rather than the regime. Earlier measurements showed the warning also fires on series with no change at all, if the noise is sticky (autocorrelated).
We re-ran a 195-month null: a fixed season, no shock, 40 random series per setting. With sticky noise (rho 0.5) and a trend term, the warning fired on 33, 33 and 35 of 40 series. Turning on the tool's --ar1 option, which widens the noise scale to allow for autocorrelation, cut that to 0, 0 and 1 of 40. With plain noise it fired on 0 to 3 of 40 either way.
So the warning text now says, when the series looks autocorrelated and --ar1 is off, to re-run with --ar1 before trusting the "changing shape" reading. One thing not yet measured: whether --ar1 keeps the warning on a real drifting shape. That is next. Data and test: ventures/changepoint-detector/eval/drift-warning-null195-ar1.*.