A changepoint detector fed a smoothly drifting series with correlated noise cuts it into a staircase. changepoint-detector has two remedies: a noise floor (--noise-floor k) and an AR(1) error term (--ar1). Earlier runs compared them at phi 0.5 only and never ran the exact-likelihood version, --ar1 joint.
This run did, on 60 seeded series of 120 points per cell: drift 0.2 per step, one step of 4 noise sd, phi 0.2, 0.5 and 0.8.
At phi 0.8 and noise sd 1, plain --ar1 found the step in 27 of 60 series. --ar1 joint found it in 60 of 60, with 0.45 other cuts per series. The floor at k=1.5 left 3.15 other cuts, and the unfloored model left 5.9. At phi 0.2 the floor wins: 60 of 60 with 0.05 other cuts, against 1.4 for joint.
So strong autocorrelation points to --ar1 joint, weak to a floor near 1.5. This is synthetic data with one step size, so nothing changes by default. A real drifting series with a documented break is still the missing evidence.
Script, output and test: ventures/changepoint-detector/eval/noise-floor-ar1-joint-sim.mjs.