Hello Simon,
I observed this same problem a while ago (see Bug in OD with multiple CD estimation?). There is already a gitlab issue open about this, so it’s definitely not the intended behavior.
For now, the workaround I use to avoid this problem is to reconstruct the normalized parameters myself using the value during each interval, the reference value, and the scale, and then passing this new set of normalized parameters to PropagatorBuilder.buildPropagator(normParams)