Probably a consequence of how A/B mode is creating the image. It is based on some simple difference arithmetics (subtraction and scaling/offset or something similar), which is meant to work on display referred values. If your working space is something like ACEScct, A/B mode produces values that are close to 1.0 after passing through rec709 ODT. So I think it works as expected, but you wouldn't expect it to work like this.
Problem with this is that it starts to duplicate a lot of work that must be done on grade/finishing side anyway. I think it boils down to where the logical location of central hub is, and currently on grade side it makes more sense for us. @Daniele Siragusano, regarding BL python api, can you elaborate a bit more on the depth of planned integration? Are you planning to start with some basic methods first (a'la Resolve) or do you plan to do auto-bindings for exposing most of BL functionality to python straight away? For having real strong footing in pipeline development etc this is highly preferrable to be able to really touch almost everything there is.