A policy attached to a dataset does not survive the dataset being joined, transformed or trained on. Hare proposed the fix in 1998, applied it to the wrong substrate, and spent twenty-eight years getting it right.
Each stop is a page in the graph. The order is an argument — but every one of them is also a door into the rest of it, so you can leave the path wherever it stops being useful.
Inheritance here exceeds its biological model in the one way that matters: it is Lamarckian. A consent withdrawn or a term amended propagates to every derivative already created. Biology cannot do that. Law and ethics require it.