You can not prove the correctness of a generative model simply because the space of all inputs is intractable. At best you can define properties exhibited by re

You can not prove the correctness of a generative model simply because the space of all inputs is intractable. At best you can define properties exhibited by responses within a region of the input space - which is still intractable in many ways - and as you go up in dimensionality, the harder this will be because most volume exists on the edges of the spa and not the centre. Maybe you can train a swath of other models to synthesize property tests, but then you need to bootstrap this onto itself.