Typically models don’t work like that... i.e changing the model or its training alters it’s behaviour generally and needs to be tested against its overall perfo

Typically models don’t work like that... i.e changing the model or its training alters it’s behaviour generally and needs to be tested against its overall performance. So whilst crash data may go toward a dataset which is then used to better evaluate or train different aspects of a system, it has none of the luster of the learning from every accident being maximised. The human mind is a quite unique in its ability to decide where one model stops and another one starts.