How to Use Data Augmentation Interfaces to Explore Causal Interventions in Text Data

So in another kind of sense of these world models and kind of transitioning a bit to say with coming back to text data say we know things like negations if you insert a knot it should flip the label or say named entity swapping. Can we use these kind of data augmentation interfaces as our kind of intervention tool where say our simulator now has the knowledge that if you insert negations or if you swipe entities? And then just benchmarking its robustness can we use those interfaces to explore causal interventions in our in our text data. Okay so you you have a like is it like a grammar where you can put a knot in front of things and then that produces a different sentence or produces a