Right, though I'm skeptical that the program would also be able to correctly identify the motivation for the typeface, which drives all of the decisions about h

Right, though I'm skeptical that the program would also be able to correctly identify the motivation for the typeface, which drives all of the decisions about how it behaves. That component of the program would end up being far more complex than generating the actual kerning pairs themselves.