I think result-filtering (based on license of search results) is gnarly enough, and likely computationally intensive, so as to break the whole feature. But it would be interesting to see if that can be crafted to fix the shortcomings of the ML model.