I dont mind how they do it, but cab ir modeling rather than hunting cab by cab speaker by speaker through thousands of irs
My sense is that a lot of this "modeling" really just applies an EQ transformation that varies with distance from the speaker, or phase, or distance from the center cone. This is how you get continuous variation from a static IR. (I could be wrong about this, though.)
Maybe what we really want is not modeling, but a tool where the program plays a recorded sample of an amp, and you can work through a bunch of IR samples quickly by saying "better" or "worse" or "no change" like you do with an optometrist.
For things like a Celestion speaker sample that includes a single speaker with a bunch of IRs with different mics, at different distances, and in different combinations you would have to have different axes of variations. One might be distance, one might be mics arranged by brightness or fullness (or high/low frequency response), and then some kind of way for getting at how multiple mic combinations differ. If you could figure out the right axes of variations, and the differences in the IRs are measurable, the program could sort IRs automatically for you.
You then could use the user interface to move in a general direction (further from the speaker, brighter, flatter) on the axes, and then the program starts offering you pairs of IRs to compare to see which you like better. If you don't know what you want, there would be a lot of variation (difference) between the IR pairs you are comparing, as you start to zero in on a sound the variations would be smaller and you could fine tune the tone, and eventually you keep picking one IR as "best" in a given context (e.g. best for an amp model played with a specific recorded sample) and then you are done and have a winning IR.
You could also have a library of recorded guitar samples in different styles, for which you select different amp models, and then fine tune the IRs as above.
If the program can analyze IR characteristics consistently, it could scan your whole library of IRs and arrange them along the different axes of variations. This could be done manually for the stock library of IRs, and then analysis could fit third party IRs into the gaps.