Quite an interesting read. I wonder how well this system would work if you used something other than Midscene.js to fingerprint the models? It was limited in scope on how the models interact with the sites for study's sake, but would be interesting to look at further.