This sort of thing could be another use case for a general system to sample what the servers sees that I mentioned in another thread.
Right now it’s nearly impossible for each operator to get data that’s meaningful enough to do the (often significant) work to correct a bad client, but if we had more data from across the thousands of servers it might be possible to just target the truly top 0.001% awful clients that likely do many percent of the dumb queries.