There's also the point that one histogram is an order or two of magnitude larger than the other one. Larger samples of normally distributed data will tend to better resemble a normal distribution.

There frequency of "game-changing" in comments would say otherwise. It's probably cron running at those intervals, not a work schedule.

I hate game-changing as much as the next guy, and was ranting about it on here just the other day, but some people really do talk like that.

Have you tried running any network analysis on these bots? I would expect to see strong clustering, and I think that's usually the primary way these things are identified. The prompt injection is an awesome approach though!

Yes I did on subsets of the data because cupy and cudf haven't implemented intersection functions yet for the GPU. But the clustering is weak because new signups are cheap so they burn/throwaway accounts after one targeted vote. Normally clustering works with more than one common vote between users?