This is a report on what the 1000 → 700 validator set size reduction on Kusama did to decentralization metrics of the network, and a prediction on what further reductions would do (computed at era 10166).
The prediction uses a static nomination model (that is, assume no nominators change their nomination), and only varies the validator set size. All data is sourced on-chain (RC, AH, People chain for identities).
The very nature of decentralization metrics requires some clustering heuristics when dealing with "anonymous" validators that might share the same operator underneath.
The operator clustering heuristic model is selectable in the "HUD" below.
I recommend studying the "levels explained" dropout to understand the assumptions going into the clustering heuristics, so that you can assess for yourself how much weight to give them.
To slightly bias though: L2 probably has some false positives (clustering too eagerly), while probably all of the models still have some false negatives (operators that couldn't be distinguished from anonymous singletons.).
So I personally consider somewhere between L1 and L2 to be a safe bet.
What I don't consider a safe bet is viewing this report on a phone - I avoid that shit ;)