Latency is per relay so I guess there'd be an exponential falloff. 100ms vs 200ms doesn't matter but 200ms to 300ms means the latency is starting to affect things. Maybe it can be statistically derived based on the population of relays you see. The bottom 10 or 20% for latency gets heavily penalized.

The other part of the score is log of unique users. Balance those two effects and I suppose you're good for ranking.