I don't know if that's actually a useful reference class for analysis

I don't know, either, but I feel that some research in this direction would be justified, given the potential payoff.

The other has to do with perceptual thresholds and just-notic[e]able differences.

This might, in fact, be one of the models: the metric being observed hides the "true growth curve". So a useful analysis, assuming it generalizes, would point to a more sensitive metric.