I fit a very simple Bayesian model in Stan to infer the direct effect of screen brightness on power draw

Thanks to the simple form of the exponential smoothing applied here, I can invert the smoothing to recover the latent true power readings and get a much more realistic looking profile:

In the post, I show that the observed readings are highly consistent with an exponential smoothing process being applied to latent "true" readings. By assuming that the latent power profile is a step function, I find a smoothing parameter that maps the idealized latent power to the observations.

By taking readings from the battery while applying a step-function artificial load, I observed that the reported power decayed toward near-constant values in a manner that appeared exponential

you win this round, Octavia Butler

Whelp, there it is - for now, at least! Interesting direct call out of Trump for the hopes of bringing it back.