something like that would be interesting. i just get the sense that doing it metadata blind would result in something that learns a somewhat brittle short-time covariance-like structure that doesn't use all available information to achieve best possible generalization and that in the end the benchmark numbers start to reflect some luck with that for predicting non-stationary behavior combined with a lot coming from a pretty good synthesizer that matches previous patterns.

that being the case, it seems the next jump in performance would come from incorporating both metadata and metadata enriched causality and maybe that next jump in performance would be the most interesting jump from a practical system that is useful perspective. (it's more valuable for a system to predict outlier events than it is for it to do an excellent job at synthesizing ordinary behavior)