What troubles me about time series foundation models is that it's impossible to get a quick answer of whether these models work out of sample for many domains which don't generate new data super quickly. So, yeah, they gave an example of power price forecasting (my domain), so I would love to built a nice model with fancy covariates and all and backtest over the last couple of years -- but I have no way of knowing whether Mistral would be leaking data form its training into my backtest, no?