I think their point is something like what's discussed here: https://news.ycombinator.com/item?id=22915837 https://news.ycombinator.com/item?id=22915837 I'd d

I think their point is something like what's discussed here: https://news.ycombinator.com/item?id=22915837 https://news.ycombinator.com/item?id=22915837 I'd distill it as: 1. Every model should be presented alongside a discussion of whether the fitting process is well-posed[1] 2. If the fitting process is ill-posed, the model cannot be used to guide policy 3. If the fitting process is well-posed, but only via Bayesian regularization of some kind, before being used to guide policy it should be checked for sensitivity to choice of regularization, and the regularization scrutinized in exhaustive detail 4. Compositions of models must be checked for well-posedness https://en.wikipedia.org/wiki/Well-posed_problem https://en.wikipedia.org/wiki/Well-posed_problem