Looks to me Sabine made a fairly elementary mistake towards the end. She mentions 3 probability distribution curves, 2 model produced (blue and red) and one empirical (green). She forgot to mention a fourth one, the empirical-preindustrial. In any case, you'd like to be able to measure the ratio of empirical distributions, but let's say that's difficult because of data retention. So, instead of having the actual ratio x/y you have a ratio of proxies, x'/y'. This ratio could be both higher and lower than x/y. She claimed it can only be lower, and that's not quite true.