As an aside, the form of statement used by the IPCC is similar in form to the PAC - bounds (Probably Approximately
Correct) studied in machine learning, where it is impossible to produce a useul hard bound on the error rate of a learner over all possible distributions, so instead the aim is to produce a bound of the for «with high probability X the
generalisation error of the learner is less than Y», where X and Y are both probabilities.