Semantic leeway

A general methodological question: what criteria do you think should
be used to distinguish a probable semantic change from an improbable
one? Is it possible that comparative linguistics has not yet developed
any such criteria?