The conclusion urged in Mr Salmon's recent article ([1], pp. 252-261) is so remarkable that it may be worth recording some difficulties. He claims to rescue Reichenbach's notorious vindication of induction. This is essentially concerned with estimating long-run frequencies. By an estimator let us mean any rule for making estimates appropriate to various bodies of information. Reichenbach thought an estimator is sensible only if it is convergent, that is, roughly speaking, only if its estimates tend to approach the truth as more and more data is collected. Mr Salmon observes that an estimator should, in addition, be independent of the language in which the information is expressed. He contends that this criterion of linguistic invariance, when added to the rest of Reichenbach's train of thought, suffices to justify the “straight rule” for estimation: if m of the n observed A's are B, estimate the long-run frequency of B among A's as m/n.