Do inputs need to be restricted on a language-specific basis? Classic Optimality Theory claims that they do not: the rich base is filtered by constraints that yield full contrast, complementary distributions or positional neutralisation depending on the ranking. The problem arises when positional neutralisation affects a gappy contrast. In Russian, voicing neutralisation works on all obstruents alike, including non-contrastively voiceless ones – but it creates voiced allophones that are otherwise disallowed. In the popular OT account of positional neutralisation, analysing these cases requires handling voicing twice: once for all segments, then again for gaps. I argue that the solution is to relax the rich base assumption by ruling gaps out at the UR level through morpheme structure constraints (Halle 1959 et seq.).