In formal analyses of productive systems of stem gradation, it is commonly assumed (1) that each alternating stem possesses a single, basic grade form from which its other grade forms may be derived, and (ii) that all basic forms belong to the same grade; certain languages, however, present predictively indeterminate systems of gradation, for which these assumptions are not apparently valid. Two approaches to the analysis of such systems are discussed here: the first approach, exemplified by Anderson’s (1976) Analysis of initial consonant gradation in Fula, allows both assumptions (i) and (ii) to be maintained at the expense of an absolutely neutralized phonological distinction; the second approach, embodied in the traditional Pãninean analysis of Sanskrit vowel gradation, allows assumption (i) to be maintained at the expense of assumption (ii). It is argued that the Pãninean approach is superior, since it is more generally applicable and does not require the postulation of any phonetically unrealized phonological distinctions, relying instead on distinctions in grade among underlying stem forms.