Panjabi is an Indo-Aryan language spoken primarily in the Punjab province of Pakistan and the Indian state of Punjab. Long vowels contrast in nasality in the language, but this contrast is neutralized before nasal consonants: [˜V˜VN], *[VVN]. This paper presents air pressure data from Pakistani Panjabi showing that predictably nasal vowels before nasal consonants in the language, [˜V˜VN], are statistically equivalent in nasality to contrastive nasal vowels, /˜V ˜V/ (Exp. 1). However, predictable [˜V˜VN] and contrastive nasal /˜V ˜V/ vowels behave differently in their interaction with a nasal harmony process, in that only contrastive nasal /˜V˜V/ vowels trigger nasal harmony (Exp. 2). This difference in phonological behavior can be straightforwardly analyzed by assuming (i) that pre-nasal vowels are underlyingly oral, even in morphemes which never show alternations in nasality, and (ii) that only underlyingly nasal vowels trigger harmony. These findings support a view of phonology in which abstract underlying representations (URs) can be completely covert, never surfacing in any allomorph without some kind of symbolic alteration. Surface-oriented approaches to phonology which reject covert URs — including usage-based models relying exclusively on observed exemplars to form lexical representations —have difficulty accounting for these facts.