This article approaches from a new angle the question of the extent to which predictable information is stored in the lexicon. By examining the ways in which morphological phenomena can be sensitive to prosodic structure, I argue that some—but not all—predictable information is stored in lexical entries. Detailed analysis of a fragment of the Armenian phonological system, focusing on the behavior of consonants at morpheme edges, supports a more abstract view of phonological representations (containing syllables, appendices, and unparsed segments) than can be inferred from phonetic facts alone, contra Ohala and Kawasaki-Fukumori (1997), Steriade (1999), Scheer (2002), and others. The Armenian facts furthermore indicate that attempts to abandon underlying representations (Flemming 1995, Burzio 1996) are misguided and that we must also retreat from the excessively abstract under specification approaches advocated by most phonologists.

Since Clements (1985) introduced feature geometry, four major innovations have been proposed: Unified Feature Theory, Vowel-Place Theory, Strict Locality, and Partial Spreading. We set out the problems that each innovation encounters and propose a new model of feature geometry and feature spreading that is not subject to these problems. Of the four innovations, the new model-Revised Articulator Theory (RAT)-keeps Partial Spreading, but rejects the rest. RAT also introduces a new type of unary feature-one for each articulator-to indicate that the articulator is the designated articulator of the segment.

Phonologists have traditionally assumed that the unmarked laryngeal state of fricatives is to be unaspirated ([−spread glottis]). However, the data analyzed here, which are drawn from Armenian, Greek, Pali, Sanskrit, Spanish, and Thai, suggest that in their unmarked state voiceless fricatives are in fact [+spread glottis], whereas voiced fricatives are [−spread glottis].