|Abstract:||This paper presents an optimality-theoretic approach to dealing with variation as well as systematicity in aphasic output patterns. Fluent and nonfluent speakers with aphasia after stroke showed different patterns of segment deletion in onset and coda cluster reduction. While onsets were consistently reduced to their nonsonorant components in both groups of speakers and nonfluent speakers showed the same pattern in coda cluster reductions, the fluent patients had a 50-50 distribution of sonorant versus nonsonorant deletions in codas. With parallels to psycholinguistic modeling, it is argued that these data are best accounted for by assuming two different levels of computation, with some but not all markedness constraints being level-specific. At the deeper, phonological level of computation, impaired in the fluent speakers, constraints on both segmental and syllable output wellformedness gain prominence, due to a lowering of faithfullness constraints. At the 'cognitive phonetic' level of impairment of the nonfluent speakers, only segmental markedness constraints are usually active, and therefore become dominant if faithfullness constraints are lowered. The paper argues specifically against a reranking among the family of markedness constraints, as a representation of aphasic deficits, primarily because allowing for this would lead to boundless overgeneration and would fail to capture the systematicity and the general adherence to mother-tongue phonotactics observed in aphasic speech.