It is unclear whether word stress in a language is stored as part of the word or whether it is generated by a rule. We test the generativist hypothesis of lexical storage stating that only unpredictable stress is stored in long-term memory against the contrasting usage-based approach assuming that all phonetic information regardless of its (un)predictability is stored in the mental lexicon together with the word. In a correctness judgment task involving correctly and incorrectly stressed penults and antepenults, we found that incorrectly stressed penults do not evoke an N400 effect, whereas incorrectly stressed antepenults do: there is increased negativity with a peak latency around 350–600 ms from word onset. Only changes to words with exceptional stress cause lexical inhibition, hence exceptional but not default stress markers are stored in the lexicon. Additionally, differences in processing patterns between the N400 and the late positivity component window point to an integration of two stages of word processing: pre-lexical stress recognition and stress-to-meaning matching. The results of the study support the view that stress should be understood as abstract phonological information.