This paper reports on two experiments that back the graded salience hypothesis (Giora 1997, 1999, 2003; Peleg et al. 2001, 2004), which postulates that early language processing contains distinct linguistic and contextual mechanisms that do not interact, yet run parallel. The results criticize direct access models, which claim that context can selectively activate the suitable meaning, irrespectively of saliency (see Vu and Paul 1998; Vu et al. 2000). While contextual processes make up an integrative, top-down mechanism that benefits from linguistic and extra-linguistic information, the linguistic mechanism is modular (Fodor 1983). Using Vu et al.'s (2000) materials, Experiment 1 shows that the sentential position of a target word (initial vs. Final) is crucial for the operation of the global, predictive mechanism, whose effects, accumulated in prior discourse, mask lexical effects in final, but not in initial position. Experiment 2 shows that even in a sentential position that favors contextual effects (i.e. sentence final position, see Peleg et al. 2001, 2004), lexical access is not affected: Salient meanings are activated upon encounter of the lexical stimulus, regardless of contextual information to the contrary. Taken together, these findings argue against direct access models, which suggest that context can selectively activate the appropriate meaning, regardless of salience (see Vu and Paul 1998; Vu et al. 2000).