Computation efficiency and feature inheritance in crash-proof syntax

Frampton & Guttmann (2002) argue that a language design that assumes “crashing derivations” would seem to be less computationally efficient than a design which outputs only convergent derivations. Therefore, they advocate a “crash-proof” syntax which requires constraining all the computational operations. This paper makes a distinction between a fatal crash/strict crash and non-fatal crash/soft crash. I will argue that in a model with Feature Inheritance (Chomsky 2000, 2001 and 2004), a mechanism that supersedes Agree, seemingly non-convergent derivations can be salvaged as long every mechanism in the grammar that’s available is exhausted. I argue, given data from Tamazight Berber, that the three logical possibilities of Feature Inheritance namely DONATE, KEEP, and SHARE, proposed in Ouali (2006, 2008), and whose application is ranked with KEEP applying only if DONATE fails, and SHARE applying only of KEEP fails, despite requiring seemingly different derivations can be accounted for within a less strict crash-proof syntax.