Grammatical and Lexical Evidence under Model Damage: A Paired Study of Hebrew, Arabic and English

Does grammatical information remain usable when lexical information deteriorates in a language model? We study this question with paired inflection selection and context retrieval in Hebrew, Arabic and English, using Qwen3-32B and Gemma-3-27B-it. A frozen experiment comprises 146 conditions and 84,096 item rows under Gaussian weight noise and random feed-forward-neuron ablation. All ten adjusted primary intervals include zero, and the two checkpoints disagree on the Gaussian language contrasts. A subsequent source audit finds spelling variants among inflection alternatives and normalization that removes Arabic orthographic distinctions. We then test whether a shared gain and response-label bias can predict damaged choices from clean scores. On held-out lexical/source components, this account improves distribution prediction but does not recover important checkpoint-specific task differences. Its evidence MAE worsens while pooled evidence RMSE improves. These findings do not establish a general language-family advantage or a model counterpart of a clinical dissociation. They show why paired task differences must be distinguished from selective preservation of linguistic information, and motivate a common-candidate experiment with independently varied grammatical and lexical cues.