Abstract
Distributional similarity between word forms does not establish that a model distinguishes morphological form from grammatical function. This study investigates that distinction in Kazakh legal vocabulary, treating Word2Vec and FastText as analytical instruments. A corpus of 2,188 documents and 12,327,449 tokens supports a reference set of 200 pairs, eight morphological relations and 4,800 analogies. Retrieval, relation-vector coherence, cross-domain transfer and error diagnostics are combined with a new comparison of identical versus different surface allomorphs. FastText achieves higher overall exact retrieval than Word2Vec (35.21% versus 28.98%), but genitive and derivational results qualify any general superiority claim. Among 195 target pairs admitting both allomorph conditions, target-matched same-allomorph advantages average 5.61 and 2.32 percentage points, respectively. For accusative FastText analogies, an apparent unmatched advantage of 11.93 points becomes −0.68 after target matching, demonstrating lexical-composition confounding. The models preserve similar inter-relation geometry despite different local coherence and error profiles. The contribution is a linguistically constrained account of morphological regularity: shared spelling, lexical retrieval and grammatical interpretation must be evaluated separately. Findings concern one specialized corpus and a rule-based reference resource, not an exhaustive model of Kazakh grammar.