Abstract
Universal and multilingual large language models cannot correctly simplify Kazakh texts in accordance with internal linguistic patterns. Using LLM without a linguistically sound simplification model leads to uncontrolled generation and increases the risk of lexical and grammatical errors. Therefore, there is a need to create formalized linguistic patterns for Kazakh text simplification, which will facilitate the creation of automatic simplification systems within LLM. This can be achieved through the analysis of manually adapted materials. A corpus of 20 Kazakh texts was annotated by the authors for the analysis, each presented in three versions such as the original (O) and two successive adaptation levels (A1 and A2). At the A1 stage, syntactic transformations predominate, while A2 demonstrates lexical stabilization. The text size is comparable at all levels, which ensures control and reliability of the analysis. The process is formalized through a two-level tagging system: lexical (LS) and syntactic simplification (SS) that records each operation. Frequency analysis revealed a limited core of dominant techniques, supplemented by peripheral layers, forming structured polyoperational architecture. Transitions from A1 to A2 demonstrate a process shift toward lexical changes while maintaining discursive coherence. The results emphasize the importance of linguistic expertise, operation typologies, and manual annotation for the interpretability and quality of systems. Hybrid approaches with humans in the loop improve the accuracy and reliability of ATS.