Reliable detection of machine-generated codehas become increasingly important for aca-demic integrity and software quality as codegeneration is largely being undertaken by largelanguage models. This paper presents our ap-proach to SemEval-2026 Task 13, Subtask A:detecting machine-generated code in a binaryclassification setting, where we propose anensemble approach combining TF-IDF lexi-cal representations with 23 hand-crafted sty-lometric features for binary classification ofAI-generated code. Our system aims to addressthe challenge of cross-language generalizationby extracting language-agnostic patterns andcombining them with TF-IDF. While we ob-served that transformer-based models (Code-BERT, UniXcoder) noticeably underperformedunder distribution shift, our analysis revealedthat AI-generated code exhibits distinct sty-lometric patterns and our TF-IDF ensembleachieved 0.5175 Macro F1 on the task submis-sion.