Detecting and Localizing Generated Segments in French-Language Hybrid Texts through Within-Document Variance of Linguistic Features

Research protocol / preprint. This manuscript presents a proposed methodology; no experiments, performance measurements, or improvements are claimed. Results remain to be completed after experimentation. Document-level classification is insufficient when human-written and AI-generated passages coexist within the same text. This paper proposes a framework for detecting and localizing generated spans in French-language hybrid documents using within-document linguistic variance and local feature changes. The hypothesis is that authorship transitions may coincide with measurable changes in lexical, syntactic, stylistic, or statistical profiles. The planned corpus covers human texts, generated texts, human texts partly modified by AI, and generated texts partly rewritten by humans. Normalized segment features are combined with adjacent-segment differences and boundary scores. Ablation studies will compare absolute features, variation features, and their combination while keeping source-document families in separate data splits. Evaluation will address classification, localization, and generalization to unseen domains and generators. This deposit contains the English-language manuscript (PDF) and editable LaTeX source with a README. The supplied PDF was typeset separately; the LaTeX source has not been compiled in the preparation environment and its pagination may differ.