Stylistic Drift as a Model Fingerprint: Measuring LLM Stylistic Signatures and Their Prompt Sensitivity in French Literary Rewrites

We study stylistic drift in LLM-rewritten French literary texts using function-word frequency vectors. For each original–rewrite pair, we define shift as the cosine distance between their L2-normalised 57-dimensional style vectors, and compare four commercial models (GPT-4o, Claude Sonnet 4.6, Mistral Small, Gemini 2.5 Flash) across 80 texts from Zola and Maupassant and two rewriting prompts. Three findings emerge. First, Gemini produces significantly larger stylistic drift than the other models (Δ = 0.23, p < 0.001 after Bonferroni correction); GPT-4 and Mistral are statistically indistinguishable (p = 1.0). Second, a comparison of six feature sets shows all methods plateau between 33% and 44% accuracy on 4-class model attribution, confirming that the signal is weak at the individual-text level. Third, inter-prompt robustness is weak across all models: per-text Pearson correlations between P1 and P2 shift series are non-significant for Gemini and Claude (r < 0.07, p > 0.6) and borderline significant for GPT-4 and Mistral (r ≈ 0.28, bootstrap 95% CI barely excluding zero). Model rankings are preserved at the aggregate level, but individual-text shift is not predictable across prompts, suggesting that stylistic fingerprints reflect distributional tendencies of models rather than stable per-document properties. Code and data: https://github.com/riadmaouchi/llm-style-fingerprints