This module implements a token‑level statistical baseline for AI vs human code detection. It relies solely on distributional properties of code tokens, such as entropy, burstiness, and predictability, without using structural or stylistic features. The goal is to evaluate whether LLM‑generated code exhibits detectable statistical signatures prior to incorporating higher‑level program structure and stylometry.

The statistical baseline achieves 58.6% validation accuracy, indicating that token‑level regularities alone are insufficient for robust authorship discrimination, motivating the inclusion of structural and stylistic features.”