Pooling Obscures Model-Specific Convergence in AI-Generated English Essays: Implications for Interactive Writing Environments

Research compendium for the paper "Pooling Obscures Model-Specific Convergence in AI-Generated English Essays: Implications for Interactive Writing Environments". This repository contains the analysis code (Jupyter notebooks) and result files for a prompt-controlled, multi-metric study comparing AI-generated essay convergence with human writing. The study uses PERSUADE 2.0 (25,996 human essays) and DAIGT V2 (17,497 AI essays from 16 model sources) corpora across 15 shared prompts. Contents: NB1: Core analysis (pooled and within-model comparison, robustness checks) NB2: Quality-stratified human baseline analysis NB3: Robustness suite (embedding sensitivity, alternative encoder, classification stability) NB4: External replication (legacy, superseded by NB5) NB5: Expanded external triangulation (ICNALE + GPT-4o + Claude Sonnet 4 + Gemini 2.5 Flash) Result files: figures, tables, and reproducibility materials for each notebook All notebooks are designed to run on Google Colab Pro with GPU. Public datasets must be downloaded separately (see Data Availability Statement in the manuscript).