Replication Package -- Code-to-Prompt-to-Code: An Empirical Study of Prompt Recovery, Adaptation, and Structural Drift

This replication package accompanies the paper “Code-to-Prompt-to-Code: An Empirical Study of Prompt Recovery, Adaptation, and Structural Drift.” It contains everything needed to reproduce the study’s results. A Jupyter notebook reproduces the main tables, the statistical tests, and the RQ1 and RQ2 plots. The analysis/ folder contains scripts for the additional analyses (effect of the repair loop, prompt-level measures, change coverage of adapted prompts with bootstrap confidence intervals, a data-leakage check based on git blame, and repository statistics) and for the final versions of the RQ2 figures. Two further scripts recompute the CodeSage similarity scores. The data/ folder contains the 200 sampled real-world Python functions, 8,400 prompt-recovery records (7 information levels, 2 models, 3 repetitions), 2,400 prompt-adaptation records (400 modifications across 6 modification types, 2 prompt-adapting models, 3 repetitions), and precomputed CodeSage similarity scores for both research questions. The figures/ folder holds the generated plots. The README maps each table and figure in the paper to the notebook cell or script that produces it.