Replication artifacts for "Can Global XAI Methods Reveal Injected Behaviours in LLMs? SHAP vs Rule Extraction vs RuleSHAP"

This record contains the reproducibility artifacts for the paper: Francesco Sovrano. 2026. “Can Global XAI Methods Reveal Injected Behaviours in LLMs? SHAP vs Rule Extraction vs RuleSHAP.” In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '26). The uploaded archives contain precomputed experiment artifacts used by the RuleSHAP replication package. They are intended to be placed in the root directory of the code repository and unzipped there before running the reproduction scripts. Files included: abstract_model_io.zip: precomputed topic abstractions, model inputs and outputs, generated explanations, output metrics, and cached SHAP-related intermediate data. cache.zip: cached LLM/API responses used by the experiment pipeline to reduce repeated API calls and support reproducibility. xai_analyses_results.zip: extracted RuleSHAP and baseline rules, SHAP outputs, evaluation summaries, plots, and related analysis outputs. Reproducibility usage: Clone or unpack the RuleSHAP source-code repository. Copy the three ZIP files into the repository root. Unzip them in place so that the repository root contains the directories abstract_model_io/, cache/, and xai_analyses_results/. Install the Python dependencies following the repository README. Run the provided experiment or evaluation scripts. The source code is maintained separately in the RuleSHAP repository. These archives provide the large precomputed artifacts needed to reproduce and inspect the reported experiments without regenerating all model calls from scratch.