Supplementary Materials for: Agentic Explainable Artificial Intelligence (Agentic XAI) to Explore Better Explanations: A Case Study in Decision Support for Rice Cultivation in Japan

This record holds the supplementary materials for the article "Agentic Explainable Artificial Intelligence (Agentic XAI) to Explore Better Explanations: A Case Study in Decision Support for Rice Cultivation in Japan". In the study, a multimodal large language model (MLLM) received the SHAP plot of a rice yield model with the list of explanatory variables and wrote recommendations for farmers (Round 0). In each of Rounds 1–10 it wrote Python code for additional figures, which the authors ran as written, and revised its recommendations from the resulting figures. Twelve crop scientists and 14 LLM judges scored the recommendations of all 11 rounds on seven criteria. Contents Prompts/: the three prompts of the loop. Prompt 1 asks for agronomic insights and recommendations from the list of variables and the SHAP plot (Round 0). Prompt 2 asks for additional code that generates more figures, gathered into one file (Rounds 1–10). Prompt 3 asks for insights and recommendations updated with the new figures (Rounds 1–10). The file names run in the reverse order: Prompt 1 is Prompt_3.txt, Prompt 2 is Prompt_2.txt and Prompt 3 is Prompt_1.txt. PythonCode.ipynb: the execution history, with the initial SHAP analysis and the code the MLLM wrote in Rounds 1–10. Figures/: the SHAP plot given in Round 0 (Round_0.png) and, for Rounds 1–10, the figures produced by the code the MLLM wrote (one PDF per round). Recommendations/: the recommendations of every round (one PDF per round), which are the texts the evaluators scored. Reproducibility/ (new in version 3): the analysis code that produces every figure and table of the article, its complete outputs, reference copies of those outputs with SHA-256 checksums, and the pinned software environment. Two data files are not included: the evaluation scores (combined_evaluations.xlsx) and the farm data (field_averages.csv). They are available from the corresponding author on reasonable request. Every output of the analysis is included, so each computation can be inspected, but the pipeline cannot be re-run without these two files. The code is under the MIT License (LICENSE-CODE) and everything else under CC BY 4.0 (LICENSE); README.md lists which files are code. Version 3 adds the Reproducibility folder and brings the title and authors in line with the article. The prompts, notebook, figures and recommendations are unchanged from versions 1 and 2.