ADVD-M/hons2 — reverse-engineered prompt

Build me an interactive Streamlit app for a research demo about Honest Random Forests and how robust they are when tabular data changes over time.

I want users to upload a classification dataset, pick the target column, and then compare a normal Random Forest against an Honest Random Forest. The app should split the data for the honest model so one part builds the tree structure and another part estimates the leaf probabilities. Then it should stress test the models by shifting important input features and showing how accuracy drops as the shift gets stronger.

Make the dashboard easy to use, with clear explanations for non experts, charts showing baseline accuracy, accuracy retention, and where the standard model starts failing compared to the honest one. Use Python, pandas, scikit learn, Streamlit, NumPy, and Plotly. Include basic data cleaning for common CSV issues and make sure it can handle typical classification datasets. Look up current docs online if you need to.