This repository provides a reproducible experimental pipeline for our submission to the AI Co-Scientist Challenge 2026, evaluating the trade-off between data minimization and model performance on the Adult dataset.

We analyze how four minimization strategies (H, C, P, R) impact accuracy and F1-score across multiple models. Strategy-specific LLM choices are intentional: GPT is used for semantically sensitive strategies (P, R), while Copilot is used for the structurally deterministic C strategy to reduce stochastic variance.

A replication of the vertical data minimization baseline (Staab et al., 2023) is included to explicitly contrast feature generalization with our staged feature-removal design space.

Project Structure

• Main entry: experiments/main_results/run_final_experiment.py
• Reference replication (Staab et al., 2023): experiments/replication/staab2023_vdm/run_replication_iterative.py
• Comparison: docs/differences_vs_reference.md
• Paper artifacts: paper/ (manuscript, final figures, tables)
• Legacy code: experiments/main_results/legacy/ (reference only)

All figures and tables reported in the paper are generated from the main entry script.