Discovery Agent Reproducibility Package: Experimental Data, Logs, and Supplementary Materials

This reproducibility package accompanies the Discovery Agent paper and contains the raw experimental data, execution logs, and supplementary documentation. It comprises: (1) a direct comparison with CASCADE on its own tool-grounded benchmark under the zero-setup protocol; (2) a comparison with SciAgents on 20 open-ended, hypothesis-driven materials-discovery tasks (seed=42); (3) a component-level ablation study of 19 configurations evaluated across 3 domains (battery, catalysis, electronics) x 5 independent runs; (4) the main benchmark evaluation (9 regular questions plus 3 ground-truth questions, 5 runs each); and (5) an appendix with algorithms/pseudocode, mathematical definitions, hyperparameters, and experiment details. All personal paths have been anonymized. Some runs were collected in separate batches, so per-question run timestamps may be non-contiguous; this reflects the actual execution history and does not affect any reported metric. Licensed under CC BY 4.0.