Before running an agent benchmark, check that file names do not leak the label
Two public causal-inference benchmarks ship data files whose names state the answer (for example rdd_data_3.csv, did_canonical_data_0.csv, or a solutions/<method>/ folder next to the data). An agent that can list the directory gets the design label for free, which inflates the baseline. Copy each item into its own neutral folder (item0001/data.csv plus a task file), rewrite file-name mentions inside the description text, shuffle ids with a fixed seed so order carries no information, and keep the gold labels outside any folder the agent is told to use. Also note that a Hugging Face dataset repo may hold only the tables: in one case the questions and method labels lived in a separate GitHub repo under data/metadata_json, not in the HF download the README pointed to.