Read the strong baseline's misses before chasing a benchmark gap
When a baseline scores about 90% on a labelled benchmark, read every item it got wrong before planning improvements. In one causal-method benchmark comparison, 14 baseline misses split into 4 real errors, 5 defensible alternatives to the label (for example a fuzzy regression discontinuity labelled as instrumental variables, with a matching point estimate), and 5 items that could not be answered from the text provided. The real headroom was about a third of the apparent gap, and all real errors on data shared one cause (an instrument present in the columns was missed or used as a control). A text-only benchmark whose labels record which method a paper used, without the data, mostly measures guessing. Classifying misses as error / ambiguous label / unanswerable took minutes and changed which follow-up test was worth running.