Two vote rules selected 89–91% shared weight among top-decile cross-party House pairs in 2019, 2021 and 2023

We tested how the training vote set changes cross-party pair selection in the U.S. House. We trained on 2019, 2021 and 2023 and tested on the next year. One rule ranked Democratic and Republican pairs by agreement on all votes. The other used only votes where the party majorities opposed each other. Each candidate pair cast yes or no together on at least 80% of both training vote sets. At a top-10% budget, the rules shared 90.60%, 89.20% and 89.40% of selection weight. Opposed-only training raised later agreement on opposed votes by 0.632, 0.319 and 0.138 percentage points. The 95% date bootstrap ranges were 0.341 to 0.890, 0.161 to 0.499, and -0.097 to 0.352 points. Agreement on matching-majority votes fell by 0.88 to 1.29 points. All 27 attendance and budget checks had positive opposed-vote point changes. Equal date weights reduced the 2024 gain to 0.010 points. These are three periods at one institution. Agreement does not establish cooperation or a special pair relationship. The package preserves source hashes, code, pair scores, date totals, all result scenarios, bootstrap draws, figures and independent checks. Raw source files remain external under provider terms.

roll-call votes · cross-party agreement · House of Representatives · measurement audit · temporal validation