Every place has twins. Find them.

Every place in a tier gets a fingerprint: a vector of standardized scores across four groups. Trajectory is the presidential margin at each cycle from 2000 through 2024 plus the 12-year trend and elasticity — level and movement, the same derivations the map and table use. Demographics is median household income, median age, college share, population (log scale), and the four race/ethnicity shares. Ancestry is the tier's twelve most-reported self-reported ancestries. Religion is Catholic, Evangelical, and unaffiliated adherence shares.

Each dimension is expressed as a z-score within its own tier — how many standard deviations a place sits from the tier norm — so counties are compared against the spread of counties, metros against metros, and filtering or re-weighting never rescales anything. The twin score (0–100) is the weighted cosine between two fingerprints, with your slider weights split evenly inside each group. Suppressed or unmeasured values sit at the tier mean, so missing data can neither attract nor repel; an ancestry a place doesn't report counts as zero, since ACS ancestry is multiple-response and unreported shares are negligible.

The chips are computed, not written: “twins on” ranks the dimensions where both places sit on the same side of the tier norm, distinctively and close together; “differs on” is the single largest standardized gap of at least one full standard deviation. Deterministic arithmetic over the baked atlas data — no model, no inference. Margins are the county-aggregated presidential record (MIT Election Data + Science Lab and precinct sources, 2000–2024); income, age, education, population, and ancestry are ACS 2024 five-year estimates; race/ethnicity is ACS 2023 B03002; religion is the 2020 U.S. Religion Census (ASARB). The same fingerprints power the twins module on every place page and the find_similar_places tool on the hosted MCP server.

Where a tier's baked record doesn't carry a dimension, that dimension is dropped for the whole tier rather than guessed — never rescaled, never imputed. Cities are the visible case: their election record begins in 2008 (so the 2000 and 2004 cycles are absent from the city fingerprint), and the city data set carries the Latino share and the unaffiliated share but not the other race or religion breakdowns. District fingerprints likewise start at 2008. Every tier's exact dimension list ships in its fingerprint file.