Lines: the U.S. House redrawn 10 ways, from gerrymanders to fair maps

The building blocks

Every alternative district is assembled from whole 2020 Census voting districts — 187,560 of them, block groups in California, Hawaii, Oregon and West Virginia — with their 2020 population and voting-age population by race, as packaged by Dave’s Redistricting. Seats follow the 2020 apportionment, the same 435 on the ballot this year.

Partisanship is the 2024 presidential vote, precinct by precinct. In the nine multi-seat states whose 2024 precinct returns are not yet in that package (Arkansas, Connecticut, Idaho, Maine, Michigan, New Jersey, Oklahoma, Oregon and Pennsylvania), each 2020 precinct result is carried forward by its own county’s actual 2020-to-2024 swing.

The rules every map obeys

Every district is contiguous — each precinct reachable from every other through shared borders, water crossings only where islands require them — and within 0.5% of its state’s ideal population. Across all 4,350 drawn districts the largest deviation is 0.50%, and half the states’ maps stay within 0.47%. A build step re-checks both rules for every district independently and refuses to publish a map that fails either.

Whole precincts cannot balance districts to the single person, as enacted maps do by splitting census blocks. Anyone adopting one of these maps would finish that last step at block level.

How a computer draws a map

Each map starts from a random legal map and is improved by two kinds of move. ReCom merges two neighboring districts and splits them again along a random spanning tree, exploring very different maps quickly; runs of ten such moves keep the best map found (“short bursts”). Simulated annealing then moves single precincts across lines to finish the job. Searches also start from the enacted map, and the better result wins.

No goal may buy its result with tentacles: the durable gerrymanders cannot cut more than 1.5 times the precinct borders of a typical no-goal map, the aggressive ones three times as many, and the proportional and competitive maps no more than a typical map’s borders and county splits. The four geography maps use no voting data at all. Nothing was tuned state by state or drawn by hand.

An even year, by default

Every map is drawn for, and shown in, an even year: a national environment of zero, the way FiveThirtyEight’s 2018 Atlas read districts by their lean relative to the nation. Trump carried the 2024 two-party vote by 1.6 points in these precincts, so an even year moves every district 1.6 points toward the Democrats. A map that splits seats fairly should then come out near 218 to 217. The slider moves the national environment, the same House margin our forecast publishes, with presets for the 2024 result and our forecast; each point of environment moves a district about 0.73 points from its 2024 presidential margin, as our forecast does.

A district’s chance comes from its lean through one curve fitted to our own House forecast (edition of 2026-09-19), whose spread is 7.2 points; at the forecast’s environment the curve reproduces its 229 expected Democratic seats on today’s map within a seat (229.8). It is an average relationship, not a race forecast: it knows nothing of candidates, incumbents or local polls — the full forecast for today's map is 231 Democratic seats as of 2026-09-29.

The no-goal test

For each of the 44 multi-seat states, a ReCom chain drew 2,000 maps under the same rules and no goal at all; 23,202 of them, spaced through the chain, form the comparison. Such ensembles are how courts and scholars now judge whether an enacted map is an outlier. They describe what neutral rules tend to produce — they are not a claim that any single neutral map is the fair one.

Two kinds of gerrymander

Durable is the gerrymander a legislature would pass and defend: as many districts as it can where the drawing party leads by at least 12 points in an even national year, seats that survive a 12-point wave against it, on lines no more ragged and no more county-splitting than 1.3 times a typical neutral map’s. Aggressive takes the most seats that hold as long as the party does not lose the national vote by more than 6 points, with no regard for shape, the Voting Rights Act or the courts: 10 of Virginia’s 11 seats for Democrats, 15 of Illinois’ 17, 37 of Texas’ 38 for Republicans. It is a ceiling rather than a map a court would let stand. After the seat count, both maximize the party’s expected seats across national swings of up to 10 points either way, so a losing district is drawn as close as the votes allow. Every state was searched with several different methods and scored by one published scorer; each map is the best any of them found.

Fair maps and geography

Proportional makes each state’s expected seats track its own vote inside a typical map’s shape and county limits (at most 1.2 times its border length and county splits), the map a good commission might draw. Competitive makes as many true toss-ups as it can, districts within a point of even in an even national year, then draws every other district as close as the votes allow, with borders no more than three times as long as a typical neutral map’s. Four maps know nothing about voters: straight lines cuts each state again and again with the shortest line that splits its people in the right ratio; town and country makes each district as alike inside as possible in how densely its people live, so cities, suburbs and the countryside each get their own districts; slices does the opposite, giving every district the same mix of city and country, cut outward from each big metro; county lines splits the fewest counties.

What is new since 2018

FiveThirtyEight’s original Atlas drew its maps by hand on 2010 geography and scored them with the Cook PVI. This edition is drawn by algorithm on 2020 geography, scored with 2024 returns through a curve fitted to our 2026 forecast, lets the reader move the national environment, draws safe-seat and opportunity-seat gerrymanders, runs an arms race by who actually controls each state, adds fair and geography-only maps, measures the efficiency gap and median seat, tests every enacted map against thousands of goal-free maps, and publishes every map as a shapefile, GeoJSON and exact precinct assignment.