On 27 October we published a number. On 3 November you can check it.
Everything else on this site is a claim about method, and you would have to take our word for it. This one is a claim about the world. It was filed before the result, the arithmetic is open, and the date is not ours to move.
What was filed.
Three probabilities, from twenty thousand simulated elections run on public polling through 26 October, with a fixed seed so anyone re-running it lands in the same place.
| Chamber | Probability of Democratic control | Expected seats |
|---|---|---|
| Senate | 67% | 51.4 of 100 |
| House | 83% | 229 of 435 (80% range 213–246) |
| Governors | 48% | 25.4 of 50 |
67% is not a prediction. It says that in roughly one run in three, it goes the other way. A forecast that cannot lose is not a forecast, and a number published without its odds is a boast. These are the odds.
What is actually under test.
Not the polls. Anyone can average polls. The thing being tested is a turnout term: a measure of how much a given voter has at stake, taken from how lopsided their distances are to the two candidates in a space built from favorability. People with more at stake turn out more.
So the model was run twice. Same polls, same error history, same simulations. The only difference is whether that term is switched on.
Six races carry the term. Everywhere else the published crosstabs were not there to build the space from, so it contributes nothing and the model is running on polls alone.
| Race | Poll average | Turnout term | Model margin |
|---|---|---|---|
| Texas | D+2.7 | +3.8 | D+5.1 |
| Iowa | D+0.3 | +1.3 | R+0.4 |
| Maine | D+1.3 | +1.2 | D+1.5 |
| Ohio | D+3.4 | +1.1 | D+4.1 |
| Michigan | D+3.2 | +0.5 | D+2.7 |
| North Carolina | D+9.9 | +0.4 | D+11.0 |
Texas is the clean test. It carries the largest turnout term of any race on the board — nearly four points, more than the polling margin itself. If Texas lands outside the polls in the direction the term points, that is the single best piece of evidence this model will ever produce. If it lands on the polls, the term is decoration.
Where it is thinnest.
Stated now rather than discovered later. These are the places a careful reader should attack first, and they are the places we would attack first.
Kansas sits at 50%. That is not a finding, it is the absence of one, and it rests on five polls.
Alaska is a projected flip resting on three polls, under ranked choice, which the model treats far more crudely than the voting actually works.
One governor's race in the set rests on a single poll of 91 people. It is carried at its face value because dropping it would be a choice too, and a less visible one.
Iowa has the most undecided voters of any Senate race, at 8.5%. How undecideds break is assumed, not measured, and the assumption moves Iowa more than any other race.
The spaces are assembled from published favorability, not from Galileo surveys. They are a reconstruction from what pollsters released, not an instrument we fielded.
The turnout term has never been tested against an election. This is the first one. That is the point of publishing it, and it is also the reason to hold the result loosely either way — one election is one observation.
What would prove this wrong.
Written before the result, so it cannot be rewritten after one.
The term is adding nothing
If the six races carrying the turnout term land no closer to the model than to the plain poll average, the term bought precision that was not there. The comparison is already published above; nothing has to be taken on trust.
The term is pointing the wrong way
If Texas, Ohio and Iowa all land on the other side of the polls from where the term pushed them, that is worse than noise. It is a sign the stake measure is inverted, and the whole approach needs rebuilding rather than tuning.
The error model is too confident
If the outcome falls outside the published range, the spread taken from past races was too narrow. The model would be wrong in the most ordinary way a model can be wrong: certain before it had earned it.
And what would not prove anything
A 67% call that misses is not, on its own, a failed model — one run in three is supposed to miss. Anyone treating a single miss as a verdict is doing the thing this whole page exists to argue against.
Why this sits on a marketing measurement site.
Because a measurement firm should be willing to be measured, and almost none of them publish anything that can come back to bite. The category's promises are overused and underbelieved, and nobody fixes that by promising harder.
An election is the only subject where the answer arrives on a published date, for everyone, at the same time. The same turnout term asks a commercial question: not who likes you, but who has enough at stake to act. That question is usually answered months later and quietly, if at all. Here it gets answered in public, on the clock.
Geodesic does not take political work. This was built to test the method, and it is here so the test is checkable.
What it was built from.
- Poll positions: every public poll in each race since 12 September, through 26 October, with any trend damped to a quarter of its slope. The damping was chosen because it beat the raw average across 185 past races — 3.93 against 4.09 points of average miss.
- Error structure: 329 Senate and governor races, 2010 through 2022. Each year's miss splits into a shift shared by every race that year (spread 2.95 points) and a remainder specific to each race (4.45 points). Because part of the miss is shared, races move together — which is why control odds are not the product of independent coin flips. Polling archive
- A standing correction of 0.93 points is applied, the average amount by which polls have overstated Democrats in competitive midterm races.
- Every change made after reading the pollsters' own post-mortems is about who turns up, not how people answer — late deciders, nonresponse, new voters, uneven turnout. None of the reviews found shy voters or herding. AAPOR 2024, 2020, 2016
- House seats are placed at the position of their published rating rather than modelled individually. There is no district-level turnout geometry, so the term does not touch the House at all.
- The mass term comes from seven synthetic partisan readers trained on 37,554 news articles, measured by how far each one moves when fed the other side's coverage. Article corpus
Probabilities are model output from twenty thousand simulations with a fixed seed. Differences of a point either way are simulation noise, not findings. Filed 27 October 2026 and left unedited.