The CountSenate control · D 18% / R 82%Median seats 49D36 calls locked0 gradedBrier —
sleyor

Issue One · Vol. I — opinion that owes you a score

The Count

Unbiased is not a claim. It is a score.

Every outlet says it is fair. Almost none of them can be checked. A forecast can: we publish a number before the event, and afterwards the world tells everyone whether we were right. This desk keeps that record permanently, including the parts that embarrass us.

Read the desk's opening piece
Now open — 3 November 2026

The midterm board is live: Senate, House and governor control math, every key race rated with a stated probability, and a call on all three chambers the desk will grade itself against on election night.

Balance of power watch
Inputs — Certified 2024 returns

The map, the national board and the race table below run on certified statewide returns from the 2024 presidential election — 155,240,955 votes counted across 51 contests, as canvassed by each state. Source, retrieved 4 August 2026. Nothing on those three panels is modelled.

The forecast layer is modelled and labelled as modelled: the 2026 Senate distribution and the tipping-seat ranking below come from a seeded simulation whose coefficients, sources and limits are printed in full on the balance-of-power page. No polling is in it yet, and the calibration record stays empty until the desk has graded calls of its own.

The boardCertified — tipping point PA
48.3%
49.8%

Share of the national popular vote as certified: 75,019,230 to 77,303,568, an electoral college of 226 against 312. The 270th elector came from PA — that state, not the national number, is where the thing was decided.

Outcome distributionSenate 2026 — modelled
39 Democratic seats — 0.0% of simulations40 Democratic seats — 0.0% of simulations41 Democratic seats — 0.1% of simulations42 Democratic seats — 0.2% of simulations43 Democratic seats — 0.5% of simulations44 Democratic seats — 1.5% of simulations45 Democratic seats — 3.4% of simulations46 Democratic seats — 7.7% of simulations47 Democratic seats — 13.4% of simulations48 Democratic seats — 19.6% of simulations49 Democratic seats — 20.6% of simulations50 Democratic seats — 15.2% of simulations51 Democratic seats — 9.0% of simulations52 Democratic seats — 4.9% of simulations53 Democratic seats — 2.2% of simulations54 Democratic seats — 1.1% of simulations55 Democratic seats — 0.4% of simulations56 Democratic seats — 0.2% of simulations57 Democratic seats — 0.1% of simulations58 Democratic seats — 0.0% of simulations61 Democratic seats — 0.0% of simulations51 seats

Democratic seat totals across every simulation of the 2026 Senate. The desk publishes the whole spread rather than one number, because a 52% call with a fat tail is a different object from a 52% coin flip.

18%
82%
Democratic majority 18%Republican control 82%

Median outcome: 49 Democratic seats, with eight runs in ten landing between 46 and 51. Fifty–fifty is not a tie — the vice-president breaks it for the Republicans, so 51 is the number that matters and the chart is drawn against 51.

Model senate-2026.1Frozen 2026-08-0520,000 simulationsSeed 20261103Holdovers 34D · 31R36 calls locked · 0 graded
The model, coefficient by coefficient
Tipping seatSenate 2026
  1. Ohio13.6%
  2. Florida13.0%
  3. Alaska10.5%
  4. Iowa10.2%
  5. Texas10.1%
  6. Kansas6.5%
  7. North Carolina6.1%
  8. Michigan4.9%

How often each seat delivers the 51st vote. This, not the national environment, is where control of the chamber is actually settled.

The mapEqual tiles — land does not vote
AK3ME4VT3NH4WA12ID4MT4ND3MN10IL19WI10MI15NY28RI4MA11OR8NV6WY3SD3IA6IN11OH17PA19NJ14CT7CA54UT6CO10NE5MO10KY8WV4VA13MD10DE3AZ11NM5KS6AR6TN11NC16SC9DC3OK7LA8MS6AL9GA16HI4TX40FL30

Tap a state for its certified margin

Every state is the same square, labelled with its electoral votes. Area maps hand empty land the loudest voice in the picture, which is the oldest lie in election graphics, and both sides have used it.

Colour depth is confidence, not enthusiasm. A pale tile means the desk does not know, and a grey tile means the interval crosses zero — a real toss-up, not a race we have quietly leaned.

LeftToss-upRight
To the county, to the districtCertified — 3,160 counties · 435 seats
435 districts — certified

Districts are the unit that actually elects somebody. Tap one for its presidential margin and the census profile of the people inside the line.

Loading district returns
Margin
R+50evenL+50
Land does not vote. A county map shades acreage, so a thousand square miles of near-empty ground reads louder than a city that cast ten times the ballots. Keep the tile cartogram above beside it, and read the two together.

RETURNS: 2024 presidential results by congressional district
GEOMETRY: US Census 119th Congress cartographic boundaries
GAP: NY-21 has no published district-level presidential return. It is drawn blank rather than estimated.

2,685 of 3,160 counties went one way and cast barely half the ballots — and 56 million people are painted the colour of their neighbours. The desk's piece on why the county map lies, worked entirely off this file.

Read it
The racesSorted by closeness
Left 226 EVToss-up 0 EVRight 312 EV
StateEVMarginVotes castResult
Wisconsin10R+0.93423k
Michigan15R+1.45664k
Pennsylvania19R+1.77059k
Georgia16R+2.25251k
New Hampshire4L+2.8826k
Nevada6R+3.11485k
North Carolina16R+3.25699k
Minnesota10L+4.23254k
Arizona11R+5.53390k
Virginia13L+5.84506k
New Jersey14L+5.94273k
New Mexico5L+6.0923k
Maine4L+6.9831k
Illinois19L+10.95633k
Colorado10L+11.03193k
Ohio17R+11.25768k
New York28L+12.68262k
Florida30R+13.110894k

Eighteen closest races shown

The record

When we said seventy percent, did it happen about seventy percent of the time?

no graded calls yetstated →

Stated probability vs. observed rate

Calls published
35

Senate seat probabilities standing on the record right now.

Calls graded
0

First grading 3 November 2026. Every call, win or lose.

Brier score
—

Lower is better. A coin flip scores 0.250.

Worst miss
—

Published in full, with the post-mortem.

A forecaster with no published scorecard is a pundit with better graphics. This board stays up in bad cycles, and the worst miss of each season gets its own piece.

See every call, locked and dated
The method
The model is published before the result

Weights, inputs and the code path are posted when the forecast is, not after. A method released after the fact is a story about a number, not a method.

Every number carries its source

The board runs on certified state canvass returns, cited on the page with the date they were retrieved. Where a figure is modelled rather than counted, it is labelled as modelled in the same size type.

We grade ourselves in public, forever

Every stated probability is logged and scored. The calibration chart and the Brier score stay on this page whether they flatter us or not, and the worst miss of each cycle gets its own write-up.

No unsourced number, ever

If an input cannot be traced by a reader, it does not enter the model. Where the data is thin the desk says the data is thin, in the sentence rather than the footnote.

Commentary never adjusts the model

The desk writes about what the numbers do. It does not nudge the numbers toward what it would rather write about. Manual overrides, if ever used, are logged as overrides on the page.

What would change our mind
  1. 01

    A three-point move in the two-party margin across the industrial north that holds for two weeks, rather than one good survey.

  2. 02

    Turnout composition in early voting diverging from the modelled electorate by more than the historical error band.

  3. 03

    A systematic house-effect correction that changes sign — the clearest signal that the pollster adjustments are the thing that is wrong.

Read the Standard