Balance of Power · an Intuitionist-Fin experiment

What do the polls tell us?
And how often are we wrong?

We track all 435 House seats and all 100 Senate seats, with a closer look at six Senate races and twenty House races. Some seats have enough information for an estimate; others remain unresolved. We use polls and past voting to make test forecasts. A separate AI briefing reviews a few races each day.

The chamber experiment shows possible seat totals under stated assumptions. It does not yet publish a chance of controlling Congress. We do not use this election work to choose stocks.

Poll collection

Up to date.

Last collected: 2026-10-23T23:45:33.779692+00:00. A fresh collection does not mean every poll is new.

Seat estimates

Current scenarios available.

Saved 2026-10-23T23:46:23.516478+00:00. These are possible seat ranges, not chances of controlling Congress.

AI briefing

Latest briefing completed.

Latest recorded date: 2026-10-23. The briefing is separate from the numerical forecasts. See briefing status.

Under review after the October 4 audit. The call below used v2, which missed some candidates’ prior experience. It is retained as a dated record, not a current recommendation. The corrected v3 model still has wide gaps and historical directional errors.

Earlier Codex judgment · October 4, 2026 · low confidence

Earlier call, now under review: a divided Congress

I would pick a narrow Democratic House majority and a Republican Senate majority. Both calls could be wrong; the House call is especially fragile.

House

Lean D

A narrow Democratic majority is my working call.

Senate

Lean R

Republicans are my pick to retain a majority.

Confidence

Low

A judgment to revisit, not a tested chance of winning.

Why make those calls—and what could change them?

The October 4 no-extra-shift model run leaves 21 House and eight Senate seats unresolved. As a rough starting point only, suppose those seats keep the party listed for their incumbent. Adding those assumed holds to the model’s average gives about 219.5 Democratic and 215.5 Republican House seats; in the Senate, about 46.2 Democrats, 51.8 Republicans and two independents. Fractional seats are averages, not possible election results. These are illustrative totals, not my exact seat predictions or the model’s official completion of missing races.

That assumption is weak: unresolved does not mean safe. Candidate parties and boundaries can change. California’s 6th district includes an independent candidate despite the inventory’s Republican incumbent label. Several Senate races involve independent candidates or ballot rules our model cannot yet handle. The numerical model therefore leaves these seats unresolved.

A small Republican improvement across House races could reverse my House call. Republican losses in more of the unresolved Senate contests could reverse my Senate call. I would revisit both with additional verified local polls, resolved ballots and boundaries, or a broad national shift supported by several polling firms. Shared polling mistakes could upset both calls together.

Outside evidence is mixed, not independent confirmation. The October 2 Senate consensus leaves five seats as toss-ups, and its prediction-market comparison points toward Democrats. Those are reasons to keep the Republican call tentative. I have not converted either source into our model’s probabilities.

Written by Codex under human stewardship, using the model run saved October 4 at 19:43 UTC. This is a dated editorial call, not an automatic daily update, a Claude Code audit conclusion or a validated control forecast. It is kept separate from the frozen model tests. Later revisions should preserve this original call.

New experiment · the whole chamber

How many seats could each party win?

We add up simulated races, allowing many races to move in the same direction. Safer-looking seats still have a chance of an upset.

These are scenarios, not settled predictions. Most seats use past voting and candidate experience, with recent local polls where we can match them. This version does not blend in national polling. The shift controls let you test a different national mood. Seats we cannot model stay unresolved. No probability of chamber control is published.

October 13 ballot check: eight Virginia districts have additional candidates missing from our earlier list. Those seats now stay unresolved rather than being assigned to one of the two major parties. Official Virginia candidate list.

Chamber What if the model misses the national mood? How much do mistakes move together?

A four-point shift changes each modeled race’s Democratic-minus-Republican margin by four points. It is not a four-point change in either party’s vote share, or a prediction of the national popular vote. The shared-error settings are assumptions we are testing.

Prior federal wins remain unverified in our archive for 26 listed incumbents. Unmatched is unknown, not proof of no experience.

How much do we know?

How wide could the result be?

Dark lines show the average seat range allowing for missing races. Pale lines show the wider possible range. The dashed line marks more than half the seats, not a chance of control.

How to read the ranges: the wider range covers the middle eight out of ten simulations for modeled seats, then allows unresolved seats to go either way. It is not a proven 80% range for the whole chamber. Each party’s endpoints describe different possible outcomes; they do not have to add up to 435 or 100.

Seat totals under the selected assumptions
Party	Average total, allowing for unresolved seats	Wider possible range	Average seats won this election, allowing for gaps

Explore the country

Select a state to see its seats. This map shows where our evidence comes from, not which party will win. State size does not represent seat count.

At least one unresolved seatSome local polling, no unresolved seatsBackground or fixed seats only

Explore a state

Loading the map. The seat list is available below.

State outlines: U.S. Census, via us-atlas. These are state outlines, not current congressional district boundaries. Alaska and Hawaii appear in insets.

Every seat, including the gaps

Blue means a Democratic lead; red means a Republican lead; purple marks independents. These are not election calls. ? means unresolved, ○ means not up for election, and ✓ means a certified same-party ballot. Unmarked seats are modeled. Select a seat for details.

Select a seat to see how it is handled.

Read the complete seat list

Seat	Basis	Details

The historical House estimates leaned too Democratic in six of seven elections. They also missed the 2018 Democratic wave in the other direction. Shared errors widen the ranges; they do not fix this bias. The model has no national-mood adjustment yet.

What has been tested so far?

We reused race predictions made in historical tests that learned only from earlier elections. We then counted party wins under different assumptions about shared errors. These tests cover eligible subsets of seats, not complete historical chambers. Missing races, changed boundaries and unusual ballots prevent a claim that the chamber ranges are calibrated. We have not chosen the shared-error setting by picking whichever looked best in this test.

Earlier-election checks on eligible subsets, 28 days before election day. Seat counts differ by year.
Chamber	Year	Seats tested	Democratic wins	Model average	Modeled range	Democratic seat error (+ means too high)	Range width
House	2012	326	143	152.5	129–179	+9.5	50 seats
House	2014	337	142	156.8	140–175	+14.8	35 seats
House	2016	327	146	154.7	139–172	+8.7	33 seats
House	2018	350	183	163.8	145–183	-19.2	38 seats
House	2020	381	194	206.1	187–225	+12.1	38 seats
House	2022	372	191	198.4	181–215	+7.4	34 seats
House	2024	341	167	167.7	156–180	+0.7	24 seats
Senate	2012	31	23	20.7	17–24	-2.3	7 seats
Senate	2014	27	11	13.6	12–15	+2.6	3 seats
Senate	2016	29	11	12.3	10–15	+1.3	5 seats
Senate	2018	29	19	19.7	17–22	+0.7	5 seats
Senate	2020	26	12	13.7	11–16	+1.7	5 seats
Senate	2022	25	10	10.6	8–13	+0.6	5 seats
Senate	2024	28	15	16.5	14–19	+1.5	5 seats

Independent candidates and some ballot formats need separate treatment. 9 contests have certified same-party ballots: 8 Democratic and 1 Republican. Those seats are counted by ballot party, conditional on the slate remaining in force—not by an assumed “safe” rating. Missouri’s district boundaries also need further review. A rating such as “safe” is never converted into a guaranteed win. The Senate includes 65 continuing seats; the 35 elections include Florida and Ohio special elections. Other source candidate lists remain provisional.

Seat inventory: 270toWin House and Senate. Same-party California ballots checked against the official certified list. Continuing seats checked against Senate Class I and Class III. Their ratings do not set our numerical estimates.

Saved 2026-10-23T23:46:23.516478+00:00. Codex is the primary builder under human stewardship. Claude Code audited earlier versions. This v4 ballot correction has passed the builder's tests but has not had an independent audit. It is separate from stock analysis.

Two approaches, one fair test

Compass and Tide

Compass looks at local polls, voting history and candidates’ past elected experience. Tide asks how the national political mood might shift an area’s usual vote. We save both so we can check which is closer after the election.

A separate test corrects four candidate-name matches

In NY-17, PA-07, VA-01, VA-05, different spellings caused our archive to miss a candidate’s past federal election win. We now save a corrected version alongside the original when a matching local-poll forecast is available. It does not replace the Compass estimates shown here.

We will compare the same races at the October 20 and October 27 checkpoints. Earlier forecasts remain unchanged. This fixes an input problem; it does not yet show that the corrected version predicts better.

Tide adds the national congressional polling lead to an area’s past lean compared with the country. For example, an area five points more Republican than the country, with Democrats three points ahead nationally, gives an R+2 estimate. That example is made up.

Only 0 eligible polling firms; Tide needs at least 3. Checked Oct 23, 2026 at 23:45 UTC. 0 eligible polling firms; at least three required. Tide does not yet give a chance of winning or an uncertainty range. Neither model is proven better.

Why were other polls left out?

YouGov: Outside the 21-day fieldwork-end window
The Argument/Verasight: Outside the 21-day fieldwork-end window
HarrisX/Harris: Outside the 21-day fieldwork-end window
YouGov - CBS News: Polling organization needs identity review
Pew Research Center: Polling organization needs identity review

The original records are preserved. We do not substitute guessed corrections.

Choose a race

Circle: Compass. Square: Tide. Both use vote-margin points: right means a Democratic lead, left a Republican lead. These dots are estimates, not confidence levels. Missing dots mean no estimate.