The five joint outcomes list the House first: Democratic House and Senate, Democratic House, Republican Senate, Republican House, Democratic Senate, Republican House and Senate, and unresolved. They are probabilities over the model’s weighted simulated draws. “Unresolved” means at least one chamber’s party control is unresolved under the modeled outcome rules. These values are model probabilities, not betting prices, guarantees, or individual race win probabilities.

The model keeps a candidate’s native party separate from any documented caucus alignment. It reports native party seat counts and caucus-aligned counts as distinct summaries; chamber control uses the explicit D, R, or no-majority classification. An independent can therefore remain an independent in the native-party accounting while contributing to a documented caucus alignment. House control requires 218 seats. Senate ties are assigned using the modeled vice-presidential party; the current definition assigns the tie-break to Republicans, so a 50–50 Senate is classified as Republican control.

The simulation models curated pairs of candidate or homogeneous party-group sides using historical election performance, district partisan lean, current race polls, generic-ballot and presidential-approval research inputs, and shared election-cycle uncertainty. The candidate outcomes are correlated through shared and race-level uncertainty rather than treated as independent coin flips. It then derives party seat counts and chamber control from those draws.

The published hybrid takes the simulated joint draw distribution and uses soft entropy reweighting to move its correlated draws toward reviewed election-result price bands from both Kalshi and Polymarket. The two venues are matched to the model’s election-result definitions; their source-specific quote ranges and overlap are retained when the evidence is combined. Market evidence is therefore a model input, not an independent validation result. If the required matching inputs are not available, the hybrid cannot be published.

The model also uses reviewed ratings from Inside Elections and Sabato’s Crystal Ball as analyst priors before polling information is applied. Rating labels map to assumed favored-side scores: Solid 0.97, Likely 0.85, Lean 0.65, Tilt 0.55, and Toss-up 0.50. When both sources rate a race, their probabilities are averaged on the log-odds scale. The analyst weight is 0.25 × (source count ÷ 2) ÷ (1 + ordinal disagreement²), then multiplied by (1 − the race’s polling weight) and a sensitivity nudge. The central release uses a nudge of 1; the model also checks 0, 0.5, and 2. This moves the fundamentals prior toward the analyst score while limiting influence when sources disagree or polling already carries more weight. The labels and scores are research assumptions, not calibrated probabilities or proven measures of source accuracy.

Candidate coverage is curated and incomplete by design. A matchup that is not modeled is not a forecast of that candidate’s chance, and no probability is assigned to unmodeled candidates through an automatic fallback. A party group has no forecast for which individual member wins. When only one side is modeled, that side wins within this research convention; this does not establish that every other ballot or write-in candidate has no chance. Reviewed ballot corrections are applied before the research features are generated and retained with their sources.

Original analyst categories and publication dates are retained. A date-only release becomes eligible at the next local midnight in its source timezone. Ratings that cannot identify a modeled side remain in the source archive and do not adjust its prior.

Conditional panels report precomputed weighted simulation results: the other chamber’s control given the selected chamber’s control, or given that one party reaches the selected seat threshold. House threshold results use a five-seat grid plus observed seat-count endpoints and the 218-seat control boundary. Senate threshold results use one-seat increments and include the 50- and 51-seat boundaries. Only thresholds with a precomputed result are selectable; unsupported or low-support cases are shown as unavailable. These are associations within the model’s simulated outcomes. They do not estimate the causal effect of one chamber’s result on the other.

The conditional panel’s raw draw count and effective draw count describe support in the weighted simulation. Its approximate 95% Wilson half-width measures Monte Carlo sampling error only, and is displayed only when support clears the model’s publication rule. It does not include uncertainty from polling, model assumptions, source quality, or future events. It is not a forecast confidence interval. The scenario probabilities likewise have no displayed confidence interval.

• Historical House and Senate results: the MIT Election Data and Science Lab U.S. House collection and U.S. Senate collection, hosted by Harvard Dataverse. Historical candidate and incumbency records also use Candidates in American General Elections.
• District partisan lean and presidential results: data from The Downballot (formerly Daily Kos Elections), with official map-specific results from the North Carolina General Assembly, California Statewide Database, and Texas Legislative Council.
• Current race-poll inputs include VoteHub polling data and manually reviewed poll releases, with individual pollster/source links retained in the underlying research records. Historical polling research includes FiveThirtyEight’s public data archive and archived source pages from the Internet Archive. Presidential approval history is from the American Presidency Project at UC Santa Barbara.
• Analyst priors: reviewed 2026 House and Senate race ratings from Inside Elections and Sabato’s Crystal Ball at the University of Virginia Center for Politics.
• Market evidence: reviewed public election-result markets and order-book data from Kalshi and Polymarket. Their prices enter the hybrid as source-specific bands after contract definitions are matched to the modeled event.
• Candidate roster cross-checks use the Federal Election Commission candidate master file.

Source inputs are reviewed, time-stamped, and retained with provenance for each release. The links above credit the organizations whose data, analysis, or market observations contribute to the research.