About Smarter.Vote

Automated research agents search public sources—including candidate statements, campaign websites, official records, voting databases, and reputable reporting—and organize the evidence into candidate summaries, issue positions, background, and race analysis. They are instructed to describe what candidates say and do without endorsing them.

Source links remain attached to the research so readers can inspect the evidence directly. A source is not treated as neutral merely because it is linked: campaign material represents the candidate's own account, while records and reporting provide additional context.

Candidates are listed with Democratic and Republican candidates first; within each group, incumbents come first, then alphabetical by last name. The order is not a ranking or an endorsement.

Confidence labels describe the strength of the available evidence—not the value of a candidate or policy. High confidence requires multiple corroborating sources or an official candidate position. Medium confidence requires at least one credible source. Low confidence marks information that is inferred, unverified, or unsupported by a source.

“No public position found” means the research process looked for an attributable position and did not find one. It does not mean the candidate has no view, and it is different from an issue that has not yet been researched.

Separate AI models review the full race side by side for the correct candidate roster, internal consistency, source quality, completeness, background accuracy, and neutral treatment across candidates. Automated checks also look for broken sources, malformed polling, placeholders, and missing required research.

The automated research score combines the review models' scores and accounts for unresolved warnings. A failing score or an unresolved error blocks publication. This is automated editorial quality control, not a promise that every statement has been independently fact-checked by a human.

Every race forecast starts from the published race record: the verified candidate list, candidate-level polling, fundraising, incumbency, the district or state's partisan lean, and the documented race environment. Prediction-market prices can add context, but they are discounted when trading is thin or stale and are never treated as ground truth.

Three AI models from different developers each estimate the race independently. None of them sees the others' answers or the previous forecast, so no single model's lean—and no stale number—decides the result. The published win probability is the median of their estimates, and the rating follows directly from it:

Toss-up

below 55%

Tilt

from 55%

Lean

from 65%

Likely

from 80%

Safe

from 95%

Confidence reflects how closely the models agreed and how much polling exists; a race without public polling is never labeled high confidence.

A separate writing step explains that consensus without changing its numbers, and a fact-check sends back any text that contradicts the race—for example, describing a primary as unresolved after it has been decided. Each forecast lists its key drivers, its main uncertainty, and the sources behind specific claims.

Chamber forecasts combine race probabilities with known holdover seats. Polling errors tend to run in the same direction across the country, so the model lets every race shift together with a shared national swing instead of treating each contest as an independent coin flip. Each race keeps its own probability, but the range of seat outcomes and the chance of party control reflect how a real polling miss plays out. The chamber analysis is drafted independently by several leading AI models and merged by an editor model that checks every figure against the forecast. For the 2026 Senate forecast, a 50–50 result is counted as Republican control under the vice president tie-break assumption.

Sources can be incomplete, positions can change, and automated analysis can misunderstand nuance. Polls capture a period in time, and forecasts change as new evidence arrives. Forecasts are estimates—not election results, voting advice, or guarantees.

Use the linked evidence to check research claims, and rely on state and local election officials for registration, deadlines, voting methods, and official results. Report errors through our corrections process. Funding never determines findings, classifications, forecasts, or publication decisions.