Methodology
Evidence before interpretation.
The Stated Record is built to put what candidates say next to how they voted, how their campaigns are funded, and who backs them, and to show where those records line up and where they do not. We use AI-assisted research to find and organize public information. Each claim about a candidate is checked against its source, and a person signs off before it is published. Official records we show directly, such as roll-call rows from the House Clerk, appear as filed. One fixed procedure runs on every candidate the same way, and it never guesses at motive.
What this site is for
Candidates tell you what they stand for. Their records show what they have done. This site is built to put the two side by side for every candidate on the ballot in the races it covers: their own words, their votes, bills and official acts, and the campaign money they filed, each with its source. We compare a candidate only with themselves, never with an opponent. Each comparison gets at most one label from a short, fixed list, or says why it has none. Nothing here says why anyone did anything. Today thirty-six races carry that material, and every card on them meets the rules below.
How the analysis stays unbiased
Not by having no analysis, but by running the same procedure on everyone. Every published card must meet these rules.
- Same fields for every candidate. Every candidate in a race gets the same sections. A gap is shown as a gap, never filled by inference and never left out for one person.
- Labels from a short fixed list. Each comparison gets one label, chosen by rule, not by taste, and every rule applies to every party alike. Today's list is matches stated direction, matches stated record claim, not comparable yet, and record does not bear out the claim. A card may instead say No assessment; it explains why.
What each label means
Matches stated direction: an official action moved the way the statement calls for. A recorded vote, a sponsorship, a cosponsorship, and a formal act of an executive office the candidate holds (an opinion issued, a suit filed, a settlement entered) all count as official actions. The card says which, and for an executive act, what kind of act it was, because an opinion and a settlement are not the same instrument, and neither is a vote.
Matches stated record claim: a claim the candidate makes about their own record is borne out by the official record. A claim to have sponsored, introduced or led a bill matches when the candidate is listed on that bill as its sponsor or as a cosponsor, because a legislature names one sponsor per bill and a co-lead appears as a cosponsor. The card states the official roles. Stronger words, such as wrote or championed, are judged on their own.
Not comparable yet: no like-for-like official action exists to set beside the statement.
Record does not bear out the claim: a claim the candidate makes about their own record is not borne out by the official record as checked. The check covers the official titles and summaries of every bill the member sponsored in the relevant sessions, and the source the campaign itself cites, read by a person when it blocks automated reading. The card shows what the record does contain and names what the check could miss. No card carries this label while the evidence that decides it is unread. It is kept for a record that contradicts the claim: a stated position the candidate voted against, a bill they are not on at all, a law that is not law, a number the record does not support. Where the record runs the same way as the claim and only a stronger word or a title overreaches, the card matches stated direction, and its note says plainly what the record shows.
No assessment: for one situation only. A like-for-like official record plainly exists, but no official source publishes the figure the claim turns on, so every label above would misstate why we are not judging it. The card still shows the whole record it did find, names the document that would settle the claim, and states that it was not read. It is not a way to avoid a hard call: a claim the record can test gets a label.
Adding a label, such as changed position, means adding its rule to this page first.
- Both sides, with their sources. A statement is quoted or closely paraphrased, with its source attached, and dated whenever the source carries a date. If a card compares two statements that differ, both appear with their dates and the difference is described. We do not call it a flip-flop, a lie, or growth.
- Money is shown, not interpreted. Campaign money appears as filed amounts, with the filing attached. If money ever appears next to a vote, it is context, not evidence of why the vote happened, and the card must say so.
What money we show today
Committee totals and, for all thirty-one federal desks, itemized receipts from political committees grouped by the FEC's recorded organization type, each with the filings it covers. Some desks also compare those rows with the campaign's own FEC summary. The five governor desks point to the state filings instead: Wisconsin's links to them, because state law limits reuse, and the other four have not added them yet. Any further itemized detail must pass the same review first.
- No score, ranking, or verdict. We publish no candidate score, no ranking, and no automated verdict. Readers draw their own conclusions from records they can check.
- Every card is checked, then signed off. One model proposes a comparison and cites the passages it relied on. A separate, independent check reads each claim against its source and reports what holds and what does not. Where a site blocks automated reading, a person opens the source instead. A person signs off on the checked set before it appears. Nothing on the site asserts why someone did something.
- Corrections are public. Errors are corrected on the record, dated, and attached to the affected page.
Different sources answer different questions
Official election records show who is on the ballot. Congressional records show what was voted on and how members voted. News reporting adds interviews, local developments, and context. Election-rating publications judge how competitive a race looks. A news article and a race rating are not interchangeable evidence.
Sources used in our current coverage
Each page groups its evidence in Sources. Webpages are linked there; files are listed by title, publisher and available filing or document identifiers. Dates, scope and limits stay with the facts.
- Ballots: state and county election authorities, identified in each race's Sources.
- Votes and bills: House Clerk, Congress.gov and state legislatures.
- Campaign money: Federal Election Commission data, with committee, report and coverage details retained for the figures.
- Statements and background: official offices, campaign platforms, original debate recordings and reporting, identified by publisher and date.
- Race ratings: Sabato's Crystal Ball, credited by name and date under its permitted use. We publish no other outlet's rating without permission.
How AI and editorial judgment work together
AI helps find sources, organize material, draft research, and propose comparisons across statements, votes, filings, and public posts, always with the passages it relied on. An independent check reads each claim against its source, and a person signs off before it is published. Human review remains part of our publication process, especially for comparisons and interpretations. We tell reporting from opinion, trace syndicated stories to their origin, and do not count repeated coverage of one report as independent confirmation. Where sources disagree, we explain the disagreement instead of averaging it into a fact.
How launch races are selected
The September 2026 launch set is six Senate contests, four governor contests, and 10 House contests. In October we added one governor contest and 16 House contests: the races Sabato's Crystal Ball rated Toss-up in its ratings of September 29, 2026 that we did not already cover. A desk stays in the set once added, even when its rating moves. Explore the races.
More detail
We do not print a category from a publisher that has not allowed its use as the reason a race is here. Ballot information is checked with the responsible election authority. Candidate names are separate from full profiles and interpretive reporting.
How the rating synthesis works
Every publisher keeps its own exact category, date, link, review state, and freshness state. When only one current approved rating exists, it is shown as that publisher's rating, not as ours. The Stated Record comparison is our reading of approved ratings, not a fourth independent rating, forecast, endorsement, or poll.
Ratings are ordinal judgments, not probabilities, so we never turn them into percentages. Only verified, current ratings from an approved publisher can enter a comparison. During the election year through election day, a rating goes stale after 45 days; outside that window, after 90 days. A stale rating stays visible with its date but is left out of the comparison.
How the scale and the comparison work
We map publisher labels to one ordered scale: Safe Democratic, Likely Democratic, Lean Democratic, Tilt Democratic, Toss-up, Tilt Republican, Lean Republican, Likely Republican, and Safe Republican. “Solid” maps to Safe. Tilt and Lean remain separate categories.
With three current ratings, the comparison uses the middle category. With two that are not on opposite party sides, it uses the ordinal midpoint and resolves a half-step toward Toss-up; two on opposite sides resolve to Toss-up. Zero current sources produce no comparison. Crossing party lines is labeled as disagreement, and a same-side spread of three or more category steps is labeled as wide disagreement.
This rule is deliberately modest: it keeps each publisher's category, handles ties the same way every time, and avoids invented mathematical precision. New rating sources are added only after their use is approved.
How reporting supports “What to know”
Every published sentence links its source or sources and carries an “as of” date. A campaign-development sentence needs at least one original report approved for that claim; higher-risk interpretations can require two genuinely independent sources. If that bar is not met, the race page says what is still incomplete instead of filling the space with generic process language.
How we weigh sources
Official records establish ballot identity, votes, dates, and filings. Original national and local reporting can establish dated campaign developments and context. Opinion, endorsements, candidate statements, press releases, ratings, and polls are labeled by type and are not interchangeable with independent reporting. Claims made on a candidate's own channels stay first-party claims unless independently confirmed.
We trace syndicated reporting to its origin and count one underlying AP, Reuters, press-release, or other origin story only once.
What the coverage list means
Every race page ends with a coverage list. It describes our coverage, not the candidates, and it never fills a gap with inference.
Included means that area has passed its source check and appears on the page. Partial means only part of it is verified, for example specialist ratings shown with their dates but past our freshness window. Not yet means nothing in that area is published for the race.
Polling and printed briefs
Polling will appear only with the pollster, sponsor, field dates, population, sample size, method, and margin of error when available. Every race page can also be printed as a brief. The printed copy carries its snapshot date, source identities and webpage addresses. File sources retain their document details.
What a comparison means
A comparison puts two documented records side by side. On its own, it does not prove motive, dishonesty, or hypocrisy. Votes can be procedural, bills can contain many provisions, and public language can change over time. So each published comparison includes its primary sources, the relevant bill context, dates wherever the sources carry them, a label from the fixed list above or the reason it has none, and a note on the limits of the evidence.
Corrections
Substantive corrections will be dated and attached to the affected record. Readers can report an error through the corrections form.