This page explains where our data comes from, how quotes and evaluations are checked before they appear in the app, and how often everything updates. It is a first-class part of the product: if you cannot tell how we know something, you should not have to trust it.
Nothing is displayed in the app before its verification path, described below, is live. Where a section describes something we intend to do rather than something we currently do, it says so.
We show receipts: verbatim quotes with links to their source, roll call votes linked to Congress.gov, donations linked to FEC records. Model-generated scores are secondary. They always appear together with their reasoning, their citations, and a link to this page. All candidates are processed by the same pipeline, with the same prompts, the same scoring rules, and the same display treatment, regardless of party.
| What | Source |
| Candidates, committees, contributions, independent expenditures | Federal Election Commission bulk data and the OpenFEC API |
| Bills, roll call votes, member records | Congress.gov API (Library of Congress) |
| Lobbying filings | U.S. Senate Lobbying Disclosure Act database |
| Member ID crosswalk | the unitedstates/congress-legislators open data project |
| Donor industry classification | OpenSecrets (opensecrets.org), used under their bulk data license. We thank OpenSecrets for making this work possible. |
| Ideology scores | Voteview (voteview.com), DW-NOMINATE |
| Candidate statements | Official congressional websites (house.gov, senate.gov), campaign websites, press releases, and public video, with archived snapshots from the Internet Archive Wayback Machine |
| District lookup | U.S. Census Bureau geocoder and TIGER shapefiles |
Every fact stored in our database traces back to a recorded retrieval: the source URL, the retrieval time, and a cryptographic hash of the raw payload. Deep links in the app point to the official record (fec.gov, congress.gov) wherever one exists.
Campaign finance figures come from two independent FEC channels, and we show our work by keeping both:
Comparing our itemized sums against the official totals is a permanent accuracy check. Divergence beyond normal bulk-processing lag is investigated before the affected numbers ship.
Accounting rules applied to itemized records:
A political action committee may give a candidate $5,000 per election, which is $10,000 across a primary and a general. There is no comparable ceiling on spending independently for or against that same candidate. The practical result is that a candidate's largest direct donors are often a row of identical $10,000 entries, while the decisive money sits in independent expenditure beside them. Party committees, and transfers from a candidate's own affiliated committees, follow different rules and can appear in a donor list for considerably more. Across the candidates we cover, outside spending exceeds the single largest direct donor for a clear majority of those who have any.
We therefore name the committees that spend independently, not just the total. Each candidate shows the largest spenders for and against them, each one linked to its own FEC record, ranked separately by side so that a lone committee spending against a candidate is never pushed off the list by a crowd of supporters. Naming them costs us nothing in accuracy: the spending committee's identifier is part of the filing, and it covers all but a fraction of a percent of the outside dollars we hold.
This is a mechanical rule applied to every candidate. We do not decide which backers are worth mentioning, and we do not annotate any candidate with a description of who supports them. The list is whoever filed the spending.
Campaigns report their totals to the FEC in summary form, and separately file the itemized records behind those totals. We hold all of the summaries and only part of the itemization, so every candidate shows how much of their individual contributions we hold as itemized records.
Contributions from committees are close to complete. Itemized contributions from individuals now stand at roughly four fifths of what the FEC's summaries report, and the remainder is mostly contributions filed since our last bulk load. Each candidate shows their own figure rather than this average, because the spread is wide.
Two things that section gets right only because they were once wrong.
The comparison covers one period. A campaign's official summary runs up to a filing date, and the itemized file we load runs to whenever we last downloaded it. Measuring everything we hold against a summary that stops in March told twenty candidates' readers that we held more of their money than the campaign had reported raising. We now count only contributions dated within the period the campaign has actually reported on.
A candidate's own money is not a contribution to them. The FEC records money a candidate gives their own campaign under its own transaction type and reports it on its own line; we were adding it to individual contributions, which made a self-funded campaign look like a campaign with supporters. That was $27.5 million across 860 candidates in this cycle. It is now shown separately and labelled as what it is, because on a site about who is backing a candidate, the answer "they are" is a real answer and not a rounding detail.
A small number of candidates still show slightly more held than reported, usually by a few percent. That is contributions filed after the campaign's last summary but dated inside it, and the card says so rather than rounding the number down and claiming to be complete.
There are three ways money reaches a candidate, and a donor list shows only one of them.
A committee can give directly, capped at $5,000 per election. A group can spend independently for or against the candidate, uncapped, which we cover above. And an organisation can bundle: it asks many individuals to give, collects their contributions, and delivers them together. Every one of those gifts is an individual's own and stays within the individual limit, so nothing about it is irregular, but the organisation that assembled them directs a total far larger than any single limit and appears nowhere in a donor list, because it never donated.
The filings identify the collecting committee, so we name it. Each candidate shows the organisations that bundled contributions to them, how much arrived that way and across how many gifts, each linked to its FEC record. The count matters as much as the total: an organisation routing $400,000 across nine hundred contributions is doing something different from one routing it across three.
This is stated as a channel, not an allegation. Bundling is lawful, disclosed and ordinary, and it is treated identically for every candidate. We report it because a reader asking who is behind a campaign is asking about exactly this, and answering with only the capped committee donations would be a technically accurate reply to a question nobody asked.
We measure that gap on individual contributions alone, and it is worth saying why, because the first version of this measure was wrong. Combining committee and individual money into one ratio reported that we held more money than existed for 388 candidates. Two causes: our committee sums include coordinated party expenditures and in-kind transfers, which the FEC's own contribution total does not count, and itemized filings routinely post-date the summary they belong to. A coverage figure above 100 percent discredits the disclosure it is trying to make. Individual money suffers from neither distortion, and it is the gap that actually matters, so that is what the number reports.
This matters for one mechanism in particular. Organizations that bundle earmarked contributions from many individuals exercise influence through exactly the channel we have least of, and their role is correspondingly under-represented in what we can show today. We would rather publish that limitation next to every donor list than let a short list imply a complete one.
Contributions below the itemization threshold are excluded from both sides of this measure, because no itemized record of them exists at the FEC either. Counting them against ourselves would describe a gap that nobody could close.
Roll-call votes come directly from each chamber's official record: House votes from the Congress.gov API (which mirrors the House Clerk's records) and Senate votes from the Senate's own published roll-call XML. For every vote we display, you can click through to the official government page for that roll call.
A promise appears in the app only after it passes two independent checks: the quote has to be real, and the statement has to be a promise.
We separate what a member says in their official capacity from what a campaign says. Official congressional websites are published by a government office under rules that restrict campaign content; a campaign website is the candidate asking for your vote. Both are that person stating what they intend to do, but only one is a campaign promise, and we label them differently rather than merging them.
We do not search for a candidate's website. We read the address the campaign gave the Federal Election Commission.
Every committee a candidate authorizes files a Form 1, and that form asks for the committee's web address. It is the campaign's own statement, made to a federal regulator, about where its campaign lives. We ask the FEC for each candidate's authorized committees and take the address from there, matched on the committee identifier rather than on anyone's name.
The reason is not convenience. The other ways to find a challenger's website are to search their name or to buy a list from someone who did. A name search has to decide which Brown is Sherrod Brown and which is Shontel Brown, and a wrong decision publishes one candidate's words on another candidate's page. Reading an address the campaign filed itself removes that decision.
Four things we will not do here.
We do not repair a mistyped address. One 2026 filing reads VONDRASFORCONGRESS,ORG, where the comma is plainly meant to be a period, and we still record that we have no site rather than invent one the campaign did not file.
We do not read committees the candidate did not authorize. A political action committee may file a website too, and publishing that as the candidate's own would put a spender's words on a candidate's page.
We do not treat a candidate's Facebook page or donation form as a campaign site. We count them and say so, because there is nothing on them a promise can be read from, but we do not pretend they are the same thing as a site with stated positions.
We do not hide it when a page was read over a connection we could not verify. Some campaigns run misconfigured certificates, which a browser repairs silently and stricter software does not. Refusing those pages would have reported live campaigns as silent, and it would have fallen hardest on campaigns with the cheapest hosting. So we read them, and every page read that way is stored with a mark saying the certificate was not verified: what the host served is still evidence, but we cannot promise the host was who it claimed to be.
Two limits worth stating plainly. First, we measured the screening rules against a hand reviewed set of 118 quotes drawn from video transcripts, and they are less well measured on written material, which currently supplies most of what we hold. Second, dropping a quote is not a claim that the candidate never made the underlying commitment; it means this particular quote is not usable as evidence of one.
For promises that are specific enough to check, we compare the promise with the incumbent's roll call voting record. Evaluations currently cite votes only. Campaign finance is displayed alongside a candidate, and you can read both on the same page, but we do not generate scored claims about a donation having influenced a vote, and no evaluation cites a donation as evidence.
Scores are the least important thing on the page, and they are the part most likely to be wrong. They exist to order and summarize evidence you can read yourself. Where the score and the underlying votes disagree, believe the votes, and please tell us.
We publish no finding that a member BROKE a promise. Those findings exist in our database and every one of them is withheld pending human review, by a rule enforced in the database itself rather than by anyone remembering.
The reason is specific. An earlier version of this scoring was shown a bill's title and not what the bill did, and it reasoned backwards from there. A bill called the "Homeowner Energy Freedom Act" repeals home energy efficiency rebates; a member voted against it, thereby protecting the rebates he had promised to protect, and we recorded that he had broken his promise. We found it in review, withdrew every score we had published, and disabled the scoring in code so it could not be regenerated by accident.
The rebuild is not a better prompt. The model is no longer asked whether a vote supports or contradicts a promise, because that is the judgment it was getting wrong. It is asked only what passing the bill would do, and the direction of the vote is then arithmetic. Voting down a repeal now counts as protecting the thing being repealed, by construction rather than by instruction. Separately, a vote on a contested question that a promise never raised cannot be scored at all: a promise about working families is not a position on abortion, guns or immigration, and treating it as one would be a political judgment we do not make on anyone's behalf.
We then screened the remaining broken-promise findings ourselves, flagging any that cited a vote AGAINST a bill that would have repealed or narrowed something, since that is the shape the old error took. About half were flagged. That is a rough automated screen and not a considered review of each one, which is precisely the point: we do not yet know which of them are sound, so none are published, and the findings you do see are limited to "acting on it" and "completed".
The app loads a published data snapshot with a version stamp, and the "data as of" date is always visible in the app. That date, not this section, is the authoritative answer to how current what you are reading is.
Our cadence is weekly for the parts that can be automated safely, and manual for the parts that cannot.
Automated, weekly: campaign finance from the FEC, roll-call votes from Congress.gov, a fresh read of members' official websites, and the rebuild and republish of the site. This is the data that goes stale, and none of it requires a judgment call.
Deliberately manual: reading new documents for promises, and assessing promises against voting records. Both need a rented GPU, so automating them would mean spending money unattended, and the second produces the one kind of finding we will not publish without a person reading it. Those runs happen when someone decides to do them, and the site always shows the date of the data you are looking at.
The map shows city and town names so you can orient yourself. Those labels cannot tell you your district, and we are careful not to imply they can: congressional districts routinely split cities. Five points across Houston fall in five different districts, so a single label tied to one district would be wrong for most of the people living there.
The address box is the reliable answer, and here is exactly what it does.
The one technical caveat, stated because it is a real tradeoff: to work on a site with no server of its own, the lookup loads the Census reply as a script rather than a normal data request. That is a deliberate trust in a US government domain over HTTPS, and it is the only external service the page talks to.
The FEC issues a candidate a new identifier for each office they seek, and nothing in the published data joins the two. A sitting House member running for the Senate therefore appears twice, and their voting record sits under the identifier they are no longer campaigning under. Fourteen candidates were in that position here, and their Senate pages showed sitting members of Congress with no record and no promises, as though we had never heard of them.
We join those identities, and the standard for doing so is deliberately high, because the failure mode is one member's votes appearing on another member's page. Two independent things must agree.
The FEC's own candidate master has to list both identifiers with a character-for-character identical filed name, the same state, and the same party, across two different offices, with exactly one marked as an incumbent. That is a single source in a single format, self-reported by the candidate. We do not match names across two different systems, which is guesswork dressed as evidence, and we do not match on surnames: Ohio has had a Sherrod Brown and a Shontel Brown at once.
Then our own database has to agree independently, because the FEC's incumbency marker is self-declared and stale filings carry it. The incumbent-side identifier must resolve to a member we hold a real roll-call record for, and the other side must have none.
Anything that fails either test is left alone rather than guessed at. On the current data that reduced twenty-nine name matches to fourteen links, and a hundred and five ambiguous groups were never candidates for it. Every link we make is stored with the evidence for it and the superseded identity is kept, so any one of them can be checked or undone.
Coverage is uneven, and the app says so per candidate rather than leaving you to guess. At present:
Where we have nothing for a candidate, the app says we have nothing. An empty section means we have not gathered it yet, never that the candidate said or did nothing.
The map draws that distinction too. A district shaded "collected, not yet analysed" is one where we hold the candidates' own pages and have not yet run extraction over them, which is a different statement from having nothing to show. It reverts to the plainer shading once the material has been read, whether or not reading it produced a single publishable promise.
If something here is wrong, we want to fix it and say that we fixed it.
Open an issue at github.com/Tally-2026/Tally-2026.github.io with the candidate, the item, and what you believe is incorrect. Quotes and votes carry deep links to the official record, so the fastest correction usually cites that record directly.
Corrections to underlying data are made by re-running the pipeline against the corrected source, not by hand editing a stored fact, so the provenance trail stays intact. Evaluations are never edited in place: a corrected or re-scored evaluation is stored as a new version, and the record of what was previously shown is retained.