Methodology
This site makes claims about how real people voted. That only means something if you can check it, so here is exactly where every number comes from and what it does not cover.
What is loaded right now
| Dataset | Rows | Last fetched |
|---|---|---|
| committees | 230 | 2026-08-21 |
| congress_gov_policy_areas | 15,114 | 2026-08-21 |
| fec_pac_contributions | 31,281 | 2026-08-21 |
| legislators | 2,687 | 2026-08-21 |
| stance_classification | 30,810 | 2026-08-22 |
| voteview_members | 14,803 | 2026-08-21 |
| voteview_rollcalls | 52,115 | 2026-08-21 |
| voteview_votes | 15,672,969 | 2026-08-21 |
| wikidata_demographics | 2,686 | 2026-08-22 |
Sources
- unitedstates/congress-legislatorsPublic domain (CC0)
Member roster, terms, parties, birth dates, cross-reference IDs
- unitedstates/imagesPublic domain
Official member portraits
- Voteview (UCLA)Free for research use
Every roll-call vote and DW-NOMINATE ideology scores
- Congress.govPublic domain
Official Library of Congress policy area for each bill
- FECPublic domain
PAC contributions to principal campaign committees (Schedule A, line 11C)
- Google Gemini (gemini-3.6-flash)Commercial API
Vote-direction classification: which way a YES vote pushes on each issue
- WikidataCC0
Race/ethnicity, religion, birthplace
How the scores are computed
Issue lean (−100 to +100)
Each roll call is tagged to an issue and given a direction: does a YES vote push toward the issue’s high pole or its low pole? A member’s lean is the share of their scoreable votes that went each way. Votes whose direction cannot be read confidently are excluded entirely rather than guessed at — a wrongly-signed vote would silently invert someone’s record. No score has to be taken on trust: every issue on a member’s page opens onto the roll calls behind it, including the ones that did not count and why, and each of those opens onto the full measure, the tally and the recorded direction.
Consistency (0–100)
The member’s lean is recomputed separately for each Congress they served in. Consistency is 100 minus the mean absolute deviation of those per-Congress scores. It is undefined — not 100 — for anyone who served a single Congress, because one term is no evidence of consistency.
Party loyalty and bipartisanship
For each vote we find the majority position of each party. Loyalty is how often the member matched their own party’s majority; bipartisanship is how often they matched the other party’s. Votes where a party split evenly are excluded from both.
Issue tagging
Where Congress.gov publishes an official policy area for the underlying bill, that is used. Otherwise a keyword rule set runs over the vote description. Every tag records which rule produced it, so any classification can be audited or disputed.
Known limits
- Race and ethnicity are not published by any official congressional dataset. Where shown, the value comes from Wikidata and is attributed to a specific entity ID. It is never inferred from a name, a photo, or a district. Members with no sourced value read “Not recorded”.
- Portraits are incomplete. The public image set covers almost every sitting member but under half of those who served since 1973, and very few before that. Members without one get a monogram.
- Roll-call votes are not the whole story.Voice votes, unanimous consent, committee votes, and bills killed without a vote leave no roll call. A quiet record is not necessarily an inactive one.
- Contributions are not bribes. A donation is a documented financial relationship, not proof that it changed a vote. This site reports the relationship and lets you draw your own conclusion.
- Contributors are not grouped by industry.OpenSecrets, which publishes that coding, now requires a paid subscription and is not used here. No free source codes contributors by industry, and guessing one from a committee’s name would invent a fact, so PAC filings appear exactly as the FEC records them. Only PAC contributions are shown — not itemized individual donations, whose self-reported employer field will not support this either.
- Vote direction is inferred, and measured. Whether a YES vote supports or opposes an issue is not published by anyone — it has to be read off the text of the measure, together with the effect of the question (a yes on a motion to table kills what it names, so the member’s position is the opposite of the measure’s). A language model does that reading, is given each issue’s own two poles rather than any notion of left and right, and returns a confidence with every answer. Anything below the confidence threshold is excluded from scoring rather than guessed at, and each roll call stores the model’s own one-sentence reason next to the direction it produced. Measured accuracy: 48 of 50 correct (96%) on a held-out sample hand-checked against the roll-call record, with no reversed directions. Both misses assigned a direction to a vote that is not really a position on that issue, rather than getting a real position backwards. The rule-based classifier this replaced measured 72%, with 8% of directions reversed. Full method and every error: docs/CLASSIFIER_AUDIT.md.
- Age is shown only for sitting members.The source data carries no death date, so for former members we show birth year instead of implying they are alive.
Reporting an error
This site makes factual claims about how named, living people voted, so being correct matters more than being complete. If something here is wrong — a vote attributed to the wrong member, a direction that reverses what a bill actually did, a name or a term that does not match the official record — it should be fixed, and the fix should be visible.
The most useful report names the page, quotes what it says, and says what the record actually shows. Vote directions carry a recorded reason on every roll-call page, which is usually the fastest way to show where a judgement went wrong.