How we rate, and who we are

Every rating on this site comes from public records you can check yourself. Party is shown but never scored.

You can download every record and the scoring rules. If we got something wrong, send a correction.

Three questions, scored separately

A single grade would hide too much. A member can champion AI research and still push hard for ratepayer protections. We think both are good. So we score three things separately:

• AI innovation & infrastructure. Do they support American AI research, chips, and the data centers and power needed to run them?
• AI safety & accountability. Do they support practical safety testing, transparency, and accountability for powerful AI systems?
• Protecting ratepayers & communities. Do they make sure new data centers pay their own way and treat local communities fairly?

Supporting safety and supporting build-out are both positive. Fighting for ratepayers is positive. Blanket opposition to all data centers counts against innovation and infrastructure, but not against the other two.

How a score is built

We collect public records: floor votes, bills sponsored or cosponsored, committee actions, signed letters and public statements. Each record links to its source. Each record gets a score from −2 (strongly against) to +2 (strongly for) on each question it touches. It is left blank on the others.

Votes count more than words

Each record is weighted by what the member did:

• 3 points: a recorded floor vote, or sponsoring a bill or amendment.
• 2 points: a committee action, cosponsoring a bill, or signing a letter.
• 1 point: a public statement, press release or discussion draft.

A floor vote is on the record and has consequences. Sponsoring a bill means writing it and putting your name first. Committee work, cosponsoring and letters are real commitments, but smaller ones. Statements are easy to make and easy to walk back. We still count them, because for many members they are the only record so far.

One action counts once

An action is one bill, amendment, vote, letter or event. A bill's companion in the other chamber, and its reintroduced or follow-up versions, are the same action. So are statements about it. Every record in our data names its action.

When we have several records of one member on the same action, only the strongest counts on each question: the highest-weighted type, then the most recent. A member who moved the Ratepayer Protection Act through committee and then voted for it gets credit for the vote. Saying the same thing again does not move a score.

The score and the label

A member's score on a question is the weighted average of their counted records on it. For example, a bill they sponsored scored +2 and a statement scored −1 give (3 × 2 + 1 × −1) ÷ (3 + 1) = 1.25.

Labels: Strong supporter (≥ 1.25), Supportive (≥ 0.5), Mixed (between −0.5 and 0.5), Skeptical (≤ −0.5), Opposed (≤ −1.25). No records means No public record yet. That is not a negative rating.

How confident we are

Every rating comes with a confidence level. It depends on how many separate actions we have, and how much they weigh together:

• High: at least 3 actions, worth at least 7 points.
• Medium: at least 2 actions, worth at least 4 points.
• Low: anything less. Any rating built on a single record is Low.

Thin evidence should not look conclusive. So a Low-confidence rating is never shown as Strong supporter or Opposed. It stops at Supportive or Skeptical. The score itself does not change. Confidence measures how much evidence we have, not whether we judged each record correctly.

The overall rating on the map

The map colors each House district by its member's overall rating. It uses the same rule as the three questions, with their records pooled: the weighted average of every counted score on every question. A record that touches two questions counts on both. Confidence counts each action once, at its strongest weight, and the same cap applies.

We pool the records so that votes and bills outweigh statements here too. A question with one statement should not count as much as a question with three votes. The tradeoff is that a question with more records has more say in the overall rating. That is why the three questions are shown separately, and first.

Until November 5, 2026, every record counted the same, and the overall rating was a plain average of the three question scores. We changed both, and the change is in our change log.

Same rules for everyone

We score the record, not the person. Party is never an input. Every member is scored by the same code with the same weights. Members can improve their record at any time, and we'll update it.

Limits we want you to know

• Coverage is incomplete. We started with the members most active on AI and data centers and are adding more.
• Judging a statement takes judgment. We read it in full context and link it so you can judge for yourself.
• A vote that passed almost unanimously counts as much as a close vote, even though it tells you less about the member.
• District lines on the map are for the current (119th) Congress. Several states redrew lines for the Congress seated on January 3, 2027. The address lookup tells you if your district is changing.
• Research and first drafts on this site were assisted by AI tools. A person reviewed every claim, and every claim links to its source.

Check our work

Everything behind the ratings is public. These files are built by the same code that scores the site, every time we update the records, so they always match what you see.

• ratings.json: the scoring rules above, as data, and every member with a record: their ratings, each record with its weight and action, and which records counted.
• ratings.csv: one row per member of Congress, with each score, label and confidence.
• evidence.csv: one row per record, with its type, weight, action, scores and source.
• Change log (changelog.json): every change to the records and to these rules.

Corrections

Think we got a record wrong, scored it unfairly, or missed one? Email bdcaunt@gmail.com with a link to the source. Each member's record on the Find your reps page has its own correction link too. We fix errors quickly and list every change in the change log.

Data sources

• Members of Congress: unitedstates/congress-legislators (public domain).
• District boundaries: U.S. Census Bureau cartographic boundary files, 119th Congress.
• Address-to-district matching: U.S. Census Bureau Geocoder.
• Member photos: unitedstates/images.

How we calculated the point of no return for elder care

The care chart on our home page is our own calculation, not an official forecast. It asks one question: when does America run out of time to staff elder care with people alone?

1. Care needed. We take the share of each age group that got paid care at home or in a nursing or assisted-living home in 2022: about 6% of people 65–74, 9% of people 75–84 and 32% of people 85 and older. These figures come from the National Health and Aging Trends Study, as read from Exhibit 2 of Freedman et al., Health Affairs, 2026. We apply those rates to the Census Bureau's 2023 population projections. Then we scale today's workforce of about 5.4 million direct care workers (PHI, 2025) by the growth in care needed. That gives about 9.2 million workers needed by 2050.
2. Projected path. The workforce grows by the 772,000 jobs the government projects for 2024–2034, about 77,000 a year. We extend that pace past 2034, shown as a dashed line.
3. Record pace. From 2014 to 2024 the workforce added almost 1.8 million jobs, about 180,000 a year. PHI calls that record growth.
4. Point of no return. This is the last year America could switch from the projected pace to the record pace and still have enough care workers by 2050. By 2050, every surviving baby boomer will be 85 or older. With the Census main series, that year is 2032. On the projected path, America is about 1.8 million care workers short in 2050.
5. After the point of no return. Once even record-pace hiring cannot catch up by 2050, the gap closes only if each care worker can give more care. Tools like AI and robots would need to add about 24% more care per worker by 2050.

How much the answer moves:

• Census low-immigration series: 2033. Zero-immigration series: 2035. Fewer immigrants means fewer older adults too, which slightly lowers the care needed.
• A paid-care rate for people 85 and older that is 4 points lower (28%): 2035.
• If a quarter of care work serves younger people with disabilities, whose numbers grow slowly: 2042.

What the model leaves out. It holds care per person at 2022 levels, but residential care per person fell from 2011 to 2022. It assumes record-pace hiring could go on for decades, even though the working-age population barely grows and immigrants are about a third of home care workers. That makes the date generous, not alarmist. It does not count unpaid family caregivers, who help many more older adults than paid workers do. Higher pay, immigration and family caregivers can all help. They make the gap smaller, but they do not change its direction. The full calculation is in scripts/build-care-gap.mjs in our source code.

Your privacy

When you look up your representatives, your address goes to the Census Bureau's geocoder to find your district. We don't store it, log it, or use cookies or trackers.

Who runs this site

This is an independent project by Ben Caunt.

Questions, corrections, or a record we missed? Email bdcaunt@gmail.com. Please include a link to the source. We correct errors quickly and say so publicly: every addition, edit and removal is listed in our change log.