How we make this

Everything on this site that is a fact — a bill's text, its sponsors and cosponsors, every action taken on it, every recorded vote and how each member voted, committees, nominations, the Congressional Record itself — is copied from what Congress and the government publish, and stored as published. No AI writes, edits or scores any of it.

It comes from five places:

• congress.gov, the Library of Congress's record of Congress — bills and amendments and every action on them, members and their terms, committees and their reports and hearings, nominations and treaties, the House's recorded votes, and the laws.
• govinfo.gov, the Government Publishing Office — the full text of each day's Congressional Record.
• clerk.house.gov — each House recorded vote, member by member, and whether it needed a simple majority or two-thirds to pass.
• senate.gov — each Senate recorded vote, member by member, including the Vice President's vote when the Senate splits evenly.
• The Census Bureau — the file that says which congressional district each ZIP code falls in, which is how a ZIP code finds the people who represent you. About one ZIP code in six crosses a district line; when yours does, we say "probably" and point you to house.gov's lookup by street address.

Every response we fetch is kept exactly as it arrived, so any fact here can be traced back to what was published — and rebuilt from it.

A few things you see are worked out from those facts by fixed rules, not by AI. Where a bill stands — passed the House, waiting on the Senate, became law, never left committee — is read from its official actions. A vote's tally and how each party split are counted from every member's recorded position, using the party the chamber recorded for that member at that vote, so a member who later switched parties is not refiled.

A copy of Congress is only worth reading if nothing is missing and nothing is stale, so both are checked by programs that report every gap they find, rather than taken on trust.

• Kept current on a schedule. New and changed bills and amendments are fetched every 10 minutes, recorded votes every 30 minutes, nominations every two hours, the Congressional Record every six hours, and the House floor schedule daily.
• Recent days are read again. A chamber sometimes corrects a vote after the fact, and a day's Record is published the day after it. So Senate votes from the last 7 days are re-read on every pass (a House vote is re-read whenever the House marks it changed), and the last 7 days of the Record are checked again for anything new.
• Our copy is compared with Congress's own lists. A check asks congress.gov — and senate.gov, for Senate votes — which records exist, and names every one we are missing, every one we hold that they no longer list, and every one that has changed since we stored it.
• Stored records are re-read against the source, field by field. Matching lists cannot tell a record stored correctly from one stored wrong or half-empty. So a second check takes a sample of records back out of our database, fetches each from the source again, and compares every field, including the lists inside a record — sponsors, actions, how each member voted. It works out what each field should be straight from what was published, not from the code that stored it, so a mistake in that code cannot agree with itself.
• An unfinished record says so. A record is marked complete only after every part of it has been stored. One whose storing was interrupted is picked up again on the next pass instead of looking finished.

On top of those facts, AI writes the plain-language layer: what a bill would do, who it would help and who it would cost, and an overview of each day in the Congressional Record. It is always labelled as AI-written, and the record it was written from is always one click away.

• Bill and amendment summaries. "What this bill would do" and "Who it would affect, and how" are written from the bill's own latest text, or from its committee report when that is what has been published. A bill too long to read in one pass is read in parts and the parts merged, and the page says when that happened.
• Scores. Each bill is scored from -10 to +10 on more than a hundred questions like "Would veterans be better off under this?", with a one-sentence reason for every score. The "who it helps" band on each bill, a member's "What their bills would do" and the assistant's answers about who a bill helps are all built from them.
• Policy area. When congress.gov has not given a bill a policy area, AI picks one from congress.gov's own list.
• The Congressional Record overview. The summary of each day's House and Senate proceedings is written from that day's published Record.
• The assistant. When you ask a question, the numbers, lists and charts in the answer come straight out of the database, through the lookups shown with each answer. AI chooses which lookups to run and writes the sentences around what they return.

AI never writes a fact: not a vote count, a date, a sponsor, or where a bill stands. The AI that summarises and scores a bill is given the bill's own text — not the sponsor's party, not how anyone voted, not what anyone has said about it — and asked what the bill would do and to whom, never whether it is good.

Which AI does the work is a detail of how the site is built, and it changes. What it is asked to do, and how that is checked, is what this page describes.

A score answers one yes-or-no question about one bill, from -10 to +10, and it always comes with the reason it was given.

• Plus is yes; minus is the opposite. A +8 on "Would veterans be better off under this?" means the bill would clearly leave veterans better off; a -3 means it would leave them somewhat worse off. The number says how strong and direct the effect is — not whether it is a good idea.
• Read the reason. The sentence beside a score is the part to judge. If it misreads the bill, the score is wrong, and we want to hear about it.
• No score means no effect was found. A bill is scored only on the questions it touches. No score on a question means the bill was not found to affect it either way — which is not the same as being found neutral.
• Scores are never added up. Each answers a different question, so +6 for veterans' health care and -6 on the national debt do not cancel to nothing. We never show one total for a bill, and never one for a member or a party.
• A short list says how long the full one is. "3 of 45" means the bill has 45 scores on that side and you are seeing the strongest three.

A new question does not appear on the site until it has been measured against bills whose answers were settled first, and until every bill has been asked it.

• Measured against a labelled set of bills. The set is chosen to put the question where it is easiest to get wrong: bills squarely on the subject, bills that only mention it in passing, near misses on neighbouring subjects, and a random handful of everything else. Each bill is labelled with a direction and a strength under a written rubric. The labels are drafted with AI help, without sight of the answers being tested; a sample is labelled twice, independently, to check the labels agree with each other; and the cases a reasonable reader could decide either way are settled by written rulings the site's owner makes.
• The worst mistakes are counted first. Answers pointing the opposite way from the label, which mislead a reader most; bills with a strong effect that the question missed; bills the labels call unaffected that were scored anyway; and whether the same bill gets the same answer when asked twice. Every disagreement is read bill by bill, not only counted: a number says how often, and only the bill says who was right.
• A narrow question gets a note. When a question's wording cannot carry where its edges are, the AI is given a short note beside it — for example, that making it easier to build anything is not the same as making it easier to build data centers. You see the question; the note only keeps the scoring on it.
• Hidden until complete. A new question stays off the site until every bill has been asked it, because a half-asked question shows a short list that reads as "few bills do this" when it means "we have not asked yet".
• Rewording is re-asking. Changing a question's wording, or its note, re-asks it of every bill and measures it again. When the way questions are asked changes, it is measured by asking bills that already have answers and comparing the two.

The feed is ordered by a fixed rule about what happened, never by what people click. Every post can say why it is there, under "Why am I seeing this?".

• What happened counts most. A law signed or vetoed comes first, then a bill sent to the President, then a bill passing the House or Senate, a vote kept from happening, or a nominee confirmed or turned down.
• A close vote is news. A vote decided by less than 10 percent of the votes cast is shown to everyone, whatever the bill is about.
• So is a bill that would change a lot. Each bill's scores are added up by size, ignoring whether they point yes or no. A law scoring at least 15 is shown to everyone. A law scoring under 10 (naming a post office or a monument) is shown only to people following it.
• A nominee is ranked by the office. The Cabinet, the Supreme Court and the head of the Federal Reserve are shown to everyone. Appeals court judges and the heads of major agencies are mixed in. Other offices are shown only to people following them.
• Newer comes before older. The feed covers the last 30 days. When Congress has been away it reaches back 90 days instead, and says so.
• What never appears. A bill being sent to a committee, a cosponsor being added, and the routine votes that come before a bill is debated are left out. They are part of a post's story or nothing at all.

Nothing about readers goes into the order: not clicks, not shares, not how many people follow something. A feed ranked by attention ends up ranking whatever provokes people most, and this site describes rather than advises.

The site describes; it never advises. Being non-partisan is built into how the record is shown, not claimed on top of it.

• Both sides of every score. Wherever scores appear, the ones pointing each way are shown together. A bill that mostly helps still costs someone something, and a page that showed only the help would be taking a side.
• Questions, not categories. Every score answers a question that names who or what and which way — "Would veterans be better off under this?", "Does this give the federal government more power?" — so a number is an answer to something you can read, not a verdict on a topic.
• No totals, no grades, no report cards. We do not add a bill's scores into one number, grade members, or score parties. A grade needs a standard for what a member ought to have done, and choosing that standard is a political position. A single number also hides what it is made of: "0 of 32 bills passed" describes nearly every member of Congress, because most bills never get a vote.
• The record as it was. A member's votes are shown as recorded, with the party the chamber recorded at that vote. The votes that decided something come first, and every procedural vote stays listed and counted.
• You choose what matters. You can narrow a member's record or a list of bills to a question you care about. The site never ranks a member against your interests or suggests who deserves your vote.
• No advice. Ask the assistant who to vote for, or which party is better, and it says it does not do that — then shows you what they have actually done.

Some things are left out on purpose.

• We don't tell you how to vote, endorse anyone, or say which party is better.
• We don't grade members or parties, or add scores up into a verdict.
• We don't predict whether a bill will pass.
• We don't edit what Congress published, and a person does not review what the AI wrote before it appears.
• We are not an official source. Nothing here is a Congressional Research Service summary or the position of Congress; the official record is on congress.gov, and every bill links there.
• We don't run ads, have data partners, or sell or share anything about you.

Treat the AI-written layer as a careful first reading, not an official one. It is not a Congressional Research Service summary, it is not reviewed by a person before it appears, and it is not the position of this site. It can misread a provision, miss one, or weigh it differently than you would — the reason beside each score is there so you can judge that for yourself. The bill's own text is always one tab away, and it is the thing to read before deciding anything.

When the way we ask changes, summaries and scores are written again, so what a bill's page says can change without the bill having changed.

We do not use personal information of any kind, and the site is built so there is very little of it to use: no account of our own, nothing about you sent to the AI, and nothing sold or shared. What is kept — your conversations with the assistant, what you follow, and anything you send us — is kept to serve you and to fix what we get wrong.

• No account of our own. Signing in uses your Google account, and is needed only for the assistant and for following members, bills and questions. We hold no password and no profile — what is kept is a sign-in id, and filed under it your conversations, what you follow and how often you want to hear about it. Your name and email are never written to our database.
• Alert emails go to your Google address, looked up when we send. If you follow something, we ask Google for your account's email address at the moment a digest goes out and hand it, with the email, to our email delivery service — which sends it and does nothing else with it. We do not keep a copy. Every email has a link that stops them, no sign-in needed.
• Nothing about you goes to the AI. When you ask the assistant a question, the AI sees your question, the earlier turns of that same conversation so it can follow along, and the site's public data — bills, votes, members. Nothing else: not your name, not your email, not your sign-in id. A new chat starts fresh; nothing carries over from the last one.
• Your conversations are kept, so you can come back to them. Each question and answer is saved against your sign-in id, which is what lets a refresh, the Back button or a visit next week return you to where you were. Delete one from the Ask page and it is gone, along with everything it showed. We read conversations — especially answers marked unhelpful — to find and fix what the assistant gets wrong, and for nothing else: not to build a picture of you, and never to sell. Anything you type is kept as you typed it, so leave out what you would not want stored.
• Sharing is your choice, and a snapshot. Nobody else can see a conversation until you press Share. The link shows it as it stood at that moment; anything you ask afterwards stays private unless you share again, and "Stop sharing" takes the link down. We count how often a shared link is opened, not who opened it.
• Apps you connect are not remembered. Questions asked from an app you have connected to this site are answered and not stored.
• Feedback and reports are what you type. "Send Feedback" and "Report a problem with this" store your message in our own database — nothing goes to an outside service — with the page you were on, your browser type, what you were reporting, your sign-in id if you are signed in, and your email only if you choose to give it.
• Nothing is sold or shared. We have no advertisers and no data partners. Nothing you do here is sold, traded or given to anyone.

Every AI-written summary and score has a Report a problem with this link beside it. It opens a short form that already knows which bill or Record day, and which summary or score, you mean — so all you need to say is what is wrong.

• Say what is wrong, and why. Quote the sentence or name the score if you can, and say what the bill's own text says instead. Leave an email if you would like to hear back; it is optional.
• Every report is read. Each one is marked addressed with a note saying what was done, and reports about the same bill are read together.
• The fix goes where the mistake came from. We do not hand-edit what the AI wrote. When a report shows it misread something, the fix goes into how it reads — the question, its note, or the instructions it is given — and the summaries and scores it affects are written again.
• For anything else — a broken page, a wrong fact, an idea — use Send Feedback at the bottom of every page.