How we know what we publish

1. How your ballot is resolved

Ballot resolution is deterministic. A language model is never asked which ballot an address gets. The pipeline is:

• Your address is geocoded to a coordinate.
• That coordinate is matched against official U.S. Census Bureau TIGER/Line boundary files to determine the county, incorporated place, school district, congressional district, and state legislative districts containing it.
• Those districts are combined with each state’s statutory election schedule to produce the contest list.

Every step is reproducible from public data. A model is used only to explain an answer that has already been determined — never to produce one.

2. What we assert, and what we don’t

Not every contest can be stated with the same certainty. Federal and state schedules are fixed in law and follow from your districts alone. Local contests are staggered differently in thousands of jurisdictions, so we mark them accordingly rather than guessing:

• On your ballot — follows directly from federal or state law and your resolved districts.
• Likely on your ballot — usual practice, but the cycle is set at a level we have not verified for your jurisdiction.
• Varies locally — the office exists in your jurisdiction, but whether a seat is up this cycle depends on local staggering. Check with your election authority.

Marking uncertainty is a feature. A guide that presents a guess as a fact is worse than one that admits the limit.

3. Source hierarchy

Every substantive claim is tied to a source, and sources are weighted by how directly they document what happened:

4. Claims are kept separate from interpretation

A claim is a discrete, sourced statement: this officeholder voted for this bill on this date. An interpretation is what a claim might suggest. We store and display these separately, and the interpretation never becomes the record.

This matters because interpretation is where bias enters a voter guide. Keeping it structurally distinct makes it visible and checkable rather than baked into a summary.

5. Confidence scoring

Every research output carries a confidence level:

6. Neutrality is enforced structurally

Neutrality policies that live only in a document tend to erode. Ours are built into how the product works:

• Candidates are always displayed in ballot order. Never sorted by fundraising, polling, popularity, or any internal score.
• There is no candidate score. No hidden ranking exists to sort by. This is a schema decision, not a policy promise.
• Explanations are generated from fixed templates over reviewed reference material, not from open-ended prompting.
• Party affiliation is never used as a substitute for research, and is not used to infer any candidate’s position.
• Unfavorable information is not omitted. Where a documented criticism exists for one candidate, we look for the equivalent for their opponents.
• Our color palette avoids red and blue. A small thing, but a nonpartisan product should not wear a party’s colors.

7. Untrusted content

Every retrieved web page — including a candidate’s own site — is treated as untrusted data, never as an instruction. A page that contains text addressed to an automated system is processed as page content and nothing more.

Retrieval is kept separate from generation for this reason. The explanation layer on this site works only from our own reviewed reference material and the deterministic resolver output.

8. Sponsorship and advertising

Sponsors and advertisers have no influence over which candidates appear, how they are described, how they are ordered, or what our research concludes. There is no mechanism by which they could — ordering is fixed and no score exists to adjust.

All advertising is visibly labeled and visually distinct from editorial content. No ad unit is placed inside a candidate comparison, a confidence disclosure, or between a claim and its source.

9. Human review

Generated research is not automatically published. Candidate profiles pass automated checks — identity, district, party, source coverage, neutrality of language, ballot-order verification — and then human review before going live, with priority given to statewide races, major measures, and any low-confidence profile.

10. When we get something wrong

We will. Election data changes constantly: candidates withdraw, districts are redrawn by courts, filings are amended, local schedules shift.

The most serious failure this product can have is showing someone the wrong ballot, which is why every page links to the official election authority and why we mark uncertainty rather than smoothing over it. Corrections are made against the underlying source record, not by editing a summary.

11. What we are not

We are not an election authority, and nothing here is legal advice about your eligibility to vote. Your state or county election office is the final word on your registration, your polling place, and what appears on your ballot.