Editorial transparency

Known bias: Self-promotional by definition. Well-funded campaigns produce polished sites; under-funded campaigns may have sparse pages or none at all.

Our mitigation: We rank website content first because a candidate’s own framing of their candidacy is the most authoritative source for voters. Source labels are always visible, and candidates without a website get a dignified empty state rather than fabricated content.

Known bias: Coverage skews toward well-known incumbents. Articles for women, people of color, and third-party candidates have historically been shorter or less comprehensive.

Our mitigation: Wikipedia fills the gap only when no website bio exists. We display the extract verbatim (never paraphrased) with CC BY-SA 4.0 attribution so voters can click through and judge the source article themselves.

Known bias: Late entries, write-ins, and withdrawn-then-reinstated candidates can be under-represented. Unusual name formatting can cause matching failures.

Our mitigation: We treat every unmatched candidate name as a real problem, not background noise, and investigate each one. Name-matching logic is tested against edge cases (hyphenated names, quoted nicknames, multi-part names).

Known bias: Language models inherit biases from their training data. Stance classification can be more confident for incumbents because more training data is available about them.

Our mitigation: Our summarizer is grounded: it only processes source material we collected, never generates from training data alone. Candidates with no source material receive no LLM-generated content. Provenance chips below every summary let you verify what was read.