Method, briefly

• Claims: checkable statements asked by established pollsters, picked by a fixed quota before any data was seen. Metaphysical beliefs are never scored.
• Answers: graded true to false with an evidence grade. Strong evidence counts fully, mixed counts 40%, and unsettled counts zero, so it never moves a group.
• Placement: a group's distance is its average share of answers the evidence contradicts, across the claims. It is shown as a plain word plus the number, e.g. Mostly supported (20% not supported). "Not sure" answers are left out, and the number of them is shown for each claim.
• Labels:
• Article checks: An advocacy site is run by a group that pushes for a cause or a party. Pages on the advocacy sites listed below, and opinion pages on any site, never count as an original record. An original court filing or bill text hosted on one of them still counts, because we check the document itself.
• Outlet grades: Visitors choose which articles to check, so their checks alone don't set a grade. When at least 5 different articles from one outlet have been checked, we check a random sample of that outlet's recent news articles with the same rules. Only that sample can move the grade, and it moves one step at a time. The same link counts once. Outlets on every side follow the same rules. Every grade change is listed on the Changes page.
• What counts toward a grade: Only facts an original record backs up, and facts an original record shows are wrong. A fact we couldn't back up is not counted against the outlet: it may be original reporting.
• Paywalls: We can't read articles behind a paywall, so outlets with a paywall can't be graded either way for now.
• Locked wording: T01–T09 wording was locked at commit f7b55bb, B01 at 689aa49, the 12 claims added or reworded on 2026-09-30 at 7568fb8, RT4 and R12 at a5aba1c, and the ten rows graded on 2026-10-01 at d082928 and fb8efa2, before blind grading.
• Named holder: a claim with no qualifying poll is listed only if a real, dated person or group is on record saying it, and we cite them. If we can't find one, the claim comes off the list. We never make up a holder or a quote.
• Mirror claims: when we add a block of claims after the list was locked, we also add matching claims from the other side, worded the same way and locked before anyone grades them, as long as each one has a named holder.
• Waiting on blind grade: new and reworded claims keep their locked wording and holder but show no rating, and are left out of every score, until a separate grader who sees only the wording rates them.
• Re-grades: a claim changed after its answer was known is marked "Changed after review; not re-run blind."
• Balance: the main view uses equal numbers of left-coded and right-coded claims, averaged over every possible combination, so no one picked which claims count.
• Clouds: the width covers that combination range plus a resampling test of the claims. Party-level polls can't tell groups within a party apart.

Limits

• Only 9 claims have a qualifying poll with a party breakdown.
• Pollsters ask about right-coded misbeliefs more often than left-coded ones.
• One poll figure (FBI and January 6) comes from the published article, not the pollster's own table.
• Polls come from 2016–2023. Opinions may have changed.