What this site does
Videos and blog posts make claims. Most are never checked. This site collects the sources on a subject, pulls out the distinct things they claim, and judges every claim against the best evidence available. The result is a page per subject that says which claims hold up, which do not, and which sources are worth your time.
How a subject is built
- Collect. YouTube videos with captions and blog posts, chosen to cover the pitches, the critics and the explainers on a subject, not just one side.
- Extract. From each source, the four to twelve claims a viewer would walk away believing, each tied to a short verbatim quote. Claims are written once in canonical form, so the same idea from twenty videos is one claim with twenty sources behind it.
- Guess. Every new claim gets an AI judgment with a one-line reason. It counts for at most 50% confidence, however sure the model sounds.
- Check. Studies, official data releases and reports go in a sources folder. Each is read against every claim, and a verdict is recorded only where a passage bears directly on the claim as worded, with a line citation. A study verdict replaces the AI guess when it is at least as confident.
- Score. Evidence belongs to the claim, not the source. A verdict on a claim counts for every source that makes it, including ones added later. Each source's score comes from the verdicts on its claims.
How to read a verdict
Each claim carries one of five verdicts. Right, wrong and partly right are judgments with a confidence from 0 to 100%. Untestable means framing, taste or a value judgment with no operational definition. Unknown means testable, but the available evidence cannot settle it yet, including predictions.
The label puts the confidence into words. The adverb comes from how sure the deciding evidence is; the colour comes from the verdict and fades as confidence drops. A leans label with a dashed border is an unverified AI guess or a weak study: it shows which way the evidence points, not a settled answer.
| Confidence | Judged right | Judged partly right | Judged wrong |
|---|---|---|---|
| 90 to 100% | Almost certainly right | Almost certainly partly right | Almost certainly wrong |
| 75 to 89% | Very likely right | Very likely partly right | Very likely wrong |
| 60 to 74% | Probably right | Probably partly right | Probably wrong |
| under 60% | Leans right | Leans partly right | Leans wrong |
| no judgment | UntestableUnknown | ||
Confidence measures how strongly that evidence settles that claim, not how reputable the publisher is. A direct reading from an official release scores high. One econometric study scores lower, and lower still when a claim reaches beyond what the study measured. Hover a label to see the raw verdict and percentage.
How sources are scored
A source's score averages its claims: right counts 1, partly right counts half, wrong counts minus 1, and the sign flips when the source argues against a claim. The average maps to four labels: mostly wrong,mixed, mostly right and right. Mixed means the source's correct and incorrect claims roughly cancel, or it only makes half-true ones.
Novelty breaks ties. It rewards a source for claims that hold up and that few other sources make, weighted by how sure the verdict is. A video that repeats what twenty others say scores the same but ranks lower than one that adds something sound.
What it does not do
- It does not judge people or politics. Each verdict applies to one claim, in one period, against one body of evidence. Claims whose truth changes over time carry a period in brackets, and the same claim in another period is a different claim.
- It does not force a verdict. Predictions stay unknown until the data arrive. Value judgments stay untestable.
- It does not treat an AI guess as evidence. The guess is a starting point that a study can overturn, and the Recap tab of each subject lists every claim where that happened.
Who
Built by Jonathan Borchardt, who writes at Always Shippable. The data, the agent briefs and this site are open at GitHub. Every claim links back to its quote and every study verdict to the lines it rests on, so you can check the checking.