How the Bluff & Spin Index works
Bluff & Spin Index: how well claims hold up, and how often questions get a straight answer. 0 to 100. Lower is better. 0 means every checked claim held up and every question got a straight answer; the more bluff and spin, the higher the number.
“When they speak to the public, do they tell the truth and answer the question?”
We do not measure whether a politician is a good person, a good leader or right on policy. We measure what they say in public, against the evidence and against the question asked.
Bluff
0–50Claims that do not hold up against the evidence.
Spin
0–50Questions that do not get a straight answer.
Reading the score
| Score | Label |
|---|---|
| 0–25 | Mostly straight talk |
| 26–40 | Some bluff & spin |
| 41–55 | Frequent bluff & spin |
| 56–100 | Heavy bluff & spin |
0 to 100. Lower is better. 0 means every checked claim held up and every question got a straight answer; the more bluff and spin, the higher the number.
How we score
We take official records (Hansard, GOV.UK, select committee transcripts, the Daily Compilation of Presidential Documents and the Congressional Record), save the raw files with a SHA-256 hash, and split them into sentences and question–answer exchanges by fixed rules. Where there is no official record, we use the broadcaster's or another published transcript, or a video's captions. Captions do not say who is speaking, so for interviews we assign each turn to the interviewer or the politician by reading them. Each transcript page says which kind of source it is and how the turns were assigned.
Our model then labels each transcript: which sentences contain a checkable claim, how each question was answered, and which emotional passages stand in for substance.
Accuracy counts material claims only: claims that support an argument, a policy or an attack. Procedural claims stay on the page but do not count: the speaker's own diary and dates, petition counts, who said what in the chamber, background, scene-setting and jokes. When in doubt, a claim counts as material.
Two more kinds of claim stay on the page but do not count in Accuracy: a clear slip of the tongue, where the context shows the speaker meant something else, and words read out from someone else, such as a minister's written answer or a newspaper's account. Each claim that does not count is marked with the reason.
Everything is append-only: a correction is a new record, and the old one stays visible.
Claims are chosen for checking by a fixed rule, not by hand: every material claim (one that supports an argument, a policy or an attack) in the transcripts that count towards a score; in other transcripts, up to 10 claims spread evenly across each; plus any claim an emotional passage relies on. Each check needs at least one source independent of the speaker, and every source is linked. Sources come in four tiers: primary data and official records; recognised fact-checkers; established news organisations; and reference works such as Wikipedia, which can point to a source but never decide a verdict. A verdict needs a source from the first three.
Every rating says "Our model rated this". We describe statements, not people. Labels never use "lie", "liar", "dishonest" or "manipulative"; the build fails if they do. Every number links to the sentences and sources behind it. Scores were last updated 8 Oct 2026.