Verdict taxonomy
Last updated: July 26, 2026
The verdict taxonomy is the set of four labels a fact-checker can assign to a claim: supported (a source backs it), contradicted (a source refutes it), disputed (credible sources disagree), and unverifiable (no authoritative source found). Four verdicts, not two, because reality is rarely just true or false.
What is the verdict taxonomy?
The verdict taxonomy is the fixed set of outcomes a verifier assigns to each extracted claim. Mephistopheles uses four: supported, contradicted, disputed, and unverifiable.
Two verdicts — true or false — would be simpler but dishonest. Real claims include ones where credible experts genuinely disagree, and ones where no authoritative source exists at all. Collapsing those into true/false forces a guess and inflates the false-alarm rate. Four verdicts let the system tell you not just whether a claim holds, but what kind of uncertainty surrounds it.
The verdicts also feed the hallucination-risk score, weighted by how much each signals real risk.
The four verdicts, each with an example
| Verdict | Meaning | Worked example |
|---|---|---|
| Supported | An independent source confirms the claim. | "The speed of light is about 299,792 km/s" — physics references agree. |
| Contradicted | An independent source refutes the claim. | "Measured from base to peak, Mount Everest is the tallest mountain on Earth" — sources say Mauna Kea is taller by that measure. |
| Disputed | Credible sources genuinely disagree. | "A daily glass of wine is good for your heart" — studies conflict, no consensus. |
| Unverifiable | No authoritative independent source found. | "The founder said in a private meeting that X" — nothing public to check. |
Note the difference between contradicted (evidence says no) and unverifiable (no evidence either way). Treating 'no evidence' as 'false' is a classic verifier error the taxonomy is designed to prevent.
Why four verdicts protect accuracy
The four-way split is not decoration — it is the mechanism that keeps the checker honest. Disputed stops the system from picking a side on genuinely contested facts and calling it settled. Unverifiable stops it from guessing when evidence is absent. Both directly protect the false-alarm rate: the system only says contradicted when a source actually refutes the claim.
Every verdict ships with the source it was grounded against, so you can inspect the reasoning rather than trust a label. See how it works and how accurate it is.
Frequently asked questions
What are the four verification verdicts?
Supported (an independent source confirms the claim), contradicted (a source refutes it), disputed (credible sources genuinely disagree), and unverifiable (no authoritative source was found). Four verdicts instead of true/false because many real claims are contested or uncheckable, and forcing them into two categories would mean guessing.
What's the difference between contradicted and unverifiable?
Contradicted means a source actively refutes the claim. Unverifiable means no authoritative source was found either way — the claim isn't shown false, just uncheckable. Confusing the two ('no evidence, so it's false') is a common error the taxonomy prevents. See unverifiable.
What does 'disputed' mean as a verdict?
Disputed means credible sources genuinely disagree — there is no settled answer. It's reserved for real controversy, not for claims the checker is merely unsure about (those are unverifiable). It stops the system from picking a side on contested facts and presenting it as settled truth.
Verify what your AI just told you.
Paste any AI answer and Mephistopheles checks each claim against independent sources — no sign-up to try.