Source-grounding
Last updated: July 26, 2026
Source-grounding is checking a claim against a specific, independent source and confirming that source genuinely supports it. It is grounding made concrete: not just 'is there evidence?' but 'does this exact source say this exact thing?' It is what turns a plausible sentence into a verified one.
What is source-grounding?
Source-grounding is the concrete act of tying a single claim to a specific independent source and confirming that the source actually supports it. Where grounding is the general principle, source-grounding is the operation you can point at: this claim, this source, this passage, this verdict.
It answers two questions in order. First, does an authoritative independent source exist for this claim? Second — and this is the step naive checkers skip — does that source actually say what the claim says? A source can exist and still fail the second test.
Source-grounding is what lets a verifier show its work. Every verdict comes with the source it was grounded against, so you can click through and judge for yourself instead of trusting a black box.
Why independence and reading matter
Two failure modes make weak grounding look like real grounding. First, dependent sources: grounding against the model's own output, or a source it was told to cite, is circular. Real source-grounding uses sources the model did not control.
Second, existence without support: a citation can point to a real paper that never makes the claimed point. This is the market-correcting truth that a real citation is not a supported claim. Source-grounding reads the source and checks the match, so a real-but-irrelevant citation is caught. See citations vs claims.
Mephistopheles source-grounds across independent channels — semantic retrieval, live web, Wikidata and Wikipedia, and tools — and a tiered judge weighs what they return before issuing a verdict. See how it works.
A worked example
Claim: "Aspirin was first synthesised by Bayer in 1897." Source-grounding pulls independent references on the history of aspirin, finds agreement, and attaches the source to a supported verdict — with the link, so you can verify.
Now: "A 2021 trial proved drug Y cures condition Z." The trial exists but reports a modest, non-conclusive effect. Source-grounding reads it and returns contradicted for the word 'cures', because the specific source does not support the specific claim. Existence was not enough.
Frequently asked questions
How is source-grounding different from grounding?
They are the same idea at different resolution. Grounding is the general principle of tying a claim to evidence. Source-grounding is the concrete operation: this specific claim, checked against this specific independent source, with the source attached to the verdict.
Why isn't a matching citation enough to ground a claim?
Because a source can exist yet not support the sentence. A model can cite a real paper that never makes the claimed point, or overstate a modest finding. Source-grounding reads the source and checks the specific match, catching real-but-irrelevant citations. See citations vs claims.
What if no source can be found?
Then the honest verdict is unverifiable. Source-grounding never invents support; when no independent authoritative source exists, it flags the claim for human review instead of guessing.
Verify what your AI just told you.
Paste any AI answer and Mephistopheles checks each claim against independent sources — no sign-up to try.