Every claim, cited

Not a citation bolted on after the answer is written. The citation is the reason the sentence exists.

Most AI tools generate an answer from what the model already believes, then attach sources that sound related afterward. That produces citations that are topically close but don't actually say what the sentence claims — a document about diuretics in general standing in for a specific dose that was never in it.

CiteWard builds the answer from the retrieved evidence first, and checks after: does the cited passage actually support this specific claim, not just the general topic. High-risk numbers — doses, thresholds, percentages, effect estimates, confidence intervals — get an extra check, because a source saying "risk increased" doesn't automatically support an exact number someone might act on.

When a claim can't be verified, it doesn't ship with a caveat bolted on next to an unverified number — the number comes out, and the supported part of the answer stays. See the full methodology for how that works end to end.