How we measure whether an answer holds up

What actually gets checked before an answer reaches you — and an honest note on what we haven't published yet.

Citation support

Does the cited passage actually address the claim it's attached to — not just the same general topic, but the specific thing being claimed.

Numeric provenance

Every dose, threshold, percentage, effect estimate, and confidence interval is checked against the retrieved source text specifically — a passage supporting "risk increased" does not automatically support an exact number.

Unsupported claims don't ship silently

A claim that fails verification isn't published with just a warning label next to it. It's either repaired against the evidence, softened to what's actually supported, or left out — not printed as fact with a caveat attached.

What we haven't published yet

We'd rather leave a number off this page than publish one we can't stand behind. Formal benchmark results — pass rates against curated clinical question sets, comparisons against other tools — aren't published here yet. When they are, they'll be published with the methodology behind them, on this page, not asserted without it. That's the same standard the product itself is held to: a claim without a checkable source doesn't get to look like a fact.

In the meantime, the checks described above run on every real answer, not just in evaluation. See our methodology for the full pipeline, and our sources for what gets retrieved from.