How BioSniper verifies every number

Every AI-generated statement on BioSniper must cite the source it came from — inline tags like [sec], [clinical] or [fda] point at the document the model actually read.

Quoted evidence is held to a stricter bar: quotes must appear verbatim in the source text. After the model answers, an independent verification step re-checks every quote against the retrieved documents, character for character. A quote that cannot be found is discarded before it ever reaches a page — regardless of how plausible it sounds.

And when the evidence simply is not there, BioSniper answers UNKNOWNinstead of guessing. Missing data is shown as missing — never estimated, never interpolated, never filled with a model’s best guess.

The citation-verification pass rate shown above is computed live from our verification logs over the past 7 days — it is a measurement, not a marketing claim.

Ingesting data is the easy part; keeping it correct is not. A daily automated audit sweeps the entire dataset across more than a dozen correctness dimensions — freshness, duplicates, referential integrity, misattribution (a real number attached to the wrong company), fabricated-success patterns, field quality, and coverage across every actively listed company.

Failures page the team automatically. When an upstream source silently changes its format — regulators do this without notice — these checks are how we find out before you do.