A fabricated citation is a sanction or a fine. Draco holds the authoritative source data and checks every claim your AI makes against it, literally.
Book a callPrimary sources, pulled from where they live: websites, logins, PDFs, scanned documents. We parse them into records, keep them current, and store the raw original.
Send a claim, quote, or citation to one API. We match it against the real source text, character for character. A model checking a model shares the blind spot. See how it works.
What was checked, against which source and version, the verdict, and when. Your reviewer signs the flags. The record shows the work if anyone asks.
AI output sounds right whether or not it is. Draco checks what your AI produced against the actual source data, not another model's opinion, so what ships is grounded in something real.
A claim, a quote, a citation, a figure: your software sends it to our API before it reaches your users.
Literally, against the source text in our dataset, not a model's guess about what sounds plausible. If the quote or number isn't in the source, it doesn't pass. That's how we catch fakes a model‑check would miss.
Verified, flagged, or needs‑review, each with a link to the raw source and when it was pulled, so your product can show exactly why an answer can be trusted.
Sources change constantly, and the check has to be right the day they do. In‑house that's a permanent team and permanent exposure. We carry that; your professional still signs.
The rule that changed last week is the one that gets you sanctioned. Scrapers run on a schedule, records are versioned, and every check names the version it matched.
We store the raw source as pulled, stamped with its origin and pull time. Every verdict links back to the exact text it was matched against, not to our word for it.
We scrape hard‑to‑get source data, structure it into clean datasets, serve it through an API, and check your AI's output against it. So a fake citation, wrong number, or made‑up claim gets caught before it reaches the person who signs off.
AI companies in regulated fields, legal, finance, insurance, healthcare, tax, whose output a professional has to review and stand behind. If being wrong costs you a lawsuit, a fine, or a lost client, we're for you.
A model checking a model hallucinates the review the same way it hallucinates the answer. We match every claim against the real source text, deterministically. It either matches the source or it gets flagged.
No. A licensed human always makes the final call. We flag only what's actually wrong, so they read the exceptions instead of everything.
Clean, current datasets from sources that are painful to collect, API access to query and verify against them, a verified/flagged result on every claim with a source link, and an audit record of every check.
