How Certavi's Verification Engine Works (No AI in the Loop)

When we tell attorneys that Certavi uses no AI in the verification loop, we mean it literally. The extraction pipeline uses a language model to find citations in brief text. The verification engine does not. Every verification status traces to a deterministic database query.

The process starts with extraction. A hybrid pipeline reads your PDF or DOCX and identifies every citation. Federal citations (U.S. Reports, Federal Reporter, Federal Supplement) are caught by regex patterns. State citations, Westlaw cites, and unusual formats are caught by a secondary pass using a private language model running in our isolated environment. Your documents are never sent to public AI systems like ChatGPT, Claude, or Gemini, and are never used for model training.

Verification is separate from extraction. For each extracted citation, the engine constructs a database query using the volume, reporter, and page number. That query goes to CourtListener, Westlaw, LexisNexis, or PACER depending on your configured source priority. The database returns a result or it does not.

The status assignment is deterministic. If the database returns a matching case with a case name that fuzzy-matches the name in your brief (above a 75% threshold), the citation is VERIFIED. If the database returns no match, it is NOT_FOUND. If the database is unreachable (timeout, error), the citation is UNVERIFIED. We never call a citation VERIFIED when the database is uncertain, and we never call it NOT_FOUND when the database might just be slow.

This conservative approach means some citations will come back as UNVERIFIED when a database is temporarily down. That is by design. A false negative (telling an attorney to check a citation manually) costs five minutes. A false positive (telling an attorney a fabricated citation is real) could cost them their career.

Every verification query is logged with the full database response, the timestamp, and the source. This audit trail is append-only. No one can modify a verification result after it is recorded. When an attorney generates a signed certificate, the certificate references these immutable records.