Every generic legal AI tool shares the same architectural failure: the model can cite authority it never looked up. It draws from training memory — a statistical blend of cases, statutes, and law review articles that may be misquoted, mischaracterized, or entirely fabricated.
For a litigator, that is not a convenience problem. It is a malpractice problem. A citation that sounds right but cannot be verified in the record is worse than no citation at all — it creates false confidence.
Retrieval as a hard boundary
Lexora inverts the default. Before the agent can cite a case, statute, or regulation, that source’s text must be retrieved into the session corpus — from CourtListener, eCFR, Congress.gov, or another connected API. The synthesizer operates only on that corpus.
If a search returns zero results, the report says so. Explicitly. With the query terms and the API that was searched. Lawyers can act on a confirmed gap. They cannot act on a plausible fiction.
Quotes, not paraphrases
Every holding in a Lexora report includes a verbatim excerpt from the retrieved text. Not a summary that drifts from the source. Not a paraphrase the model prefers. The actual language, shown to the user, tagged as retrieved.
Verification as a backstop
Even with retrieval-first architecture, a second pass matters. After the draft, Lexora runs a verification step that checks each citation against the session corpus and flags anything it cannot match. Confidence is scored per citation and per section.
This is not a feature list. It is a design principle: the competitive position of any serious legal AI product rests on one rule. Cite only what you retrieved. No exceptions.