How Litco catches hallucinated citations
Litco tests the field on LitigationBench. Asked to work from memory, most models tested invented at least one legal authority, and the worst fabricated several. Every legal AI product is built on models like these. Litco checks everything the Matter Agent produces, using verification code that runs outside the model.
Litco runs this check on every quote before the quote leaves a draft.
The quote check is ordinary software that compares the draft’s quotation with the source text, letter by letter. A lawyer who runs the check twice gets the same result both times.
Quotes
Litco compares each quotation in a draft, character by character, against the source document or the reporter text. If the source does not contain the quotation, Litco flags the discrepancy or repairs the draft.
Pin cites and form
Litco checks every page reference against the actual pagination of the reported case. Litco also holds each citation to Bluebook form, inline in briefs and memos and in footnotes in pleadings.
Treatment
The Matter Agent reads the citing decision twice, through two independent passes. Litco flags a case as overruled only when both passes agree.
Characterization
When a draft attributes a proposition to a case, Litco checks whether the cited passage supports it.
Absence
Before the Matter Agent reports that something does not exist in the record, the Matter Agent must search the full record.
Delivery
When a draft fails a check, the Matter Agent either repairs the draft or delivers the draft with the failure marked so the lawyer can see exactly what needs attention.
Working from memory alone, Litco’s self-hosted drafting model fabricated authority four times in fifty-nine tasks, and most of the models tested fabricated at least once. The models Litco selects for its products did not fabricate at all, whether working from memory, using the platform’s tools, or facing benchmark tasks designed to provoke fabrication. LitigationBench publishes every fabrication, regardless of model or setting. Any model that fabricates loses twenty points and becomes ineligible for that setting rather than for Litco products altogether.
Check a brief against the cases and the record.
Hand the Matter Agent a brief and ask it to check the citations. It resolves every authority against LitLex, Litco’s own case-law corpus, verifying each quotation against the reported text and each pincite against the opinion’s star pagination. It checks the cites to the record too, opening the document in your LitKit review database or your LitSpace case files and reading the page you cited. Most cite checkers do only the first half, because they have neither the evidence nor mastery of it.
Every citation is checked against the case it cites.
LitLex runs two independent checks on every cited case and flags the case as bad law only when both checks agree. That requirement raised the rate of correct red flags from 2.8 percent, with a single check, to 98.7 percent on the benchmark set. LitDraft compares every quotation in a draft, character by character, against the reporter text and the documents in the record. LitDraft keeps a finished document in its drafting sandbox until every quotation passes. When the lawyer cites a case Litco has never seen, LitDraft reports a blank result so the lawyer can spot that gap before filing.
Litco tests every model twice.
Litco tests every model twice on the same litigation tasks through LitigationBench, once with no help and once inside the platform with the verification stack active. Litco publishes each model’s score alongside the difference between the two runs. In a blind indistinguishability test, independent AI models from several vendors read appellate introductions and tried to distinguish the platform’s drafts from counsel’s filed versions. Those models identified the machine-written version at near-chance rates.
Limits of verification
Litco checks every quotation, citation, treatment, and characterization against the cited source. A weak argument stays weak afterward, and when no document supports a claim, your team still decides what to do with that claim.