Who Is Building Their Own Citator?

· LawDiver News Tracker

Filevine LOIS, Paxton, vLex Cert, Descrybe Cytator, Midpage, Free Law Project, and others building citators.

Who Is Building Their Own Citator?

LawDiver’s CiteDiver is our citator, built from the ground up on our own corpus — not a licensed extract of Shepard’s or KeyCite — with no restrictions on use. A live example shows the CiteDiver operation on a real brief; the same stack is available through the caselaw API.

Shepard’s and KeyCite defined “is this still good law?” for generations. A growing set of companies — legal tech vendors, research startups, and open-source projects — are now building their own citators, often AI-native, and often with claims that graph-only tools miss silent or semantic conflicts. This living post aggregates that coverage. Every product claim and benchmark number below comes from the cited reporters and vendor publications; LawDiver did not independently re-run those evaluations.

Why citators became the “holy grail”

Bob Ambrogi has repeatedly called a reliable citator the holy grail of legal research platforms — the feature that turns a case corpus into something lawyers will trust for filing decisions (LawNext on vLex Cert and Paxton, 12 July 2024; LawNext on Midpage, 10 Feb 2025).

Academic baselines keep showing up in these product stories. Paul Hellyer’s 2018 Law Library Journal study of 357 citing relationships labeled negative by at least one of Shepard’s, KeyCite, or BCite found all three agreed only 53 times — about 85% disagreement on whether negative treatment existed — with overall failure rates Hellyer calculated at 33% (Shepard’s), 38% (KeyCite), and 72% (BCite) (as summarized in Ambrogi’s Filevine coverage and Paxton’s citator announcement). Susan Nevelow Mart’s work on citator overlap is likewise cited for the finding that Westlaw and Lexis citator overlap on relevant results can be as low as about a third (LawNext on Filevine, 2 Sept 2026; Paxton AI Citator announcement).

Filevine LOIS: passage-level citator + brief auditor

Ambrogi’s LawNext report is the clearest independent write-up of Filevine’s September 2026 launch: an AI-native citator and companion anti-hallucination / brief-checking tool inside LOIS (Legal Operating Intelligence System). Per that reporting and Filevine’s own releases:

  • The citator starts from a user-highlighted passage, then runs citation-graph traversal plus semantic similarity search, handing candidate opinions to LLMs to distill conflicts — including sub silentio conflicts where later courts never cite the earlier case (LawNext; Filevine / PR Newswire, 1 Sept 2026; Filevine June 2026 release).
  • Filevine’s unpublished evaluation paper, as described by Ambrogi, started from 1,314 matters with judicially identified hallucination concerns, narrowed to 68 federal attorney filings, and checked 2,073 citations: 61.8% verified, 38.2% flagged for review, 8.4% marked severe (LawNext).
  • Filevine’s PR states its legal team confirmed 174 severe errors attorneys had missed in that sanctioned-filing sample, and claims a head-to-head where LOIS caught 14/14 “silent” negative treatments while KeyCite and Shepard’s caught none — a vendor-reported comparison, not an independent lab study (PR Newswire; critical context in US Tech Automations’ LOIS write-up).
  • Ambrogi notes Filevine’s corpus is built on CourtListener plus licensed reporter pagination, and that statutes/regulations are not yet included (LawNext).

Paxton AI Citator

Paxton announced a patent-pending Paxton AI Citator, reporting 94% accuracy on the Stanford CaseHOLD overruling/upholding task (precision/recall tables published on its site) and 86% accuracy on an internal attorney-reviewed caution/good set. Paxton positions the product as grouping cases by similar legal principles even without direct citation, and offering explanatory treatment text rather than flags alone (Paxton; Ambrogi context in LawNext, July 2024).

vLex Cert

Ambrogi reported that vLex’s nationwide Cert citator finished development in May 2024 and blends AI with human editorial review. Per that reporting, vLex found AI could tag relationships at high confidence only about 60% of the time; editors in Charlottesville spent two years reviewing more than 700,000 references for the remainder (LawNext, 12 July 2024).

Descrybe Cytator

In June 2025 Ambrogi covered Descrybe.ai’s paid Legal Research Toolkit and its AI-driven Cytator, which flags treatment on search results and adds a backward citator showing how an opinion treated the authorities it cited — a direction Ambrogi noted other new citators generally omit (LawNext, 25 June 2025).

Midpage

Ambrogi reported Midpage preparing an AI citator with claimed accuracy comparable to Shepard’s/KeyCite and an independent review planned with Vanderbilt’s AI Law Lab; Midpage’s own blog frames AI citators against Hellyer’s findings and positions its product as a forthcoming alternative (LawNext, 10 Feb 2025; Midpage — Future Citators).

Free Law Project / CourtListener

Free Law Project has published progress reports on an open-source AI citator for CourtListener: eyecite anchors citation spans, LLMs analyze surrounding context for negative treatment, and an expert-annotation project is building a public benchmark dataset focused on negative treatments (FLP, 1 May 2025; FLP annotation update, 30 Sept 2025). CourtListener already ships citation lookup/verification APIs that can help catch fabricated cites (FLP wiki — citation lookup).

Reading the claims carefully

Several themes repeat across this coverage:

  1. Graph ≠ proposition. Passage-level and semantic approaches (Filevine, Paxton’s similarity grouping) try to answer “is this holding still good?” rather than only “was this case cited negatively?”
  2. Vendor benchmarks ≠ market studies. Filevine’s 14–0 silent-treatment comparison and Paxton’s CaseHOLD scores are valuable signals, but they are not substitutes for multi-vendor, third-party audits (Midpage’s Vanderbilt plan is notable for naming that gap).
  3. Humans still in the loop. Ambrogi quotes Filevine CEO Ryan Anderson: a human still has to look; the tool points you in the right direction (LawNext). vLex’s Cert goes further and keeps editors in the critical path for low-confidence AI tags.

Sources (selected)

This article is a secondary digest of publicly reported material. It is not legal advice. Always verify citator results against primary authority before filing.