Citation integrity.
Leverage statistical analysis with multiple LLMS to verify and score every citation in your federal filings, before you submit.
Then run the same analysis on opposing counsel's.
Mata v. Avianca, 2023: six fabricated citations, sanctions on the docket. In re Prince Global Holdings, 2026: around twenty-eight erroneous citations on an emergency motion by one of the country's most prestigious firms, same district, three years later. The lesson nobody needed to learn twice has been learned several times.
The tooling has to be different now.
Run Sentinel Citation on the brief you're drafting. Then run it on opposing counsel's. The report tells you what's fabricated and what's weak. Your first filing is on us.
Citation review cannot be a vibe check.
You can't trust just one.
A single model's intuition is not evidence. A naive citation checker passes a citation through a single language model and asks "does this look real?" That answers the wrong question. The right question is whether the cited authority exists, whether you read it accurately, and whether your treatment of it matches what the court actually said.
Four phases. No shortcuts.
Find every citation
Three language models read the filing independently and return every legal citation they identify. We take the union of their output and combine it with the results from Eyecite, not relying on any single model's word for it
Confirm each one exists
Each unique citation is checked against authoritative sources: CourtListener for case law, Cornell LII for statutes and rules, eCFR for regulations, with a Scholar fallback for opinions only available behind paywalls. Anything that does not resolve is flagged, not silently dropped.
Read the cited authority in context
Once the cited authority's text is fetched, three judging models — Claude Sonnet, Gemini Flash, and GPT, each from a rival vendor — evaluate the filing's use of that authority across four dimensions: Existence, Quote Accuracy, Holding Support, and Citation Strength. The multi-model jury is the core of Sentinel's patent-pending methodology (US 64/056,937).
One verdict per citation, one narrative per filing
A fourth model, the meta-judge, reconciles the three verdicts into a final score and reasoning for each citation, then writes a summary paragraph for the filing as a whole. Something a partner can read in two minutes.
Why hand checks miss this.
Purpose-built legal AI tools still err on 17%+ of queries — because a single model checking citations is still a single model. Sentinel was built on the opposite premise: no single model's word is ever taken for anything.
58%
Hallucination rate, generic LLMs on legal queriesStanford "Large Legal Fictions" (Dahl et al., 2024)
17%+
Error rate in purpose-built legal AI toolsStanford RegLab "Hallucination-Free?" (Magesh et al., 2024)
~28
Erroneous citations on a single emergency motionIn re Prince Global Holdings, S.D.N.Y. Bankr. (2026)
Why this is different.
Existence
Every cited case and statute is cross-referenced against authoritative sources: CourtListener for case law, Cornell LII for statutes and federal rules, eCFR for regulations. If the cited authority does not exist, the citation is flagged before anything else happens.
Quote Accuracy
Anything the filing presents as a quotation is matched against the cited authority's actual text. Misquotations, paraphrases dressed up in quote marks, and outright fabrications surface here.
Holding Support
The cited authority is read in full and compared against the proposition the filing offers it for. We name the gap when the case does not actually say what the brief says it says.
Citation Strength
Above and beyond hallucination defense. Each surviving citation is rated on how well it supports its proposition given the full context of the authority: binding versus persuasive, current good law versus narrowed or criticized, on-point versus distinguishable. On your own filings, this surfaces soft spots before a clerk does. On opposing counsel's, it surfaces them before oral argument.
Even a clean brief earns its fee.
Suppose the report finds zero fabrications. You still walk away with a Citation Strength map of the entire filing: which citations carry the argument, which are binding versus merely persuasive, which would fold under a motion to strike — in your brief, and in theirs. The fabrication check is why you run it the first time. The strategic read is why you run it every time.
Citation review is the litigator's job.
We are not interested in a product that hides behind "AI did it." Citation review is your name on the filing. Sentinel Citation does the parts that don't reward judgment: extracting every citation, verifying each against authoritative sources, reading the cited authority alongside your use of it. The judgment stays with you. Every output is traceable, every verdict has a reason, every model that touched the filing is named in the report.
One question
Would you sign a brief whose citations you didn't verify?
The cases exist or they don't.
They say what you think they say or they don't.
The court will check. So should you.
Your name still goes on the filing.
Questions we hear from litigators.
Do you replace citation review?
No. Sentinel Citation is the first pass. It catches fabrications, misquotes, and treatment errors that humans miss when fatigued. The final read is yours, every time.
Can I run it on opposing counsel's brief?
Yes. The same analysis runs on any filing you can read. On your own briefs, you get the report before you submit. On opposing counsel's, you get a map of where the citations are weakest. Citation Strength scoring is included on every analysis.
What courts and filings do you cover?
Initial release covers US federal court filings: District, Circuit, and Supreme Court briefs, motions, and memoranda. State court coverage is on the roadmap.
Which models do you use?
Sentinel Citation uses a fixed three-plus-one ensemble: three judging models from rival vendors — Claude Sonnet (Anthropic), Gemini Flash (Google), and GPT (OpenAI) — plus a meta-judge model that synthesizes their verdicts. The exact model versions are pinned and recorded on every run, so every verdict is attributable. The composition isn't user-configurable. Cross-vendor model diversity is what gives the verdicts statistical weight; varying it would undermine the consensus that makes the methodology work. The output names every model that contributed to each verdict, so you can see the spread.
What sources do you verify against?
CourtListener for case law, Cornell LII for statutes and federal rules, eCFR for regulations, and Google Scholar as a fallback for opinions available only behind Westlaw or Lexis paywalls. We do not invent authority.
How long does a review take?
A typical 30-page brief runs in a few minutes. Longer filings scale linearly because verification calls and model calls run in parallel.
What happens if a citation cannot be verified?
It lands in a "not judged" bucket with the reason ("CourtListener returned not-found," "no opinion text available," and so on). It is not silently dropped, and it is not silently passed.
Is the filing stored?
Filings are processed in your account and retained according to the controls on your tier. Enterprise customers can configure shorter retention, custom export, or zero-retention.
Where does this fit alongside Sentinel?
Sentinel is our broader evaluation engine for AI-generated work, built on patent-pending drift-detection methodology (US 64/056,937). Sentinel Citation is the same engine, specialized for legal filings. If you already use Sentinel, Citation runs in the same workspace.
See what a verified filing looks like.
Pricing is straightforward and starts at the per-filing level. There is no demo gate. No card required.