Common questions about independent legal AI verification.

An LLM cannot reliably detect its own errors. Cross-vendor independence means the Adversary agent has no systemic reason to overlook errors the Advocate might be inclined to defend. This is structural, not configurational, and is protected by the VariLegal patent filings.

The deterministic auditor is a rules-based system with zero LLM involvement. It produces the final verdict. If a citation does not exist in CourtListener, the auditor overrides any LLM opinion. Hard failures are never subject to model judgment. This is the system of record.

VariLegal is platform-agnostic. Draft with Harvey, Lexis AI, Westlaw AI, CoCounsel, Copilot, Claude, ChatGPT, or another tool. Upload the output and VariLegal runs the verification independently of its source. VariLegal has no commercial relationship with any of these platforms. That independence is structural.

False leniency means a VariLegal certificate was issued for an output that contained a material error. In benchmark testing against all 225 overruled cases in the Stanford RegLab preregistered dataset (the complete set, not a sample), VariLegal achieved 0% false leniency. No fabricated or materially erroneous output received a passing certificate. This metric is the sole pass/fail gate for any model upgrade to the verification pipeline.

VariLegal detected 217 of 225 overruled cases in the Stanford RegLab dataset. The 8 missed cases involve implicit overrulings where no court has explicitly stated the case is overruled. These require citator-level treatment data (KeyCite, Shepard's) that VariLegal does not yet have access to. When citator integration is live, this number goes up. We report the actual number, not the aspirational one.

Yes. The benchmark runs against the Stanford RegLab preregistered dataset, which is publicly available on HuggingFace. Results are locked at a frozen Cloud Run revision. The methodology is disclosed in full on every compliance certificate. We do not publish numbers that cannot be independently verified.

Documents uploaded to Rubicon are used exclusively for verification of the submitted AI output. They are not retained after the session, shared with third parties, or used to train any model. Your document is the ground truth for that verification and nothing else.

Verity verifies citations against CourtListener's federal case database, operated by the Free Law Project. VariLegal has a commercial agreement with the Free Law Project. All citation data is open-source and publicly available. No proprietary citator data (KeyCite, Shepard's) is used in the current production system.

Verity verifies a citation. Rubicon verifies AI-assisted work product against private documents. Retrospect verifies your brief before filing and supports enterprise corpus review. Retrospect runs a backward audit across previously filed briefs to surface citations that were fabricated, overruled, or mischaracterized. Same adversarial architecture, three different applications.

Every verification produces a PDF certificate containing the full deliberation trail, methodology disclosure, and a SHA-256 cryptographic hash computed over the certificate content. Any modification to the certificate invalidates the hash. The certificate is designed to be attached to a court filing or matter file as documented due diligence.

Yes. Enterprise and white-label deployments are available for law firms and legal technology platforms. API access for programmatic certificate delivery. Contact info@getvari.ai or request access through the site.

Verity currently verifies against CourtListener's federal case database. State court coverage is expanding as CourtListener's corpus grows. Rubicon is jurisdiction-agnostic because it verifies against your uploaded document, not an external database. Retrospect inherits the coverage of whichever verification mode it applies.