C

Clearbrief

The models are the product, and the choice of architecture is what sets it apart. Citation detection across a document, the patented scoring of how well a sentence is supported by the page it cites, and concept search across uploaded discovery all run on machine learning and natural language processing. The vendor deliberately keeps generative models out of the checking function, saying classic AI means the checker cannot introduce hallucinations of its own, while shipping separate generative features for drafting.

Remove the models and nothing remains but a Word toolbar. Not being generative is not the same as not being AI, and that distinction is the point of this product.

Grounding is real, documented and automated, short of published measurement of the checker itself. The retrieval method is concrete. Citations are detected in the Word document and matched against case law databases: LexisNexis where the user holds a subscription, and Fastcase and vLex where they do not, with Westlaw citations mapped across to those databases. Where a source cannot be displayed automatically, the product still reports that the citation is verifiable and accepts a manual upload.

Output links every citation to the specific page and text the reader can open. The failure mode is named clearly: because checking runs on classic AI rather than generative models, the vendor says there is no risk of the checker adding hallucinations of its own. What is missing is measurement of the product's own performance. A search of the vendor site, blog, sandbox site and trade coverage found no false positive or false negative rate for citation detection, no accuracy figure for the support score, no test set and no independent evaluation.

A real published commitment with concrete review surfaces, and the product exists to serve oversight rather than to reduce it. The vendor states that source linked outputs keep humans in control and the evidence in view, and the Cite Check Report is explicitly built for the partner signing a filing, listing every citation with hyperlinks so each flagged issue can be reviewed in context. The founder frames the problem as partners being sanctioned for citation errors they did not personally make, which is a supervision argument.

Not located: the threshold at which a semantic score is treated as low, whether that threshold is configurable, and what the vendor commits to when its own checker misses something. That last point matters here, because a verification tool that fails silently leaves a lawyer more confident than before they used it.

User categories and testimonials stand in for deployment evidence. The vendor states the product is used by courts, Am Law firms, government agencies, arbitrators and legal teams handling high stakes matters, which is a notable set because courts and arbitrators are adopters rather than customers in the ordinary sense, and an attributed testimonial describes the LexisNexis backed hallucination detection in practice. Searched the vendor site, the blog, the sandbox site and trade coverage and located no named organization paired with figures and a date, no case study, and no assessable method. Categories of adopter are not the same as a named deployment.

Substantive commitments rest on a design choice, short of the full picture. The strongest element is Bring Your Own Storage, an enterprise option under which all uploaded documents stay in the organization's own cloud storage. That answers the confidentiality question by taking the vendor out of custody rather than by promising care. The vendor says the product is built for sensitive privileged legal work with enterprise grade security, privacy and access controls, and that work stays in the environment the team and IT already trust.

A policy of not using client data for model training is reported in third party review material, not on a vendor page that was reached. Two gaps remain. Retention and deletion terms for customers not using Bring Your Own Storage are not published. Attorney client privilege and work product are used to describe the work rather than addressed as a handling question, and no segregation model between users or matters is published.

The vendor engages substantively with the professional duty the product exists to serve. It publishes material on how large firms put responsible AI into practice through citation checking policies, describes the courts' expectation that every sentence carries a verifiable citation in a standard format, and frames the risk as sanctions and reputational damage to the signing partner. That engages with candor to the court and with supervision of others' work, beyond a disclaimer.

The audience is professional throughout: litigators, in house teams, courts, arbitrators and government agencies, with no consumer surface located. Competence is not addressed as such, no statement on the advice line is published, and no jurisdiction limits are stated beyond references to Bluebook and local style rules.

A published position on responsible AI without a governance mechanism a buyer could audit. The vendor publishes on how firms should put responsible AI into practice through citation checking policies. Its decision to keep generative models out of the checking path is a real governance choice with a stated reason. The vendor site, blog, sandbox site and trade coverage were searched. No AI governance framework, AI management certification such as ISO 42001, named owner of model governance, or testing regime before release is published, and nothing addresses uneven output across matter types, parties or populations.

That last point matters for a scoring product: a score that behaves differently across drafting styles or document types would be invisible to a user, and nothing published addresses it.

A general posture with one strong design element and little published detail. Bring Your Own Storage is real and material: an organization can keep uploaded documents in its own cloud. The vendor also says it works with customer security and IT teams through procurement and review. Beyond that the published material is general, describing enterprise grade security, privacy and access controls that IT and security teams can trust.

A search of the vendor site, blog, sandbox site and trade coverage found no retention period, deletion control, encryption detail, named subprocessor list, or incident or breach notification practice. Published filings are also hosted in the vendor's secure cloud for judges and clients to read, a second data location with no published terms attached.

No published indemnity, liability cap, carve out, warranty on output or insurance position is published, and no customer terms of service is published as published on the property. Recorded as a pure absence on the surfaces reached. The shape of the gap is worth naming: this product is bought specifically to prevent sanctionable citation errors reaching a filing, and it publishes nothing about what happens if it misses one.

A verification tool carries a different recourse question from a drafting tool, because its value is the assurance itself.

The vendor names four document repository integrations in its own material: Relativity, iManage, NetDocuments and Clio. It describes them as letting a firm use its existing secure infrastructure while adding the checking layer on top. That is a fitting connector set for a litigation product, covering document management, ediscovery and small firm practice management. Delivery is a native Microsoft Word add-in, where drafting and signing happen, rather than a connector beside Word.

Research integrations are named individually with their commercial terms: LexisNexis where the customer holds a subscription, and Fastcase and vLex where none is required, with Westlaw citations mapped across so they stay checkable. Not located: an integrations index page, or documentation for each integration describing what it moves, in which direction, and what an administrator configures.

The deployment model is stated clearly, with a real customer controlled option, short of residency detail. Two models are published. One is the standard cloud service. The other is an enterprise Bring Your Own Storage option, under which uploaded documents stay in the organization's own cloud storage, which the vendor presents as giving more control over data residency and confidentiality. With customer held storage, the customer keeps custody.

A search of the vendor site, blog, sandbox site and trade coverage found no hosting provider for the standard service, no named regions, no tenancy model, and no statement of where processing happens as distinct from storage. That matters because processing still occurs even when the customer holds the storage.

The vendor states SOC 2 Type 2 certification on its own blog, describing it as showing that it meets rigorous data protection standards and helping a lawyer meet the duty to protect client confidentiality through verified security controls. The same material describes the Bring Your Own Storage option as letting a firm control where client data is stored to meet jurisdictional requirements or firm retention policies, tying that feature to professional obligations.

No coverage period, audit scope, report date, named auditing firm, trust portal or published request route is published. The vendor's stated diligence path is to work with a customer's security and IT teams through procurement, a sales route rather than self service. The certification is asserted on a marketing page without scope, currency or a route to evidence.

The architecture is described clearly, and the components are not named. The vendor says the checking path runs on classic machine learning and natural language processing with a patented scoring system, not generative models. So a buyer knows the checking function does not depend on a third party foundation model. Generative features exist separately for drafting. A search of the vendor site, blog, sandbox site and trade coverage found no named model or provider behind the generative features, no statement of where models run, no subprocessor list, and no commitment to notify customers when the supply chain changes.

Knowing the critical path avoids generative models is useful, but it does not tell a buyer what the rest of the stack is.

No pricing page is published, no rate is published, no unit of charge is stated and no tier structure appears on the surfaces reached. Third party review material states that pricing is custom quoted, which confirms the absence rather than filling it. Every commercial path located terminates in a demo or contact request. One published commercial fact worth recording, though it is not pricing: a recipient of a hyperlinked filing, including a judge or a client, does not need a Clearbrief subscription to open it.

Segment coverage is described with substance and the practice focus is well defined. The stated users are litigators, in house legal teams, courts, arbitrators and government agencies, so the published adopters include the tribunal as well as the advocate. Practice focus is stated narrowly and consistently as litigation documents heavy with facts. Named uses cover briefs, tables of authorities, exhibit management, investigation reports for in house compliance teams, and real time trial support including cross examination outlines.

Not located: firm size segmentation, jurisdictional coverage beyond references to Bluebook and local style rules, and any statement of which document types or practice areas the product is not built for.