B

BriefCatch

The models are the engine of named capabilities layered on a product that functions fully without them, which is the B band, and the vendor draws the line more sharply than any other record in this corpus because the line is a switch. AI is off by default and, with it off, the suite runs entirely on its rule-based engine: more than 12,000 rule sets covering grammar, clarity, style and structure, document scoring and narrative reports, all of which the vendor states are unaffected by the AI toggle.

What the AI adds is three named things: BriefChat, a conversational writing adviser; the AI half of CiteCheck, which refines rule-based citation corrections with pattern recognition; and the second layer of RealityCheck, which assesses whether quoted language appears in an opinion and whether an authority supports the proposition cited. The first layer of RealityCheck is itself deterministic, verifying that a case or federal statute exists by cross-checking reporter volume, court identifiers and case names against authoritative databases.

So even the flagship verification feature is only half model-driven. RealityCheck became standard on every plan in September 2026 and is metered in AI credits, which makes the AI a paid, quantified component rather than a background one, and that is what holds this at B rather than lower. Recorded under ground rules section 3 and applied throughout this record per R126: what this product is for is catching other people's AI errors, and that purpose is not evidence about its own models. Verified 13 September 2026.

Grounding is real and documented with a described retrieval method and linked primary sources, short of a testable accuracy figure for the vendor's own output, which is the B band. The retrieval method is described in more detail than anywhere else in this lane and is genuinely two-stage. Deterministic lookups first confirm that a cited case or federal statute exists, cross-checking reporter volume, court identifiers and case names against authoritative legal databases to catch phantom or misidentified authorities.

AI models then read the opinion and assess whether the quoted language appears in it and whether the authority supports the proposition. Outputs resolve to three states, correct, needing attention, or unverifiable, and every finding points back to the authority it concerns. What is absent is measurement of BriefCatch's own accuracy. No precision or recall figure, no test set, no false-positive or false-negative rate and no third-party evaluation is published for RealityCheck, CiteCheck or BriefChat.

The agreement runs the other way, clause 8.2 of the Master Terms stating that the company does not guarantee that any output is correct, complete or current and that all output should be reviewed before use. Recorded and expressly not credited under R126: the long enumerated list of defects RealityCheck detects, from altered quotations to dissents presented as majority holdings, is a catalogue of other people's failure modes, not a disclosure of this product's own, and the A band asks for the latter.

A reported test on a Fifth Circuit brief in Fletcher v. Experian is trade-press sourced, not first-party, and is not graded. Verified 13 September 2026.

A written commitment that the models work alongside the user with real review surfaces, short of the full control structure, which is the B band. The oversight here is unusually concrete for a drafting tool. Every suggestion arrives as an individual card the user accepts or rejects, so nothing changes a document without a human act. RealityCheck resolves to three states rather than two, separating what it confirms from what needs attention and what it cannot verify, which keeps the unresolved cases visible rather than silently passing them.

And the position is contractual as well as behavioural: clause 8.2 of the Master Terms requires that all output be reviewed prior to use to verify accuracy and suitability, and clause 8.5 goes further, prohibiting the customer from relying on the output as legal advice. What the A band asks for is not published, and one limb of it does not bite. There is no threshold at which the system acts alone because the system never acts alone: it is a suggestion engine inside a document the user controls, so under R15 that limb is inapplicable rather than failed and is recorded as such.

The limbs that do bite and fail are the absence of any published account of what happens after an output is found wrong, and the absence of any statement of when the AI half of CiteCheck overrides or defers to the rules engine. The default-off AI switch and its administrator control are real and substantial, and under ground rules section 3 they are spent once, on AI Governance and Deployment, not again here. Verified 13 September 2026.

Real deployment evidence with substance, short of joining the names to the figures, which is the B band. The naming is strong and is concentrated where it matters for this product: the customers displayed on the pricing page are litigation firms of the first rank, being Skadden, Wilson Sonsini, Nixon Peabody, Kellogg Hansen and Bartlit Beck, which for a brief-writing tool is the relevant proof. Scale is published at more than 25,000 lawyers using the product and 40 of the Am Law 200, and the estate states adoption by state and federal courts, which is a distinctive second constituency and is corroborated by dedicated solution pages for courts and for government.

A Microsoft AppSource rating of 4.9 is published, which is third-party attested even if the vendor reports it. What holds this off A is the familiar split. The named customers carry no figures, and the figure that exists, more than 50 per cent of editing time saved, is unattributed, undated and carries no method. Nothing states how that was measured, against what baseline, or across how many documents. A Case Studies section exists in the site navigation and was not opened; under R25 it corroborates a grade that stands on the customer wall and the published scale, and it is precisely what would move this row if any story joins a named firm to a measured result with a stated basis. Verified 13 September 2026.

Substantive published commitments across most limbs, defeated on two, which is the B band. The commitments are unusually strong and several are contractual rather than promotional. Retention is zero and is written into the agreement: clause 7.2 of the Master Terms records that the company does not store or otherwise retain Customer Content, so there is nothing to retrieve on termination, and the Trust Center adds that document text is processed in RAM only and promptly cleared.

Training is excluded in express terms across several surfaces. Architecture supports both claims: documents never leave the local Word environment unless the user enables AI citation review or BriefChat, and BriefChat has access only to the user's prompts and never to the document or email. Customer Content is defined as the customer's Proprietary Information under clause 5.1 and protected by a mutual confidentiality obligation, and clause 5.2 confirms the customer owns it.

Two limbs fail. Privilege and work product are not addressed by name anywhere located, which under R33 forecloses A on its own. And no model provider is identified, so nothing states what any third party sees. One conflict is recorded because it is the sharpest finding on this record and a buyer must weigh it: the EULA states that the products are not intended for processing confidential or personal data and that the user bears all risks related to data privacy and security, while the product FAQ states the opposite, that the product is safe for use on confidential client documents. Two published instruments in direct conflict on the central question. Verified 13 September 2026.

A real published position on advice versus tooling, short of the supervision and competence dimension, which is the B band and among the strongest instances of it in the corpus. The position is contractual and has two limbs rather than one. Clause 8.5 of the Master Terms provides that the customer shall not rely on the services or their suggestions and outputs as legal advice, and shall not communicate or claim to any third party that any part of the services or their outputs is legal advice.

The second limb is rare: most vendors tell the customer not to rely on the output, and almost none prohibit the customer from representing it to others as advice. Clause 8.6 adds that the services in no way provide legal advice nor create an attorney-client relationship. A third provision reaches the end audience, requiring that recipients of ClientCatch output be advised to consult a qualified attorney for advice tailored to their circumstances, which is a downstream disclosure instruction of a kind this corpus almost never sees.

What the A band requires is absent. Nothing addresses supervision or competence: no statement of who within a customer may use the AI features, what review a supervising lawyer should apply, or what training is expected. No rule of professional conduct or bar authority is named in any jurisdiction, and no jurisdictional limit is stated for a product whose citation checking is United States-specific. Verified 13 September 2026.

A published governance position with a real mechanism behind it, short of testing results and a named owner, which is the B band, and the mechanism is what lifts it there. Under R126(4) the central fact is a shipped control rather than a usage policy: AI features are turned off by default, and administrators can enable or disable AI across an organisation from the user portal. A buyer therefore does not have to trust a commitment about how the AI behaves, because it can decline to run it at all, firm-wide, while continuing to use the rule-based engine unaffected.

That is a product capability, so R124(3), which was scoped to policies governing customer conduct, does not cap it. Default-off is rare enough in this corpus to be a finding in itself. Around it sits a dedicated AI Disclosure page, a stated principle that no more is sent than is necessary to produce a suggestion, an express account of what the models were and were not trained on, and a scope statement that the AI toggle affects only BriefChat and citation suggestions and leaves grammar, clarity and style entirely rule-based.

What the A band requires is missing. No individual, committee or function is named as accountable for model behaviour, no pre-release testing regime is described, and no evaluation result is disclosed. Bias is not addressed in any form, which matters on a product that scores writing against the styles of named judges and advocates. Verified 13 September 2026.

Substantive published policy covering most of the ground, short of the full set on subprocessors and incident practice, which is the B band and both of the gaps it names. What is published is strong. Retention is zero and contractual, clause 7.2 of the Master Terms recording that Customer Content is not stored or retained, with the Trust Center adding that document text is processed in RAM only and promptly cleared, which makes deletion moot rather than unaddressed.

Encryption is stated as AES-256 at rest and TLS 1.3 in transit. Access control is tiered and specific: two-factor authentication and Google authentication on every plan, with single sign-on, SCIM provisioning and role-based access management at enterprise. Resilience is addressed through redundancy, automated backups and a tested disaster recovery plan, and personnel controls through mandatory confidentiality agreements and annual security training.

One capability is named as a first for this corpus: RealityCheck scans documents for concealed text, including white-on-white type, formatting and zero-width characters. That is adversarial input handling, aimed at material planted to mislead a reader or a model, and nothing else in the corpus discloses it. Two limbs fail. No subprocessor list is published and no processor is named beyond Azure, which is infrastructure.

And no incident or breach notification commitment to customers was located, on an estate that otherwise documents itself carefully. Verified 13 September 2026.

A real published position on liability, short of the full picture, and the most complete one located in this pull. The Master Terms are published in full and address every commercial limb a buyer would test. A warranty exists rather than being disclaimed outright: clause 8.1 commits the company to reasonable efforts consistent with prevailing industry standards to maintain the services so as to minimise errors and interruptions, and to perform in a professional and workmanlike manner, with advance written notice of scheduled disruption.

A vendor indemnity exists and is scoped: clause 9 requires the company to indemnify, defend and hold the customer harmless against third-party claims that the services infringe a United States patent or copyright or misappropriate a trade secret, with the company controlling the defence and, if the services are held infringing, obliged to replace them, obtain a licence, or terminate and refund prepaid unused fees. Liability is capped at fees paid in the preceding twelve months, and the cap is disapplied for gross negligence, wilful misconduct and breach of confidentiality, which is a customer-favourable carve-out set this corpus rarely sees.

A service level agreement is incorporated by reference and a 99.9 per cent uptime commitment is published at enterprise tier. What is absent keeps it off A: no warranty of any kind attaches to AI output, clauses 8.2 and 8.3 disclaiming responsibility for output and for anything based on it, no insurance position is published, and the indemnity does not reach a claim arising from a bad verification. Verified 13 September 2026.

Real integrations exist and named connections are documented, short of the depth the A band describes, which is the B band. The integrations are into the systems this work actually happens in rather than into adjacent platforms. The suite runs as a Microsoft Word add-in, as an Outlook add-in at Pro and above, and as a browser-based editor, and it works on Word documents held in Google Drive and OneDrive, so a user does not leave the drafting surface.

Distribution is through Microsoft AppSource with a Microsoft 365 App Certification listing published, which is third-party verified surface area. Identity integration is documented and tier-gated with specifics: two-factor and Google authentication on every plan, single sign-on and SCIM provisioning for automatic account creation, update and deactivation at enterprise. Learning content integrates into a customer's own learning management system using the SCORM standard, which is an unusual and concrete integration for a legal tool.

What the A band asks for and is not established is depth on the one integration a firm would press hardest: document management system connections are listed as an enterprise feature and described generically as connecting to supported systems for document access and workflow, and not one document management system is named, no direction of flow is described, and no configuration prerequisite is stated. The features page was not opened and, with the supported-DMS list, is what would move this row. Verified 13 September 2026.

Region is stated and the tenancy model is not, which under R38 is B because tenancy and region are co-equal limbs and publishing either clears C. Region is answered plainly and restrictively, the Trust Center stating that the service runs across Azure services at United States geo-located systems only. For a buyer with United States requirements that is a direct answer; for one outside the United States it is equally informative in the other direction, and the record should be read that way rather than as a gap.

Delivery is published as a browser-based application, a Word add-in and an Outlook add-in, with distribution through Microsoft AppSource or direct installation, and the vendor states it supports enterprise security configurations and provides a security and privacy white paper for IT review on request. Under R126(4) one further fact is credited here and is a genuine deployment property rather than a policy: **AI is off by default and documents never leave the local Microsoft Word environment unless the user enables AI citation review or BriefChat**, so in its default configuration the product processes text locally and sends nothing.

That is the strongest form of data residency there is, and it is a shipped state rather than a promise. What is not published is tenancy. Nothing states whether accounts sit in a shared or isolated environment, and no private or on-premises option is offered or refused. Verified 13 September 2026.

Certification is real and stated and a trust centre exists, short of accessible evidence, which is the B band. What is published is genuine: a public Trust Center reachable without a sales call or a form, stating SOC 2 certification and setting out encryption at AES-256 and TLS 1.3, a tested disaster recovery plan, cloud firewalls, intrusion detection and continuous logging, and human controls including mandatory confidentiality agreements and annual security training.

The pricing comparison table lists SOC 2 Type II compliance as a feature available on every tier, which is a more specific claim than the Trust Center headline. Three things keep it from A. No report, certificate, audit period or auditor is obtainable or named, so a buyer can read the claim and not the evidence. No scope statement identifies which systems or products sit inside the audit boundary. And the estate contradicts itself on the type: the pricing page and its FAQ state SOC 2 Type II, while the BriefCatch 3 support FAQ, still live, states the company is SOC 2 Type 1 compliant.

Those are materially different attestations, Type 1 testing design at a point in time and Type II testing operating effectiveness over a period, and the discrepancy is named here rather than resolved because nothing published resolves it. Recorded and expressly not credited under ground rules section 3: the ISO 27001, PCI DSS Level 1 and SOC 2 compliance referred to on the Trust Center attaches to the Azure services the product runs on, which is the infrastructure provider's attestation and not this vendor's. Verified 13 September 2026.

The vendor refers to AI models without identifying what sits underneath, which is the C band, and the note sets out precisely which link in the chain is named and which is not because R126 asked for it. What is disclosed is real but sits either side of the models. Upstream, the training basis is described: models are stated to be trained exclusively on proprietary, licensed or publicly available non-customer data, and the pricing FAQ characterises that as real filings, real judicial opinions and decades of legal-writing expertise rather than scraped internet data, with BriefChat trained on the founder's published writing corpus.

Downstream, RealityCheck's deterministic layer is said to check against authoritative legal databases. Between them, nothing. No model is named, no version, no provider, and no distinction is drawn between models the company built and models it licenses. One phrase implies third parties without identifying them: the pricing FAQ states that customer documents are never used to train shared AI models or to improve third-party AI services, which concedes third-party AI services are in the chain while naming none of them.

Recorded and expressly not credited on two grounds. Azure is named as the hosting environment, which is infrastructure rather than a model supplier under ground rules section 3. And trade-press reporting of the RealityCheck launch attributes its citation-verification infrastructure to a supplier named Counsel Stack; that supplier appears nowhere on the vendor's own estate, including the Master Terms and the Trust Center, so under ground rules section 2 it is not first-party and is not credited. Verified 13 September 2026.

A buyer can learn what this costs without entering a sales process, which is the A band. Two of three tiers carry published per-seat rates: Individual at 60 dollars per seat per month billed annually for a single seat, and Pro at 75 dollars per seat per month billed annually with a five-seat minimum, with Enterprise custom above twenty seats. The unit is stated twice over, because consumption is metered as well as seated: each tier includes a stated monthly allotment of AI credits, 650 on Individual and 700 on Pro, and the vendor publishes a worked conversion rate, stating that RealityCheck typically consumes about 70 credits to analyse a brief under 30 pages with fewer than 100 citations.

That lets a buyer estimate throughput before purchase, which is more than a rate card usually gives. The mechanics around it are published too: credits reset monthly and do not roll over, additional packs can be bought at any time and remain valid for a year, annual billing is discounted against monthly, and plans and seats can be changed with credits and settings carrying over. A full feature comparison table sets out what each tier includes across core access, AI limits, learning, security and administration.

Purchase is self-serve with a seven-day trial and no credit card. Two limits are named: enterprise pricing is withheld, and the get-started page carries a different structure, a Solo tier at 599 dollars per year, which does not reconcile with the canonical pricing page. Verified 13 September 2026.

Coverage is described with real substance across buyer types, with the boundaries left open, which is the B band. The segmentation is published as three dedicated solution pages, for law firms, for government and for courts, and the third is what makes this record unusual in the corpus: the bench is addressed as a buyer in its own right, with the product sold to courts to run over filings they receive, and the estate states adoption by state and federal courts.

Beyond that, the tier structure itself describes the range, running from a single seat for an individual lawyer or clerk, through a five-seat minimum for a practice group, to twenty or more for an enterprise, and the vendor states subscriptions are available to solo practitioners, law students, educators, law firms, universities, corporations and government agencies. Named customers evidence the litigation end concretely, with Skadden, Wilson Sonsini, Nixon Peabody, Kellogg Hansen and Bartlit Beck displayed on the pricing page.

What is left open is every limit. No practice area is named as better or worse served, no jurisdiction is stated, and that omission has bite here rather than being formal: citation verification is described against United States reporters and federal statutes and the writing rules derive from United States judicial practice, so the product is jurisdictionally bounded in fact while saying so nowhere. Nothing states which document types beyond briefs, memos and emails are supported. Verified 13 September 2026.