Annlex vs Descrybe: how they compare in 2026

The metered unit is the model. Every paid plan, the credit pack and the free allowance are counted in AI queries, which the pricing page defines as each Annlex Chat question, each AI case analysis and each synthesis request, while Boolean, semantic and hybrid search are given away free and unlimited. What a buyer pays for is Annlex Chat's retrieval and synthesis over 10.5 million opinions and the Opinion Lens tools (Key Passages, Summary, Case Brief, Ask AI), and even the free semantic search runs on ModernBERT embeddings fine tuned for legal text by the Free Law Project. Remove the models and what remains is a free search box over a public corpus the vendor does not own. Checked the Annlex Chat page, pricing, the FAQ, the terms and the Opinion Lens page on 15 September 2026.

The corpus itself is a model output, which is as central as this axis goes. The product was built by AI summarizing judicial opinions, a corpus the vendor's July 2026 coverage page puts at 10,969,112 opinions across 492 courts and tribunals, and that summarization is not a feature layered over a database, it is how the database was built. On top of it sit DescrybeLM as a purpose built legal reasoning engine, Cytator as an AI driven citator producing issue level treatment analysis, and an engine layer exposing retrieval and verification tools to external assistants. Remove the models and there is no corpus, no citator and no product, only public domain opinions the company never had rights to sell. Third on this index at A, after Reveal and Jhana.ai. Corpus figure updated 11 September 2026 from the 3.6 million of an earlier milestone; grade unchanged.

Grounding is real, documented and unusually candid about what it does not fix. The FAQ (version 1.2, March 2026) describes the method: each question is classified, matched to one of ten research patterns and routed through a pipeline that retrieves opinions before the model writes, every claim must cite a case from the database, and every citation opens the full opinion, which the vendor states it holds on its own servers from the CourtListener bulk export. Answers carry a Verify link into Opinion Lens, where a characterization can be checked against the opinion passage by passage, and Opinion Lens answers only from the open opinion and says so when the opinion does not answer. The vendor names the failure that grounding leaves behind rather than claiming it away: a real case cited correctly and described wrong, a dissent quoted as the majority, a rule applied in the wrong jurisdiction. The terms list misread precedent, missed or inappropriate authority, outdated law and missed controlling authority as expected failures. Held at B on the band: no measured accuracy of any kind is published, with no test set, no citation resolution rate and no characterization error rate, and returned authority is not checked for subsequent history. One overstatement noted: the FAQ's opening answer says the design eliminates the conditions that cause hallucination, while the same page and the terms describe a constraint that shifts the failure mode, which is the accurate claim. Read 15 September 2026.

The best structured grounding disclosure in this category and still short of a published measurement. DescrybeLM answers carry a table of authorities with citations linked back to the primary law, the reasoning is stated to be visible at every step, and the product includes quote verification and citation resolution as first class tools rather than as claims. The vendor also discloses against its own interest: the terms of service state plainly that Descrybe does not review or validate outputs to confirm they are accurate, truthful, relevant, reliable or not misleading, and earlier product material warned users not to assume summary accuracy without checking against the original opinion. A vendor telling a reader that its output is unverified is doing the opposite of overclaiming, and it is rare enough on this index to name. Held at B on two gaps. No measured accuracy figure of any kind is published by the vendor: no summarization fidelity rate, no citation resolution accuracy, no hallucination rate, no test set. And a benchmark claim is circulating in third party material that DescrybeLM outperforms named frontier models on bar exam benchmarks, which was not located in vendor material with any methodology, evaluator or result table attached, and is therefore not credited here. Checked the home page, the FAQ, the user guide, the terms of service and the product pages on 29 Aug 2026.

The product is built around a reader checking the machine, and the review surfaces are real rather than asserted. Opinion Lens keeps the opinion on screen at all times, the generative tools sit beside the text and point to the paragraph an answer rests on, Key Passages anchors the question presented, rule, holding and limitation to paragraph numbers, and Ask AI declines when the opinion does not contain the answer. Annlex Chat links each citation to its opinion and offers a Verify route into Opinion Lens. The FAQ states that attorneys remain responsible for independently verifying all authority before reliance, that the tool suits preliminary research rather than replacing verification, and that it should not be anyone's sole source. Nothing acts on a matter; output is research text and downloadable summaries and briefs. Held at B because the control structure stops short: no threshold at which Annlex Chat declines is described, the FAQ's promise to show confidence levels where applicable is not specified, and nothing addresses what happens after an answer is found to be wrong. Read 15 September 2026.

Oversight is described as product behavior rather than as policy, which is the useful form. DescrybeLM is documented to clarify jurisdiction, posture and key facts with the user before proceeding, which puts a human checkpoint at the start of the task rather than only at the end, and it is documented to tell the user when it needs more information and to flag uncertainty rather than filling gaps with guesses. Reasoning is visible at each step and work product is exported deliberately to Word rather than acted on. In the engine configuration the user can see which Descrybe tools an external assistant called and read the data each returned, which is oversight of a tool chain and not merely of an answer. Held at B because none of it is quantified or bounded: no statement of when the system proceeds without clarification, no confidence threshold, and no description of what the uncertainty flag actually triggers.

No production evidence located. No named customer, firm, law school, library or user count, no case study, testimonial or usage figure, and no outcome measure appears on the Annlex Chat page, pricing, the FAQ, the terms or the Opinion Lens page, and the static sitemap lists no customer, news or blog page. The product is live with published self serve pricing, and the founder's public profile records a pitch at the Dig South Tech Summit in Charleston, which is launch activity rather than deployment evidence under the band. Checked those five surfaces, the static sitemap and a general web search on 15 September 2026.

Adoption claims are institutional and specific, and none of them was located in vendor material. Third party research material states that Descrybe was added to the NSLT curriculum replacing Casetext, and selected as an approved NELLCO e-resource reaching nearly 150 law libraries. If published by the vendor with detail those would be strong evidence, because a law library consortium and a curriculum adoption are procurement decisions by institutions that evaluate research tools for a living. As located they are third party assertions. Independently verifiable and dated: sustained coverage by a named legal technology journalist across four separate product stages from 2024 to 2026, an ABA Woman of Legal Tech recognition for cofounder Kara Peterson in 2024, and an Anthem Award for Responsible Technology. Held at C because no vendor published customer, deployment, usage figure or outcome measure was located, and because the corpus milestones the vendor does publish describe its own build progress rather than any customer result. Source basis Third Party Estimated on that footing.

Substantive commitments on training and on the model provider, written into the terms rather than a separate policy page, with segregation silent and the storage account internally inconsistent. Section 10 of the terms (last modified date shown as 27 February 2026) states that saved conversations are not used to train any AI model, the vendor's or anyone else's; that Annlex Chat history saves by default to the vendor's own servers for registered accounts, is kept up to one year and then deleted, can be switched off with a Save history toggle and can be deleted immediately and irreversibly; that nothing is saved for users without an account; and that queries go to Anthropic's API, which does not train on API data and may retain query content up to 30 days, longer where a safety system is triggered. The vendor states plainly that Zero Data Retention with Anthropic is not active and sits on its roadmap, which answers the provider retention question most records leave open. Privilege is engaged in the FAQ, but as a positioning argument tying attorney client privilege to that future arrangement rather than as a stated handling position. Held at B on three gaps. Nothing addresses segregation between users or any team or matter structure. The third party section of the same terms says Annlex stores the text of Annlex Chat queries and conversations in no database, and the pricing page FAQ says queries are not stored, both contradicted by the history paragraph, so a reader has to work out which statement is current. And no data processing addendum or separate customer agreement exists for a firm to sign. Read 15 September 2026.

Substantive published commitments on the two things a firm asks first, and still nothing on privilege itself. The security page (updated 30 July 2026, read 11 September 2026) states that customer prompts, files and conversational history are not used to train Descrybe owned models; that the primary model provider is OpenAI's API, opted out of training, with abuse monitoring logs retained up to 30 days; that uploaded briefs are not kept once the text is extracted for analysis; that research history can be deleted by the user at any time; and that the product is built on individual user accounts with no shared or matter based workspaces, so segregation is per user and walls at matter level are stated not to exist. The binding terms of service, framed as the legal equivalent of a signed contract, still say nothing about any of this, so the commitments a buyer can read are policy rather than contract. Held at B on two gaps. No treatment of legal professional privilege or work product was located anywhere, including in the privacy policy (updated 29 September 2025) and the security page, for a product that now ingests a user's own briefs and pleadings. And the position on the model provider covers training and log retention without stating where OpenAI processes the request. Regraded from C on 11 September 2026: the C rested on the privacy policy not having been read and on the security page not having been located, and both are now on the record.

A real position on the advice line, carried onto the product surface itself, short of full treatment. The terms open with a section headed as a statement that the service is not your lawyer, calling it an experimental research tool that does not provide legal advice and directing users to a licensed attorney; the FAQ adds that payment does not create an attorney client relationship and lists prohibited uses including reliance as legal advice, sole source research and FCRA purposes; and the Annlex Chat page, which names pro se litigants first among its audiences, carries a line stating it is a research tool only, not legal advice, and to consult a licensed attorney. That consumer facing disclosure at the point of use is what most records selling to the public lack. The competence dimension is present in outline, with attorneys told they remain responsible for verifying all authority before reliance. Held at B because jurisdiction limits are not named beyond the corpus being American case law, bar guidance is deferred to each practitioner's jurisdiction without engagement, and nothing speaks to supervising the paralegals the FAQ names as users. Read 15 September 2026.

Not addressed, and this is the record on the index where it most obviously should be. The vendor states directly that the product brings legal reasoning to anyone who needs it, lawyers and non lawyers alike, and the founding mission is access to justice for members of the public without formal legal training. A guided reasoning engine that takes a member of the public's question and returns a citation backed analysis is the unauthorised practice question in its clearest form, and no position on it was located: no statement that output is not legal advice, no guidance on when to consult a lawyer, and no engagement with any bar rule. The closest thing is the terms disclaiming that outputs are reviewed or validated, which allocates risk without addressing the professional question. Checked the home page, the FAQ, the user guide, the terms of service and the product pages on 29 Aug 2026.

Nothing published. No AI policy, no evaluation or release testing, no bias or accuracy monitoring, no model card, no named owner and no external standard. The nearest material is a list of limitations in the terms and FAQ and a design statement on the Opinion Lens page that credibility rests on falsifiability, which describes a product choice rather than a governance mechanism. The gap has a specific shape here: the ten research patterns that route every question decide which opinions the model ever sees, and nothing describes how those patterns were built, tested or changed. Checked the Annlex Chat page, pricing, the FAQ, the terms and the Opinion Lens page on 15 September 2026.

Nothing published. No AI policy, no evaluation methodology, no bias or fairness assessment, no accuracy monitoring, no drift statement, no model card, no named governance body and no external standard. The gap has a specific shape here that is worth recording: the entire corpus is machine generated summaries of judicial opinions, so any systematic bias in summarization propagates into every search result and every citator treatment downstream, and it would be invisible to a user reading the summary rather than the opinion. The vendor rebuilt all its summaries once already, on newer models, which is an implicit acknowledgment that summary quality is model dependent and improvable, and no evaluation of either generation was published. Checked the home page, the FAQ, the user guide, the terms of service and the published product announcements on 29 Aug 2026. Added 11 September 2026: the privacy policy and the security page (updated 30 July 2026) were read; the security page states that models may be updated over time for quality, safety, performance or reliability without user notice unless the change is material, which is a supply chain statement credited on that axis, not a governance mechanism, a testing regime or an owner. Grade unchanged.

Retention and deletion are specific enough to hold the vendor to; the rest of the set is thin. Published: saved conversations kept up to one year then automatically deleted, a user toggle that stops saving, immediate and irreversible deletion of saved conversations, no history saved for users without an account, Google Analytics data deleted after 26 months with IP anonymization, server access logs said to hold request metadata only, and named third parties with their data terms, being Anthropic (no training on API data, retention up to 30 days, longer on a safety trigger), Stripe for payments and Google Analytics. Access, correction and deletion requests route to a published privacy address. Held at B on four gaps. The server infrastructure provider is referred to but not named, so the subprocessor list is incomplete. No access control practice is described. Security is stated only as reasonable measures including HTTPS, with nothing on encryption at rest. And no incident or breach practice is published. The third party section's statement that chat text is kept in no database also conflicts with the default on history described earlier in the same document. Read 15 September 2026.

Most of the ground is covered, and specifically enough to hold the vendor to. The security page (updated 30 July 2026, read 11 September 2026) states what is retained (prompts, research history, outputs, metadata, operational logs), what is not (uploaded files are discarded once the text is extracted), who can delete what (users delete history entries at any time, accounts on request per the privacy policy), that no zero retention mode is offered, that production data access is restricted to authorized technical operations personnel on a need to know basis for named purposes, that data is encrypted in transit and at rest on DigitalOcean managed services in the New York region, and it names the subprocessors with links to each one's attestation: OpenAI's API, DigitalOcean, Clerk and Stripe, plus Google Analytics in the privacy policy. It also states the model provider training position: OpenAI API, opted out. Held at B on two gaps. No incident or breach practice is published: nothing on detection, notification timing or who is told. And no retention period is stated for the operational logs and metadata that sit outside the user's deletion control. Regraded from D on 11 September 2026: the D rested on the privacy policy not having been read and on the security page not having been located, and the privacy policy on its own (updated 29 September 2025, five clauses on account data and named third parties) would have moved this to C at most; the security page is what moves it to B.

Liability is addressed through a standard clause that disclaims all of it. The terms provide the service as is with every warranty disclaimed, and exclude liability for any direct, indirect, incidental, special or consequential damages arising from use, with no cap figure because nothing is retained to cap. Disputes go to binding AAA arbitration under the Consumer Arbitration Rules seated in Charleston, South Carolina, with class action and jury waivers, under South Carolina law. The recourse that does exist concerns availability rather than output: prorated credit or refund for a substantial outage, full refunds for billing errors, and 30 days notice with prorated refunds if the service is discontinued. The terms list the ways AI output can be wrong and place reliance at the user's risk. No indemnity, accuracy undertaking or correction route for wrong output was located. Read 15 September 2026.

An actual published liability position exists, which most of this roster lacks, and it runs almost entirely one way. The terms of service state that the user indemnifies and holds Descrybe harmless from liability, loss, claim and expense including reasonable attorneys' fees related to use of the service or its outputs, and that any liability found on Descrybe's part is limited to the amount the user actually paid in the twelve months before the action giving rise to it. Read against published pricing that caps vendor exposure at $300 to $600 per user. The same document states that Descrybe does not review or validate outputs for accuracy. Graded C rather than D because a specific, quantified and readable cap is genuine disclosure that a buyer can price, and it is more than the vendors on this index who publish nothing. No warranty, accuracy undertaking, service level or correction mechanism was located.

No integration into practice systems located. No document management, practice management, Word or Outlook add in, API or connector is published. Summaries and case briefs download as formatted documents, which is export rather than integration. The absence reads as a position rather than an oversight: the FAQ argues that tools reaching Claude through MCP connectors cannot qualify for Anthropic's Zero Data Retention while a direct API architecture can, so the vendor has chosen not to be an assistant plugin. Checked the Annlex Chat page, pricing, the FAQ, the terms and the Opinion Lens page on 15 September 2026.

Integration is the product rather than an afterthought, and it is documented at tool level. Descrybe Legal Engine exposes citation lookup, quote verification, treatment review and source retrieval to external assistants over an MCP interface, with named supported hosts including Claude, ChatGPT and Perplexity, and an Open Connector for building further tools. A user guide describes the call pattern: the assistant invokes one or more Descrybe tools, Descrybe performs the search, lookup, verification or retrieval, and returns focused results, with the user able to see which tools were called and read what each returned. Work product exports to Word. Held at B because the integrations are to AI assistants rather than to legal practice systems: no document management system, no practice management platform, and nothing connecting to a firm's existing matter estate. Notable as the first record in this pull whose integration story is an interoperability protocol rather than a connector list.

Cloud delivery is plain and little else is stated. The product is a hosted web application, the corpus is stated to sit on the vendor's own servers, and research queries are processed through Anthropic's Claude API, but no hosting provider, region, tenancy model or processing location is named, and the terms refer only to an unnamed server infrastructure provider. There are no deployment options to compare. Held at C because what processes a query and where the corpus is held are described, while the region and tenancy questions go unanswered. Read 15 September 2026.

Tenancy and storage region are stated clearly; processing location is half stated. The security page (updated 30 July 2026, read 11 September 2026) states that customer data is stored on DigitalOcean servers in the New York City region, that the product runs on provider managed multi tenant infrastructure rather than dedicated hardware, that it is structured around individual user accounts, and that no single tenant, private or shared workspace option exists. That is a plain answer to where the data sits and how the software runs, which is more than most records in this category publish. Held at B because processing is only partly located: research requests are processed by OpenAI's API, and nothing states which region that processing happens in, which matters more here than usual because the engine architecture also routes queries through a third party assistant host such as Claude, ChatGPT or Perplexity, and nothing describes where that leg runs. No residency options exist to choose between, so the A test of what changes between tiers does not arise. Regraded from D on 11 September 2026: the D rested on the security page not having been located.

No independent security attestation located. No SOC 2, ISO 27001 or other certification is claimed, no trust center exists, and security is described in the terms only as reasonable measures including HTTPS encryption, with no guarantee of absolute security. Stripe's PCI DSS standards are cited for payment handling, which covers Stripe and not Annlex. Checked the Annlex Chat page, pricing, the FAQ, the terms, the Opinion Lens page and the site footer on 15 September 2026.

No attestation of its own, stated by the vendor rather than inferred by the index, on the most candid security page a D carries on this index. The security page (updated 30 July 2026, read 11 September 2026) states that Descrybe does not currently maintain its own SOC 2 or ISO 27001 certification and has begun work towards SOC 2; that it relies on providers with their own programs, linking DigitalOcean (SOC 2 Type II, SOC 3), OpenAI (SOC 2 Type 2, ISO 27001, 27017, 27018, 27701), Stripe (PCI Level 1, SOC 2 Type 2) and Clerk (SOC 2 Type 2) to their trust portals; that data is encrypted in transit over TLS and at rest on DigitalOcean managed services; and that production access is restricted on a need to know basis. Held at D on the letter of the band: no independent attestation of Descrybe itself exists, and a provider's report covers the provider's infrastructure, not the application layer where Descrybe's code, access controls and data handling live, which is the layer a firm's security questionnaire asks about. The grade moves to B the day a SOC 2 report is stated with scope and date, and to A when it is reachable. Context recorded on 29 August still applies: a self funded two founder company pricing at $25 to $50 a month, and a SOC 2 examination is a material cost against that model. The 29 August finding that no security page existed was wrong; the page is dated 30 July 2026 and sits in the site footer, and this note is corrected on 11 September 2026 with the grade unchanged.

The models are named to a level most of this index does not reach, and change notification is the gap. The FAQ states that Annlex presently runs Anthropic's Claude Sonnet as its leading model, calling the Claude Messages API directly, that semantic search uses ModernBERT embeddings fine tuned for legal text by the Free Law Project, and describes the vendor's own contribution as an orchestration layer that directs leading models. The provider's data terms are stated alongside: no training on API data, retention up to 30 days, Zero Data Retention not yet active. Held at B because no model version is pinned, the word presently signals change without any commitment to tell customers when it happens, and where the provider processes requests is not stated. Read 15 September 2026.

The provider is named outright and the change notification commitment is the gap, which is exactly the B band. The security page (updated 30 July 2026, read 11 September 2026) states that Descrybe uses OpenAI API models for AI generated summaries and research related outputs, that OpenAI's API is its current primary LLM provider, that the exact model may vary by feature and may be updated over time for quality, safety, performance or reliability, and that user by user notices are not provided for every model update unless the change materially affects terms, privacy posture or data processing commitments. That settles the conflict this record carried on 29 August: the terms of service disclose third party models without naming them, a third party directory claimed DescrybeLM was the company's own model, and the vendor's own security page now says OpenAI. The vendor's position on the provider's data handling is also stated, opted out of training and abuse logs retained up to 30 days, with a link to OpenAI's data controls. Held at B because no specific model or version is named, the vendor states rather than commits that it will not notify on model changes, and where the provider processes requests is not stated. Regraded from C on 11 September 2026; the C rested on the security page not having been located.

The full price list is public and the unit is defined. Starter is $25 a month for 200 AI queries and Power is $75 a month for 600, both with a seven day trial that requires a card; a one time pack of 100 queries costs $10 and never expires; three AI queries are free without an account, and Opinion Lens allows three questions per opinion before signup. Boolean, semantic and hybrid search are free and unlimited. The pricing page defines what counts as a query, and the terms publish trial conversion, the absence of rollover, 30 days notice of price increases and the refund rules for packs and subscriptions. Signup is self serve, so nothing is added for implementation, and only volume beyond the Power plan is by contact. One loose end: the refund terms refer to annual subscriptions, and no annual plan is published. Read 15 September 2026.

The first A on this axis in the pull, and it is earned on the full set. Exact prices are published for every tier: $25 per month for Descrybe Legal Engine and $50 per month for the full Descrybe Platform with the Engine included. The unit of charge is a monthly subscription per user, stated plainly. Feature gating is explicitly ruled out, with the vendor stating that every feature is included at either tier, which removes the usual variable where a headline price buys an unusable subset. The vendor goes further and explains the pricing basis, stating that it is self funded, without outside investor timelines or enterprise sales overhead, and prices for the work rather than the market. A buyer can determine total cost of ownership from the public site without contacting anyone, which is the top of this axis. Every other record in this pull to date requires a demo request or a login.

Audience and use boundaries are stated precisely; corpus boundaries are not. The vendor names pro se litigants, law students, solo practitioners and small firms priced out of Westlaw and Lexis, paralegals and journalists, lists appropriate uses (case discovery, preliminary research, academic work, journalism) and prohibited ones (legal advice, sole source research, FCRA purposes), and says plainly the tool suits preliminary research rather than replacing verification. The corpus is described as 10.5 million full text opinions from 3,355 federal and state courts, 1658 to the present, with 71 million citation links and a taxonomy of 20 legal areas and more than 1,125 concepts, updated quarterly from CourtListener's bulk export. Held at B because nothing on the vendor's own estate enumerates courts or date ranges per court, states what curation from CourtListener left out, or says that statutes, regulations and secondary sources sit outside the corpus. The FAQ also describes opinions as retrieved live from CourtListener, while the Annlex Chat page and the quarterly cadence describe a stored copy. Read 15 September 2026.

Corpus coverage is now enumerated at the level a researcher checks, and the boundaries are stated, which is the top of this axis for a primary law product. A coverage page in the site footer (descrybe.com/coverage, data as of 18 July 2026, read 11 September 2026) lists 492 courts and tribunals and 10,969,112 opinions, grouped the way researchers look for them: federal (137 courts, 3,221,272 opinions), state (201 courts, 7,324,869), bankruptcy (97 courts, 80,941), specialized and agency tribunals (47 tribunals, 340,850) and tribal courts, with statutes, regulations, constitutions, session laws and court rules under a separate laws and rules section. Every court carries its own opinion count and a first to most recent date range, so the test this record set on 29 August, whether last quarter's appellate decision in a given state is present, is answered by reading a row: most active supreme and appellate courts run to the first days of July 2026, and where coverage stops the page says so, with Connecticut Superior Court ending in December 2014 and the Alabama Court of Appeals in 1999. Limits stated rather than glossed are what separate A from B here. Regraded from B on this evidence: the B rested on the absence of a per court breakdown, a date range and any recency signal, and all three are now published. Two things a careful reader will still want: a stated refresh cadence rather than a single as of date, and a look at a handful of rows whose most recent date sits in the future (the New Hampshire district court, the Minnesota and Arizona appellate courts), which read as source metadata rather than coverage claims. Segment coverage is unchanged and broad by design, the vendor stating that it serves lawyers and non lawyers alike. The vendor pointed the index to the coverage page after publication; the record was regraded on what the page publishes, not on the request.