How we verify what we deliver

This page describes the controls that stand between an AI-generated draft and a document that reaches a client. It does not claim that a language model cannot make mistakes. It explains what happens so that those mistakes do not leave the firm, and what we can and cannot show you about it.

The claim we make, and the one we don’t

We do not promise that a model will never hallucinate. Our claim is narrower and verifiable: no regulatory citation reaches a client without having been checked against its primary source, and no deliverable leaves without a partner’s signature. The claim concerns delivered work, not the drafting process, which produces errors that the controls below exist to catch.

The context matters. Stanford RegLab’s study of commercial legal research tools (Magesh, Surani, Dahl, Suzgun, Manning and Ho, Journal of Empirical Legal Studies, 2025) measured hallucination rates between 17% and 33% in leading products, and found vendors’ “hallucination-free” claims overstated. We cite it because it explains why verification has to be a separate step with its own agent and its own record, not because it measures our system: the study and our internal counts do not share a test population.

The corpus

Our system reasons against a verified regulatory corpus, not against a model’s memory. The corpus contains eight years of CNBV filings obtained through Mexico’s transparency laws, the Diario Oficial de la Federación (DOF), and the applicable laws and general provisions, indexed article by article so that a citation resolves to a specific provision rather than to a document title. Marisol®, our regulatory monitoring agent, reads the DOF every publication day; when a reform reaches a client’s documentation, the affected documents are updated, reviewed and signed within the same week. Jurídicon® is the platform that runs this; the models are Claude, on Amazon Web Services Bedrock.

Human review gates

Every matter runs through human review gates designed for its process type. The number and content of the gates depend on the entity and the deliverable, and they are never fewer than four; every matter ends with a partner’s signature. In an IFPE authorization file, for example, a lawyer reviews the scope and inputs before analysis begins, the requirements matrix before drafting, the consistency of the full set of documents before finalization, and the finished file before delivery. Each gate decision, with its comments and corrections, is written to an append-only audit trail.

Adversarial verification and the halt

An adversarial QA agent, with no access to the drafting context, verifies every regulatory citation against our primary-source corpus. The pipeline halts automatically if more than 5% of citations are misattributed or more than 3% cannot be found, and it does not resume until a lawyer has reviewed and corrected the file; if it halts twice on the same matter, the matter leaves the pipeline and is done by hand. Nothing reaches a client before the human review gates and a partner’s signature, each recorded in an append-only audit trail. Across two full cycles on real regulatory files, the adversarial check surfaced 100 findings and 20 blocking issues; all were resolved before delivery, and 33 documented human decisions shaped the final documents. Result on delivered work to date: zero citations delivered without primary-source verification. Controls mapped to NIST AI RMF 1.0 (15 subcategories); ISO/IEC 42001 certification in preparation.

Two cycles are two cycles. The counts above are what happened on those files; they are not an error-rate estimate for every future matter, and findings, blocking issues and human decisions are three different things.

What happens when it halts

A halt is not an exception path; it is the design. When the thresholds trip, the file stops, a lawyer reviews every flagged citation, corrects the draft and records the correction, and only then does the pipeline resume. A second halt on the same matter removes it from the pipeline and the matter is completed manually. The thresholds decide when the machine stops; they do not lower the standard for what leaves the firm, which is verification of every citation, not of most of them.

If an error is found after delivery, we correct it at our cost and notify the client. How it is handled beyond that depends on the matter and the engagement terms.

Human contribution, documented

Mexican jurisprudence on transparency in the use of AI in legal contexts (Jurisprudencia J/2 K) requires that a human decision be identifiable as such. Our audit trail records, for every deliverable, which gate decisions were taken, by whom, and what changed as a result; in the two cycles above that came to 33 documented decisions. On request, a client receives an extract of the audit trail of their own matter: gate decisions, partner sign-off and the citation-verification record, with third-party confidential material redacted. Personal data in the trail is also accessible under ARCO rights (Ley Federal de Protección de Datos Personales en Posesión de los Particulares).

Where your documents live, and where they don’t

The AI models we use run under commercial zero-data-retention terms: no prompt, output or document is retained by the model provider, and nothing is used to train any model. Your matter file lives only in Legal Paradox®’s own cloud environment, isolated per matter, encrypted at rest under a dedicated key per client that Legal Paradox® controls, with an append-only audit trail of every access and decision.

Legal Paradox® is registered in the Claude Partner Network. SOC 2 Type I and ISO/IEC 27001 certification programmes are underway, and ISO/IEC 42001 is in preparation; none has been obtained yet, and we will say so on this page when one is.

What we publish and what we don’t

We publish the halt thresholds, the counts per production cycle and the number of documented human decisions, and we will update them as cycles complete. We do not publish prompts, the agent architecture or the corpus itself, for reasons of professional secrecy and intellectual property. Publication is not the same as access: what a client can see of their own matter is described above, and it is more than what appears on this page.

Frequently asked questions

What happens if the model gets something wrong?Inside the pipeline, the adversarial check catches misattributed or non-existent citations and halts the file above the thresholds; a lawyer corrects it before anything resumes. After delivery, we correct the error at our cost and notify the client.

Does a human read everything?Every deliverable passes through at least four human review gates, each with a defined scope, and a partner signs it. That is different from claiming that every reviewer reads every intermediate artifact; the gates are designed so that the parts that matter are read by the person accountable for them.

Do my documents train the model?No. The models run under zero-data-retention terms with the provider; nothing you send is retained or used for training. Your file lives in our environment, under a key we control.

Who answers to the CNBV for an error?A partner of Legal Paradox® signs every deliverable and is accountable for it under the engagement, in the same way as in any law firm. The signature does not transfer the client’s own regulatory obligations; it does mean a named lawyer stands behind the document.

Can I see my matter’s audit trail?Yes, on request: an extract with the gate decisions, the partner sign-off and the citation-verification record, with third-party confidential material redacted. Personal data is also accessible under ARCO rights.