Topic hub

Legal AI: From Corpus and Retrieval to Professional Decision

Legal AI needs a versioned corpus, citation control, and a clear boundary between retrieval, interpretation, and professional decision.

01

Position

How I use this term

Legal AI here is not a synonym for quickly finding similar text or for automated legal advice. Its centre of gravity is material whose scope must be named: jurisdiction, time in force, source type, corpus completeness, and the route to a citation. A system can help organise questions and move a reader toward evidence, but it must not cover a database gap with a confident tone. The articles and cases show how to distinguish retrieval from interpretation, record a negative result, and identify when a responsible process must hand the matter to a person.

A practical test is whether a reader can move from a sentence in the interface to the appropriate document, version, and context, and whether the system honestly marks a case where the material does not support a conclusion. This protects both the user and the person responsible for further interpretation.

A corpus has time and boundaries

A legal corpus is not a neutral collection of files. It needs a stated jurisdiction, period of force, unit of analysis, source version, and update rule. Without those boundaries, a system can find language that is similar but unsuitable for the question. The corpus scope should be as visible in an output as the answer or citation itself.

Retrieval does not issue an opinion

Retrieval can locate material, but it does not decide its meaning, authority, or relation to the facts of a case. An output should separate document, quotation, fact, rule, and interpretation so a user does not confuse textual relevance with a professional conclusion. The distinction is especially important when complete context is unavailable or competing sources support different readings.

Citation as a control

A citation is useful only when it leads to an inspectable passage in an identified version of material. Control covers not only the presence of a link but its source, date, scope, and the fit between the quotation and the claim. A negative case can still look persuasive: a rule from another period, a similar authority, or a quotation missing an exception. That is why citation QA belongs to the process, not to interface decoration.

The professional stopping point

Legal AI can support research and preparation, but it does not remove the responsibility of a person qualified to interpret. A workflow needs a recognisable moment when a lawyer can stop, correct, or reject an output, and that decision must not be treated as a failure. This stop point prevents an automatic leap from textual similarity to legal qualification.

01A

In practice

What distinguishes this topic

In Legal AI, retrieval is not interpretation. A corpus needs an explicit scope, date, and version, and a citation must lead to material a professional can inspect. A lawyer stopping a result is not an interface failure; it is part of an accountable process.

The Legal AI pipeline designs corpus boundaries, citations, and a professional stopping point.

The event ontology helps keep document, fact, rule, and interpretation separate.

02

Projects

Where the method is used

Pilot / Legal AI and Computational Law

NormaLab

Conventional legal search retrieves similar documents. NormaLab asks a different question: how did a specific rule operate across a body of judgments, where is practice stable, where does it diverge, and what supports each conclusion?

The public method exposes the path from question and corpus to citation, report, and evaluation verdict—even when the work is returned for revision.

My role

Data engineering · Agent systems · Reproducibility · Citation engine

Evidence

live demo · article

CASE / NORMALABcase-v1.1Updated

Concept / Legal AI and Computational Law

Legal Copilot Ukraine

A large centralised system spends a long time gathering requirements, averages out distinct needs, and can be obsolete before launch. Legal Copilot Ukraine proposes ready-to-adapt agent workspaces that each professional connects to admissible sources and their own workflow.

The public design specifies access layers, agent-workspace components, and a seven-part package another professional can inspect.

My role

Concept · Systems architecture · Artifact standard · Pilot design

Evidence

live demo · article

CASE / LEGAL-COPILOTconcept-v1.1Updated

Prototype / Legal AI and Computational Law

AdmissusCase

A prototype for a controlled workflow that organises case files, facts, timelines, and sources before materials are prepared for the ECtHR or UN treaty bodies.

The work produced a product hypothesis and public presentation; a private-repository review confirmed the technical document flow, which is not presented as public evidence.

My role

Co-Founder · Technical Lead · Product architecture

Evidence

article

CASE / ADMISSUScase-v1.1Updated
03

Start here

Cornerstone material

Law as a chain of events

An event model separates documents, facts, actions, and legal qualification instead of treating law as a flat collection of text fragments.

research essay
04

Reading

All writing in this area

Pressing Record Is Easy. Explaining It to a Regulator Is Harder

Google Meet, Zoom, Fireflies.ai, Otter.ai, and Gong can place a single conversation on a path through several systems—turning it into a transcript, an AI summary, a CRM record, and part of an organisation’s memory. In 2025, the French data protection authority fined a call-centre company €250,000 for infringements involving data minimisation, retention, and security. An interactive publication about seeing the consequences before recording begins.

Read the publication →
Interactive publication · Legal AI

Law as a chain of events

An event model separates documents, facts, actions, and legal qualification instead of treating law as a flat collection of text fragments.

research essay

Verifiable AI Readiness Audit

Apply the method to your workflow

Check sources, data, evaluation, oversight, and security before selecting a model.

Start the diagnostic