A new layer of legal infrastructure
We turn millions of court decisions into a verifiable map of how norms are applied — so courts, the state and lawmakers can see what neither official statistics nor ordinary search reveal.
Research infrastructure, not a chatbot. Every conclusion comes with a source, a quotation and a stored trace.
A tool for those who shape practice, decisions and policy — not just read them.
See how a norm is applied across the entire body of practice: where case law has split, which argument most often grounds a refusal, how stable cassation really is.
Ministries of justice, judicial councils, regulators — a tool for the ex-post assessment of enforcement: how a norm actually worked, not how it was intended to.
An evidence base for legislative initiatives and for monitoring the impact of norms — decisions on data, not on assumptions.
International institutions, academia, donors and technical partners — access to reproducible empirical jurisprudence at corpus scale.
Not search over case law — but a measurement of how a norm actually worked.
You see the structure of enforcement, not isolated cases: where one approach dominates and where practice is split.
It becomes clear which outcome is typical for a norm, and why — a basis for consistent, foreseeable decisions.
Every conclusion is a figure with context and a source that can be checked and reproduced — not an impression.
Not a “here’s what turned up” digest, but a problem map and an evidence matrix for the decisions of courts, reforms and policy.
A cheap deterministic filter removes the bulk of the repetitive work, and the expensive model is brought in only for the disputed cases and the final check. Example from a run: analysing 16,460 rulings cost roughly $1.5 instead of an estimated ~$233 — and yielded higher reproducibility.
The methodology operates inside a versioned agentic workspace. A research protocol and six types of sources pass through eight stages of analysis. The result is routed either to the adjudication track or to the lawmaking track and is stored as a verifiable work package containing sources, facts, analysis, checks, a log and metadata.
We do not ask a language model to 'tell us what the law says.' We first build a research protocol: we frame the question, fix the version of the norm, the period and the jurisdiction, and we define the unit of analysis, the boundaries of the corpus, the criteria for including and excluding materials, the classification rules and the ways of verifying the result.
Only then does an LLM join the work with the documents.
The source of law is the document, not the model. The output of the research is a verifiable chain of reasoning, not a persuasively written paragraph.
Here a language model is not a legal authority, a judge or an autonomous expert. It scales intellectual work: it helps process large bodies of decisions, locate relevant passages, compare arguments, surface recurring patterns in how a norm is applied and form hypotheses that a human then verifies.
How the method and the citation engine work
NormaLab is a public research platform. NormaTrace is its methodological and technological core for analysing the application of the law.
In a typical AI scenario, the unit of interaction is treated as a single query or chat. In our methodology, the workspace is the atomic unit of legal work: a bounded, versioned project environment deployed in OpenAI Codex, Anthropic Claude Code or a compatible agentic tool.
OpenAI CodexAnthropic Claude Code
The method lives not in a chat's memory, but in a versioned workspace.
Such a workspace stores more than the exchange with the model. It holds together:
In OpenAI Codex, persistent project rules can be anchored in AGENTS.md, repeatable procedures in skills, and specialised roles in agent configurations. In Anthropic Claude Code, the same logic is implemented through CLAUDE.md, skills, subagents, hooks and permission settings. For interoperability, a single set of project rules can live in AGENTS.md, while CLAUDE.md imports them and adds only platform-specific instructions.
Methodological precision is not achieved through one long instruction text. Persistent rules are anchored in version control; repeatable procedures are moved into skills; roles separate analysis, critique and verification; deterministic scripts check citations and structure; permissions restrict access; and legally significant conclusions require mandatory human sign-off.
Instructions guide the model's behaviour, but they are not an absolute technical constraint. What must happen without exception is enforced by validators, hooks, tests, access policies and human review. Instructions guide; validators and human review enforce.
Not a chat, but a workspace. Not a lucky prompt, but an executable methodology. Not a model's answer, but a verifiable work package.
legal-research-workspace/
├── AGENTS.md # shared project rules
├── CLAUDE.md # @AGENTS.md + Claude-specific instructions
├── protocol/ # question, scope, taxonomy, criteria
├── sources/ # primary materials and register
├── .agents/skills/ # skills for OpenAI Codex
├── .codex/agents/ # specialised Codex roles
├── .claude/skills/ # skills for Claude Code
├── .claude/agents/ # specialised Claude Code roles
├── tools/ # search, computation, citation resolver
├── verification/ # tests, samples, checklist, human sign-off
├── artifacts/ # intermediate and final outputs
├── logs/ # run log
└── metadata/ # versions, run_id, model, cost, statusThe structure is illustrative: the specific files depend on the platform and the organisation's policy.
The methodological core is stored in portable text rules, skills, data schemas and tests; platform-specific adapters connect them to OpenAI Codex, Anthropic Claude Code or a compatible environment.
The workspace executes the method. The work package preserves the evidence that it was executed. The lawyer makes the decision.
Depending on the task, the research corpus may include six types of materials, each with a different legal status.
The exact text of a norm, with its version and date of effect.
Court decisions from different instances that form the research corpus.
Legal positions that set reference points for practice.
Case law of the ECtHR and other relevant international sources.
Scholarly publications and doctrinal sources.
Lawmaking materials, statistics and other data on the effects of regulation.
These materials are not treated as equivalent. For each source we record its type, jurisdiction, date, version, identifier and place in the legal argument. Doctrine does not stand in for statute, a single court decision is not presented as settled practice, and a statistical correlation is not presented as proven legal causation.
The exact text and version of the norm, the subject of the research, the period, the jurisdiction and the unit of analysis are fixed. Before any bulk processing, the hypothesis, the relevance criteria and the scheme for the forthcoming classification are defined.
Human checkpoint: The question, the boundaries of the research and the definition of done are approved.
Court decisions are collected, normalised and deduplicated. The inclusion and exclusion criteria, possible data gaps and the degree of corpus coverage are documented. Before a full run, the method is tested on a small sample.
Human checkpoint: The quality of the sample, the sufficiency of coverage and the acceptable limitations are confirmed.
The outcome of the case, the key arguments, the factual and procedural context and the legal constructs applied are extracted from each decision. Unambiguous operations are carried out by deterministic rules wherever possible; an LLM is brought in where semantic classification is required.
Human checkpoint: A sample of the classifications is reviewed, and errors are fed back into the rules, the taxonomy or a skill.
On the basis of recurring patterns, a verifiable hypothesis about the problem is formulated. Exceptions and competing explanations are analysed so that a peculiarity of the data or a classification error is not mistaken for a defect in the application of the law.
Human checkpoint: The hypothesis has withstood adversarial review and a check against alternative explanations.
The presumed nature of the problem is determined: ambiguous interpretation; inconsistent application; the absence of a single legal test; a gap or a conflict of norms; a lack of clarity in the norm or its obsolescence; systemic inefficiency of the regulation; or a peculiarity of a specific category of cases that does not amount to a general problem.
Such a classification is a reasoned research hypothesis, not an automatic verdict produced by the model.
Human checkpoint: The legal reviewer confirms or rejects the characterisation.
Possible approaches to resolving the problem are developed: an interpretive test, a legal position, a change in the practice of legal reasoning, a clarification of procedure or an amendment to the norm itself. For each option, the grounds, counter-arguments, relevant case law and limitations are set out.
Human checkpoint: The options do not conflate a description of the law as it stands with a normative proposal.
The possible consequences of the proposed solutions are examined for the consistency of practice, the predictability of the application of the law, the rights of the parties to proceedings, the workload on courts and public authorities, and transitional regulation.
This is a scenario-based assessment, not a promise of a precise prediction of the future.
Human checkpoint: Assumptions and uncertainties are stated explicitly.
The user receives not only the final text but also the materials needed to verify it: a register of sources; a description of the corpus and the selection criteria; the classification scheme; tables and an evidence matrix; links to specific passages in the documents; a log of the key decisions and rejected hypotheses; a statement of uncertainties and limitations; and details of the version of the methodology, the workspace configuration and the research run.
The result can not only be read but also traced to the primary materials, checked on a sample and, if necessary, re-run.
Human checkpoint: The final package is signed off by the responsible specialist and carries a verification status.
After the analysis, a fundamental fork appears.
Adjudication track
Lawmaking track
The system helps to prepare and substantiate such a characterisation. The final decision on the nature of the problem and the way to resolve it is made by a human.
result/ result sources/ sources facts/ facts analysis/ analysis verification/ checks logs/ log metadata/ metadata The document is the interface to the result. The work package is the object of verification and handover.
No source, no strong conclusion.
Every material claim must lead to a specific document and passage.
Versions of the data, the methodology, the instructions, the skills and the configurations, along with the outputs of intermediate stages, are preserved.
The task context, the rules, the roles and the checks must outlive a single session and remain available for audit.
Analysis, verification, critique and final sign-off must not merge into a single operation.
Where a task can be reliably solved by a rule, a search or a computation, an LLM is not used.
Project instructions and skills guide the work; validators, hooks, permissions and tests provide the mandatory constraints.
Results are presented with regard to data quality, sample size and the uncertainty of the classification.
The system measures and structures; a human interprets, verifies and decides.
Technological persuasiveness is no substitute for conformity with the law, case law and doctrine.
The methodological core must not depend on a single model provider or interface.
Legal Copilot Ukraine
NormaLab shows how to turn a large body of legal information into verifiable research.
Legal Copilot Ukraine extends the same principles to everyday legal work: personal AI workplaces built on ready-made agentic environments, verified sources, project instructions, roles, skills, structured work packages, logging, checkpoints and mandatory human review.
Personal AI workplaces for judges, attorneys, government lawyers, universities and legal-transformation teams.
This diagram describes a research process and a way of organising work in an agentic environment. It does not imply the automatic generation of a judicial, administrative or legislative decision.
Each case is a real run with stored artifacts. Figures are given as facts about the runs, not as a final legal result.
Part 2 of Article 61 of the Code of Ukraine on Bankruptcy Procedures, 2025. Funnel 260→210→70 substantive rulings; liability denied in ~60% of cases; the dominant ground — "causation not proven" (~76% within the subgroup). Run cost — $3.92. The conclusion was confirmed by a second, independent run.
16,460 rulings; ~58% of decisions survived, ~42% were destabilized; chamber spread from ~68% to ~43%; taxonomy coverage 98%. Cost — about $1.5 instead of an estimated ~$233.
The system itself wrote a caveat into the report: the "stability rate" must not be used to assess a judge's quality without adjusting for case mix. Cost — $0.
79 positions from 8 cases with precise paragraph references; unverifiable cases were excluded, and early "from memory" quotations were caught and replaced by the QA loop. A prototype for preparing materials for international tribunals.
The flagship report passed internal control with a "sufficient" status, but an independent external audit returned the verdict "return for revision" (conditionally positive, five blocking comments). This is our most valuable honest result.
Task: how many cassation filings reached the Supreme Court in 2025 and how many were accepted for review. Corpus — 112,983 SC rulings (64,812 cases). Result: accepted for cassation proceedings ≈22% of cases (≈14.3–15.3k); the dominant outcome is denial of opening (~46%), and the administrative cassation denies roughly 7× more often than it opens. Run cost — $1.66.
Download reportTask: whether the bankruptcy institution performs its rehabilitative (restorative) function. Result: of 580 cases closed in 2025, 81 ended with restored solvency (13.97%), but in the business perimeter (legal entities and sole proprietors) only 4 of 362 (1.1%); in 2018 — 0 of 526 (0%). Conclusion: for business, bankruptcy remains a predominantly liquidation, not restorative, mechanism. Figures verified against primary court acts.
Download reportTask: how courts apply Part 2 of Article 61 of the Bankruptcy Code — the subsidiary liability of persons controlling the debtor. Result: of 161 decisions reviewed, only 20 (12.4%) were relevant; 45% were upheld, 20% were first-instance refusals, 5 decisions were changed on review, and 9 showed defects of application. Key finding: the problem is not how often the norm is applied but its instability — there is no single legal test linking real control, specific conduct, causation, and financial effect. The brief proposes a unified test of six-to-seven mandatory elements, grounded in ECtHR standards (Lekić v. Slovenia, Agapov v. Russia and others): piercing the corporate veil is permissible not under a “controlled, therefore liable” formula, but only where a person genuinely exercised control, committed a specific bad-faith act, and that act caused the shortfall in the bankruptcy estate.
Task: how the Grand Chamber of the Supreme Court reviewed, in 2025, decisions of the High Council of Justice (HCJ) on the disciplinary liability of judges. Method: from 1,805 Grand Chamber rulings, 42 relevant acts (25 cases) were selected. Result: the complaint was granted (the HCJ decision quashed) in 42.9% of acts and 44% of cases; the rest were denied. Practice is binary — the Grand Chamber either fully quashes the HCJ decision or fully denies the complaint. The ≈43% quashing rate is no rubber stamp: the court genuinely intervenes in nearly half the cases. Applicants most often challenge insufficient reasoning (54.8%), incorrect assessment of evidence (47.6%), and wrong qualification of the offence (47.6%). Conclusion: outcome-stability is moderately high, but there is no single standard of “sufficient reasoning”. The report proposes a four-element test of judicial review, grounded in ECtHR case law (Oleksandr Volkov v. Ukraine, Albert and Le Compte and others).
NormaLab designs reproducible workflows that turn case files, case law, dissenting opinions, normative acts and procedural documents into not a short memo but a full, structured draft court decision — verifiable by its structure, argumentation and sources.
It analyses the case materials, relevant case law, dissenting opinions, legal norms and procedural documents in a single pipeline.
The output is not a search result or an abstract, but a structured draft of a judicial act: the facts, the parties’ positions, the reasoning, the holding and the operative part.
A separate explanatory document is produced: the logic of the draft, its advantages and how it differs from the existing act or alternative approaches.
Every element of the draft traces back to a norm, a fact of the case and a source — the decision can be checked, not just read.
Demonstration case
A judge’s disciplinary case: a critical procedural deadline missed during leave and self-isolation. The system generated a full alternative draft of the Grand Chamber ruling that keeps the correct outcome of the existing act (quashing the HCJ decision) but strengthens it — and added an explanation of its advantages that integrates the majority position and the dissenting opinion.
Key idea. It proposes a two-step test of disciplinary liability: (1) did the judge know of the critical deadline and have a real opportunity to act before leaving; (2) was the failure to act material, culpable and causally linked to the consequences.
Trust is the core currency. So we build it architecturally, not with promises.
The model does not write the text of a quotation — it only points to a source segment, and the verbatim text is supplied by a deterministic resolver. There is nowhere to enter an invention.
Every run stores a run_id, prompt versions, requests and cost — the research can be not only read but reproduced to the figure.
Conclusions are checked by independent review. Our most valuable honest result is when an external audit returned the flagship report for revision.
The system measures, the human decides. Internal review is useful but does not replace an external expert’s signature.
We publish not only conclusions but the way they were reached — methodologies, prompt versions, runs.
A plain-language, educational walkthrough with no tech jargon: how LLM agents are changing the analysis of case law — using model corpora from 2018 and 2025.
How many cassation filings reached the Supreme Court in 2025 and how many were accepted: a corpus of 112,983 rulings (64,812 cases), ≈22% accepted, with denial of opening (~46%) the dominant outcome.
Whether the bankruptcy institution performs its restorative function: 81 of 580 closed cases (13.97%), but only 4 of 362 (1.1%) in the business perimeter; 0 of 526 in 2018. Verified against court acts.
An interdisciplinary core at the intersection of law, data and public policy. Working with us means access to rare expertise, not just another service.
Law & case-law analytics
Leads legal methodology, the interpretation of case law and the quality of conclusions.
GAA8791@gmail.comData engineering & agentic systems
Owns the pipeline over large decision corpora, the reproducibility of runs and the citation engine.
LinkedInWe welcome co-authors, reviewers and partners across law, ML and public policy. Get in touch.
Pick an article or a category of cases — we will show you a verifiable analysis on a real body of practice. We reply personally.
Write to us about a pilot, research, or collaboration.