Architectural concept / Legal AI and Computational Law / Agentic Engineering
Legal Copilot Ukraine: Verifiable Agentic Legal AI
An architectural concept for exchanging verifiable work packages: sources, versions of legal rules, checks, and rejected hypotheses.
A public architecture and the related open minius prototype for a personal legal workspace.
- Status
- Architectural concept
- My role
- Concept · Systems architecture · Artifact standard · Pilot design
- Case published
- Updated
Starting point
Legal institutions operate with different sources, procedures, accountabilities, and access boundaries. Shared infrastructure therefore need not mean one central product or a single model provider.
Legal Copilot Ukraine proposes a network of personal agent workspaces. The result standard is shared, while data, permissions, and working practices remain under the control of each institution or professional.
Problem to solve
Answer text alone does not reveal the sources used, rejected hypotheses, checks, or decision history. When that result reaches a second lawyer, verification can mean repeating the work from the beginning.
Interoperability therefore requires a transferable work package, not merely a shared model interface.
My role
- author of the concept and supporting material;
- systems architecture and institutional boundaries;
- verifiable work-package standard;
- bounded-pilot and training-layer design.
Personal agent workspace
An environment combines AGENTS.md, roles, skills, workflows, templates, and evaluations. These elements define a process contract rather than an agent personality: admissible sources, required artifacts, checks, and the point of human decision.
Access layers separate public law, internal institutional knowledge, case material, confidential data, and generated artifacts. A model does not gain access merely because access is technically possible.
Transferable work package
The proposed package preserves the result, sources, facts, analysis, verification, logs, and metadata. A second professional can see the inputs, applied rules, uncertainty, and the person responsible for approving the conclusion.
source → evidence → fact → rule → conclusion → human decision → handoff| One final file | Verifiable work package |
|---|---|
| sources disappear into correspondence | sources and versions are listed |
| rejected hypotheses are invisible | rejections and reasons are preserved |
| verification has to restart from zero | another professional continues from a recorded state |
| accountability is unclear | the human decision point is explicit |
Bounded pilot design
A pilot should cover one real workflow with one explicitly unacceptable failure. Criteria include package completeness, citation fidelity, confidentiality, reconstruction by another professional, and a visible point of human accountability.
What the project demonstrates
The project publishes the architecture in several languages, a material library, a synthetic package demonstration, and a pilot framework. It shows how an artifact standard could connect different tools and institutions without creating one monolith.
What the project does not claim
This remains a Concept and pilot proposal, not an operating national infrastructure. It does not claim deployment, savings, or quality from landing-page figures alone. The architecture deliberately avoids treating one universal workspace or a higher agent count as a substitute for accountability.
Limitations
This case presents Legal Copilot Ukraine as a distinct architectural concept and does not transfer properties from other projects to it. Quantitative claims are published only after separate fact-checking, and the material is not legal advice.
08 / Evidence
Public artifacts and related material
Legal Copilot Ukraine
Public concept overview.
- Access
- public
- Publisher
- Legal Copilot Ukraine
- Date
Swarm instead of a platform
Article on distributed architecture.
- Access
- public
- Publisher
- Sergey Avdeychik
- Date