Case analysis platform · lawyers, in-house counsel, firms of 1–50
Cassation reads multi-volume case files, answers with a reference to volume, page and paragraph, puts its analysis through model debates and leaves the decision to the lawyer. The subscription is hosted on our server or on your firm's hardware; a fully local model stack is built as an integration project.
No card required. We run the pilot together with you — on one real case.
How do the security guards' testimonies differ from the inspection report, and what follows from that for the motion?
The workspace interface. The model does not write the volume and page number — they are inserted by code from a closed list of fragments.
The difference that decides in court
A general-purpose AI confidently names articles, dates and sheet numbers. In the courtroom such confidence is costly: one unverifiable reference destroys trust in the rest of the analysis.
Cassation is built the other way round: first the case fragments are placed in a closed list, then the model formulates a claim, and the sheet reference is inserted by deterministic code. If there is no match, the claim is honestly marked as unverified.
Not a single unverifiable fact in the document you take to court. Core platform invariant · under test
An ordinary AI assistant
Cassation
A closed list of fragments · Analyst — Sceptic — Opposing counsel — Judge debates · an environment chosen by the firm · a draft approved by a human.
How this is bought
A subscription means seats and storage deployed with us or on the firm's hardware. An integration project is for when you need your own perimeter and a model stack: from commercial models under your control to a fully local stack.
We are responsible for the infrastructure, updates and backups. The organisation's space is isolated, and you can start on the day of your request.
The same platform on your hardware: files never leave the office. Installation, updates, backups and monitoring are our job.
Deployment in your perimeter with a model stack: commercial models under your control, a hybrid stack, or a fully local stack with no external calls.
A server is included
The server can be yours or ours. A Mac mini with 8 GB of memory is a working configuration: the system is built from the start without heavy dependencies. A fully local stack needs spare memory for models — we size it during the project audit.
Models and roles
In the subscription the platform works with commercial models: your own keys or our balance. If nothing at all may go outside, that is a separate integration project in your perimeter. The setting for "what goes where" is visible in the interface, not hidden in the configuration.
Claude, Gemini, GPT, GLM — directly or via OpenRouter as a universal gateway. The judge in the debates never coincides with the author of the answer: otherwise it praises itself.
your own keys or our balance · file caching lowers the bill
Fact extraction and recognition run locally, legal analysis and debates run on external models. The full text of the case is sent nowhere except the service you choose.
a balance of quality and privacy
All roles, recognition and arbitration of disputed pages run on models inside your perimeter. There are no external calls at all.
spare hardware needed for models · lower quality
| Role | What it does | Model requirement | Typical setup |
|---|---|---|---|
| Analyst | Legal analysis of the case, answers with citations | Maximum depth | Claude · Sonnet/Opus |
| Sceptic | Attacks the logic of the answer point by point | Cheap and high-volume | Gemini · Flash |
| Opposing counsel | Builds an independent position for the opposing party | A different vendor, a different corpus | GLM |
| Judge | A verdict on each claim, with references | Not the model that wrote the answer | configurable |
| Fact extraction | Populates the fact graph from the volumes | Cheap, batch | any compatible one |
| Recognition | Double OCR with arbitration on discrepancies | Local, free | Tesseract + EasyOCR |
| Embeddings | Hybrid search: full text + meaning | Local or external | configurable |
The "role → provider and model" table is configured in the organisation's admin area, with a fallback in case a provider fails. Anthropic, OpenAI, Google, Z.ai, OpenRouter and any service with a compatible interface, including a local one, are supported.
Levels of involvement
Solo
One seat, 25 GB of storage. Your own cases, chat with citations, debates, drafts, deadlines. Works in the browser, the app and Telegram.
Team
Payment for seats and storage per environment. Roles and permissions, shared cases, a team chat per case, limits per lawyer and per case, and a breakdown of spend by case for re-billing the client.
Company
Individual terms and a contract with data processing annexes. If required, your own perimeter with a local model stack: an integration project from 30 000 €.
How we sell
We do not start with a price list. We start with a review of your case and training for the team — because the platform's strengths only become visible on live volumes, not in a demo environment.
We look at how the case is structured and where you lose time: the volume, the quality of the scans, the deadlines, who prepares the documents.
We deploy the hosting — our server or the firm's server — load the files and work together with you, rather than giving a guided tour.
We configure roles, keys, limits and the linking of Telegram and the apps. We run training by role: lawyer, assistant, administrator.
Updates, backups, monitoring of model spend, and reviews of new scenarios. The Academy remains available for new employees.
No prepayment. We delete the pilot's files on your request.
Cassation Academy
The platform is strong in the hands of a lawyer who understands exactly what they are checking. That is why training is part of the delivery: short courses on working with citations, debates, deadlines and drafts — plus materials for administrators of your own server.
6 modules: from loading volumes to approving a draft.
Installation, updates, backups, model keys, limits.
How to read the debate verdict and the contradictions radar.
Regular sessions on practical scenarios.
Privacy
On the firm's server, data goes nowhere except the model provider you choose. On our server, organisation spaces are isolated from one another.
The relay carries an encrypted stream and does not decrypt it — the key never leaves the firm's machine. In the office the connection runs directly over the local network.
Model keys are encrypted in the database and are never shown in full once saved. Every significant action — approval, deadline confirmation, chat linking — is written to the audit log.
We will show you the platform on your files, calculate model spend and decide together whether you need a firm's server or ours is enough.