Public sector
AI agents an institution can account for.
For ministries, agencies, municipalities and state enterprises: what the platform records, what it lets a person stop, and what it does not claim. Each item below is something the software does today.
Poseidonka
A reminder that you set the boundaries.
A system card for every agent
Each agent has an EU AI Act Annex IV system card, generated from its configuration: intended purpose, model and provider, context size, tools and guardrails. It is built on request from the settings in force, so it cannot drift from what the agent actually runs with.
Risk level and a fundamental-rights assessment
An administrator records the agent's EU AI Act risk level with its Annex III category, justification, assessor and review date. For limited and high risk the platform keeps a Fundamental Rights Impact Assessment (FRIA) with the agent, and refuses one for an agent that has no such level.
A person approves the tools that matter
Mark a tool in the agent's settings and every call to it pauses the run until a person approves or rejects it.
Written rules, honest about enforcement
The workspace keeps a constitution: prohibitions, requirements and permissions, each with its scope and consequence, with its amendments kept beside it. Every rule is marked by whether the platform actually enforces it, so a rule nothing checks never passes for a live constraint.
A record you can check for tampering
Governance events go into a ledger where every entry is hashed and linked to the one before. A verification recomputes the whole chain and names the first broken entry if there is one; the workspace shows the result next to the ledger.
Every run can leave with you
Each run's full event stream exports as a file, and each conversation exports with its messages. An auditor reads what happened from the platform's own record, not from a summary written afterwards.
Your choice of model, on your own key
Agents run on the model you choose for them, open-weight models included, and a workspace can connect its own key for a model provider instead of the platform's. The system card names the model and provider each agent uses.
What this page does not claim
No certification is held, so none is claimed: no FedRAMP authorisation, no ISO 27001 or SOC 2 report, and no attestation of an information-protection system under Ukrainian law (KSZI). If your procurement requires one, we will say so plainly before a pilot rather than after it.
The hosted service processes data in the United States and Ukraine, as set out in the Privacy Policy. When an agent uses a hosted model, the request goes to that model's provider. If your rules require data to stay in one country or on your own infrastructure, raise it at the start: what is possible there is agreed in a contract, not switched on in settings.
The platform fills in what it knows. A system card, a risk level or an impact assessment is still your institution's document and your institution's judgement.