Technology & Architecture
Three layers, one principle: accesses to models and protected knowledge are mediated and recorded before execution.
Three layers
The userspace comprises the portal, queries and project management — the layer users work with. The governance kernel is a dedicated service that mediates access to language models and protected data; the architecture is designed so that model credentials live with the kernel, not in the userspace. The data and model layer holds documents, the knowledge graph, vector indexes and the locally operated models.
Mediation before execution
A request that would touch a model or protected knowledge goes to the kernel first: it checks, among other things, deployment model, target zone, budgets and data class. The decision is recorded before the effect takes place — only then does execution proceed. Denied operations are recorded just like admitted ones.
Knowledge spaces and controlled sources
Knowledge is organised into projects or mandates, with separate stores per project. Queries operate within the respective project; access rights are granted at project level.
Audit and governance decision records
Every kernel-mediated decision produces a durable record: which operation, admitted or denied, which checks carried it, which data class was touched and who asked (pseudonymised). The page “Anatomy of a governance record” shows the structure using two real examples.
Deployment models
Local: processing and models in your infrastructure. Managed private: operation in a dedicated environment managed by us. External model connections: only in deployment models configured for them; what is claimed there is stated as a separate, deployment-scoped claim on the evidence page.
Current limitations
Access denials outside the kernel evidence the operation, the resource and the enforcement layer; the exact rule is currently not stored verbatim there.
Independently maintained knowledge bases with controlled activation and rollback are designed but still in development.
Technical claims with evidence tracking
The following claims are projected onto this page from the machine-readable claim source; their current state is kept on the evidence page.
Connections to external models are possible only in deployment models configured for them, and only after a kernel-mediated, auditable approval.
Kernel-mediated decisions name the decisive rule identities (decisive_rules) and the policy engine version. Access denials evidence the blocked operation, the affected resource and the enforcement layer; the exact rule is currently not stored verbatim there.