Policy and validator registry
Version reusable policies for prompts, retrieved context, responses and agent tool requests using deterministic rules, customer-approved models and domain-specific validators.
Sovereign trust · Roadmap
Define, test and enforce approved AI behaviour.
Validate AI inputs, outputs and agent actions; evaluate changes before release; and monitor behaviour in operation while policies, evaluation data and evidence remain inside the customer-defined boundary.
Status: Roadmap. This application is being developed for open-source release. Its repository, OSI-approved software licence, model-specific terms, version, provenance and supported deployment modes will be published with the applicable release.
Core capabilities
Designed around accountable organisational work.
Exact coverage, deployment requirements and limitations are documented for the selected version and use case.
Structured output contracts
Require schema-valid responses and organisation-defined constraints, then reject, redact, retry, correct or route failures for human review according to approved handling rules.
Evaluation and release gates
Compare prompts, models, retrieval configurations and agent workflows against versioned datasets, expected outcomes, expert annotations and customer-approved thresholds.
Inline runtime enforcement
Apply prioritised policy stages before model calls, before responses are released and before agents execute tools, with explicit monitor-only, timeout, fallback and enforcement modes.
Observability and temporal evidence
Trace sessions, retrieval, model and tool steps; monitor selected signals; and connect policy versions, evaluations, reviews and outcomes through the Sovereign Unified Knowledge Graph.
Product boundary
Quality and security remain distinct
AI Guardrails governs the expected behaviour and quality of approved workflows. AI Defense maps the wider AI estate, inspects its supply chain and responds to adversarial activity. Both share customer-controlled evidence and enforcement interfaces.
Adoption outcome
Begin with explicit failure modes
Select one workflow, define what acceptable behaviour means, build representative evaluation data and agree monitoring, escalation and rollback conditions before enforcing policy in production.
Sovereignty matrix
Evaluate the deployment across six dimensions.
Product-specific answers are established through the deployment architecture and dated evidence record.
DataOperationalTechnologyJurisdictionKnowledgeExit
Review the trust model
Connected Roadmap capabilities
See how this responsibility fits the platform.
Every connected capability shown here is part of the current Roadmap.
SovereignEU AI Defense
Map the AI estate, inspect its supply chain, attack-test systems and contain runtime threats.
SovereignEU Assistants
Permission-aware assistants, agents, workflows and connectors with human approvals.
Custom Model Engineering
Model selection, adaptation, evaluation, documentation and transfer for defined use cases.
A practical first step
Begin with explicit failure modes
Select one workflow, define what acceptable behaviour means, build representative evaluation data and agree monitoring, escalation and rollback conditions before enforcing policy in production.