Each row mirrors a domain that procurement officers and CISOs assess. On the left the procurement requirement, on the right the concrete AIonicOS mechanism — verifiable, not asserted.
Criterion
Requirement
AIonicOS
EU data residency & inference
All data and every model inference must physically reside and run in the EU/Germany — at rest, in transit and during inference (GDPR Art. 44–49, BSI C5, Schrems II).
Single-tenant architecture: the full stack is deployed per customer on a dedicated EU host, with its own isolated Postgres/Apache-AGE, Qdrant, Valkey and SeaweedFS — no shared data plane. The residency gate (`residency.py`) enforces a strict provider allow-list under `data_residency=eu`, including a fully-EU path (Mistral FR + self-hosted Ollama); a per-model block-list catches models not covered by regional processing.
Operational autonomy & third-country access protection
No provider operator, hyperscaler or third-country authority may gain access or compel disclosure without your authorisation — the system must allow no silent operator override (CLOUD Act / FISA 702, BSI 'Cyber Dominance').
Fail-closed residency gate enforced at preflight (HTTP 409 before any spend) and at runtime (`DataResidencyViolation`) — with no operator bypass: `?preflight=skip` cannot override residency FATALs, and every skip is audited. The dedicated host offers no provider-side control plane to subpoena; the fully-EU model path removes any CLOUD Act nexus for inference. Chinese-jurisdiction models are hard-blocked under `eu`.
Exit & portability (EU Data Act Ch. VI)
Full data and workflow export plus a contractually and technically supported exit path — no vendor lock-in (EU Data Act switching and portability rights).
Workflows are declarative, human-readable YAML DAG templates — portable, inspectable artifacts the customer holds, not opaque vendor config. Data lives in open, exportable stores (PostgreSQL/Apache-AGE, Qdrant, Valkey, SeaweedFS); artifacts are addressable via stable `/api/v1/artifacts/<wf>/<label>` JSON endpoints. Multi-provider routing (LiteLLM) prevents model lock-in; the whole platform is the customer's dedicated deployment (git pull + docker compose).
Transparency & auditability
Every AI action, model call, cost and policy decision must be logged as durable evidence and be queryable — suitable for EU AI Act record-keeping (Art. 12) and GDPR accountability (Art. 5(2)).
A SQL-queryable `notification_log` records every governance event with structured categories (residency block, runtime violation, skip veto, MCP connector). A cost ledger writes every LLM call to `llm_cost_events` (model, cost in EUR, latency); Temporal provides full event replay; Prometheus counters expose enforcement metrics. Langfuse, the Temporal UI and Grafana are pre-wired.
Model & supply-chain control
The buyer must control which models and providers process their data — with a documented EU allow-list and a path to a fully-EU / self-hosted supply chain free of third-country dependency (BSI C3A, NIS2 supplier risk).
Multi-LLM routing across multiple providers with an enforced EU allow-list under `eu` and a fully-EU path (Mistral FR managed + self-hosted Ollama on the customer host). New model IDs are verified against official provider docs and registered before use — no hallucinated endpoints. A per-model block-list catches models not covered by regional processing; managed failover keeps availability without violating residency.
Defence against AI-specific attacks
The system must detect and flag AI-specific threats — prompt injection, jailbreaks, output-side data exfiltration, tool abuse, package hallucination (OWASP GenAI Top-10, EU AI Act Art. 15 robustness, BSI).
PromptGuard, a defence-in-depth architecture (seven layers live), detects and flags via a typed middleware: an input scanner, an injection classifier, an output rogue-string detector, a package-hallucination validator (PyPI/npm check) and an egress sanitiser, plus memory-write quarantine and SHA256 integrity checks of MCP tool descriptions. A continuous Garak red-team CI tests effectiveness. Honest framing: several phases currently run in observe-only mode (detect/flag); the metrics come from internal benchmarks.