The 2026 Digital Omnibus on AI simplifies the paperwork — and sharpens the controls that only a runtime layer can deliver. Aperion already enforces those controls and produces the machine-readable evidence regulators can now demand on the spot.
It's a simplification package — lighter process, same (and in places stronger) substance:
Net effect: less form-filling, but a higher bar for live enforcement and on-demand proof — the layer Aperion provides.
| What the Omnibus expects | How Aperion delivers it | Status |
|---|---|---|
| Runtime prompt guardrails against prohibited generation | Semantic policy engine + destructive-prompt rules enforced on the request path before the model is called | Covered |
| Content classification & filtering of inputs and outputs | PII / toxic / regulated-content classification at ingest and egress, with data-leaving-the-perimeter detection | Covered |
| Usage restrictions & abuse detection | Per-identity and per-route allowlists, budgets and adaptive rate limiting; multi-turn abuse detection | Covered |
| Notice-and-action & corrective measures on misuse | Every blocked action is sealed into the audit chain and fanned out to alerting; human approval/override closes the loop | Covered |
| Obligation | How Aperion delivers it | Status |
|---|---|---|
| Art 9 — Risk-management system | Per-route, per-identity policy graph with tiered risk rules (Critical → Low), surfaced as live evidence | Covered |
| Art 10 — Data & data governance | Inference-time data-class controls, information barriers (group × content-label), signed agent data-class scopes | Covered |
| Art 11 — Technical documentation (Annex IV) | Auto-generated model inventory, examination reports and a signed AI bill-of-materials, exportable on demand | Partial |
| Art 12 — Automatic record-keeping | Per-request logs (input, output, model, identity, decisions) in a tamper-evident, independently verifiable audit chain | Covered |
| Art 13 — Transparency to users | Response-side disclosure banners and headers for downstream interfaces to surface | Partial |
| Art 14 — Human oversight | High-risk actions route to a human approval queue; government-grade identity step-up (IAL2/AAL2 + liveness) holds the action until a verified human approves; oversight analytics measure override rate, approval latency and outlier reviewers | Covered |
| Art 15 — Accuracy, robustness, cybersecurity | Response-quality & hallucination scoring, regression detection, degradation alerting; destructive prompt/tool blocking | Partial |
| Art 17 — Quality-management system | Continuous multi-pass review program (PII, rescans, adversarial, anomalies, nightly report) producing audit-ready findings | Partial |
| Obligation | How Aperion delivers it | Status |
|---|---|---|
| Art 26 — Deployer obligations | Verified natural-person identity per action; every request attributed to a user, group and department; tamper-evident logs | Covered |
| Art 72 — Post-market monitoring | Live usage analytics, perimeter-egress monitoring and a nightly organisation report feeding the monitoring plan | Partial |
| Art 73 — Serious-incident reporting | Critical events and high-severity findings captured and fanned out, with a pre-formatted incident envelope ready to file | Covered |
| New regulator power | How Aperion answers it | Status |
|---|---|---|
| Order access to the AI system & retain all data/documents | Read-only Conformity Console surfaces live, per-article evidence; the tamper-evident chain is the retained record | Covered |
| Evidence in machine-readable format by electronic means | Structured, per-article JSON evidence plus a regulator-facing dashboard | Covered |
| Take a system offline / limit blast radius during enforcement | One-switch emergency stop (off / read-only / halt) across model, tool and agent-to-agent traffic, fully logged | Covered |