The EU AI Act sets obligations that vary by risk tier and use case. This page maps ValGuard's runtime controls to operational questions reviewers ask about logging, human oversight, and failing closed. It is not legal advice, a conformity assessment, or a claim that ValGuard makes a system Act-compliant by itself.
What ValGuard covers in this context
ValGuard is a runtime validation and policy-enforcement layer for LLM calls and agent steps. Deterministic rules run on every completion before the next model, tool, or customer-facing action. Audit rows carry rule IDs. Shadow mode measures would-blocks before enforce.
| Act-oriented concern | ValGuard mechanism | Limit |
|---|---|---|
| Traceability of model outputs that drive actions | Per-request validation audit with rule IDs | Does not replace full system logging |
| Preventing known-bad structured actions | Fail-closed block / reask on packs | Only encoded rules fire |
| Human oversight hooks | Playbook branches and human approval where configured | Not a full case-management system |
| Pre-production evaluation | Shadow mode on hosted or Enterprise self-host | Hosted shadow still processes content |
Suggested control mapping (engineering)
- Inventory high-risk agent steps (refunds, eligibility, medical messaging, credit).
- Attach schema, policy, and domain packs; start in shadow.
- Enforce with fail-closed on those paths.
- Export audit evidence for reviewers; keep payload logging policy explicit. Trust.
- Prefer Enterprise self-host when residency requires VPC. Self-host docs.
Related packs and policies
Honest limits
- ValGuard does not classify your system under the Act.
- Fluent false claims without a rule signature can pass.
- Cloud provider choice and training-data obligations sit outside this product.
- Assessments and legal opinions stay with your counsel and notified bodies where required.
Diagram
flowchart LR
A[Agent / app] --> V[ValGuard rules]
V -->|pass| T[Tool / customer action]
V -->|block| H[Human oversight path]
V --> L[Audit with rule IDs]
Next step
Contact for enterprise diligence, or start quickstart in shadow mode on one high-risk agent.