Invarians provides audit inputs, not compliance. Whether an obligation is met is determined by your compliance function, on your facts; accountability is not outsourceable. What follows is the mapping: what each text asks, what changes when the actor is an AI agent, and the input a route record provides.
| Obligation | What changes with an agent | The input Invarians provides | In your files |
|---|---|---|---|
| MiCA record-keeping RTS (CDR (EU) 2025/1140), art. 7(1): records kept immediately after the transaction | The actor is software; there is no human contemporaneous witness | Route-state records timestamped at observation; third-party verifiable signatures arrive with the next version | Audit trail |
| MiCA record-keeping RTS, art. 2(1): reconstitution of each key stage, corrections ascertainable | Agent decisions interleave with market state at machine speed | Signed, timestamped context series; integrity between the parties today | Records reconstitution |
| DORA, art. 10(1): detection of anomalous activities and potential material single points of failure | Route anomalies can stay invisible on status pages; the September 3rd case is the documented example | Per-route envelopes and dated deviations | Anomaly detection |
| DORA, art. 11(8): records before and during disruption when continuity plans activate | Reconstruction after the fact is forensics, not testimony | Continuous archive, no missing day on the ten CCTP corridors measured | Records before and during |
| DORA, art. 17(2): all ICT incidents recorded | Agent-induced slowdowns must be evidenced or they do not exist | Deviation census against each route's measured baseline | Incident records |
| AI Act, art. 26(6): deployer keeps system logs at least six months, for high-risk systems; whether a settlement agent is classified high-risk is for your legal analysis | Internal logs are produced by the agent's own operator | An independent source of operational evidence alongside those logs | Independent evidence |
| FSB sound practices on AI agents (Oct 2026; non-binding guidance, paraphrased): oversight scaled to autonomy, human-approval thresholds | Thresholds need a neutral input or they reference the agent's own view | Signed context consumable by your own policy, before the agent acts | Guardrails input |
| SR 26-2 (Federal Reserve, Apr 2026, superseding SR 11-7): ongoing model monitoring and effective challenge | The agent's view of the world needs external challenge data | Route-state series usable as model monitoring MI | Ongoing monitoring MI |
Everything above describes the service as it runs. The one exception is stated where it occurs: third-party verifiable signatures belong to the next version, built with our first design partner.