Models that reason over real regulation — the statute, the sanctions list, the rule as written — instead of a summary of it. Compliance logic that a supervisor can audit line by line.
The model's authority comes from primary sources, versioned and traceable to their origin.
The model is trained on regulation as published — statutes, directives, and supervisory guidance — not paraphrased blog summaries. Its reasoning maps back to the text an examiner would open.
It reasons over sanctions regimes and designated-party lists as structured, versioned data. The model distinguishes a listed entity, an alias, and an ownership-linked party rather than matching on surface names.
National, regional, and cross-border rules are held as distinct, sometimes conflicting layers. The model reasons about which regime governs a given transaction rather than flattening them into one.
Every rule the model reasons over is pinned to an effective date and version. When a regulation changes, the model can tell you which text applied at the time of a past transaction.
The model's job is to show its work so a human officer can accept or overturn it.
Every screening or classification output carries the specific rule and clause it relied on. An officer reviews the citation, not just a risk score, before acting.
The model reasons about structuring, layering, and typologies against the ledger's tamper-evident transaction history. It surfaces the chain of evidence behind a flag rather than an opaque signal.
When it clears a match, it explains why the alias or ownership link does not apply. False-positive suppression is transparent so supervisors can test the logic.
The model recommends; a designated compliance officer decides. It is built to inform sanctions and AML judgments, never to auto-reject or auto-approve a party on its own authority.
The design assumes the output will be examined by a regulator, and structures itself accordingly.
Screening decisions, the rule version used, and the reviewing officer are written to the hash-chained ledger. The audit trail is tamper-evident by construction, not by policy promise.
Given the same transaction and rule version, the model's reasoning can be reproduced and re-examined. A supervisor can rerun a past decision against the exact inputs it saw.
Regulated data and the compliance model stay inside the jurisdiction's borders. Sensitive customer and transaction data never transits an external cloud to be screened.
The compliance model's weights and rule corpus are signed under ML-DSA-65 (FIPS 204). An examiner can verify that the model and rules in production are the ones that were reviewed.
Honest compliance means the model states its limits instead of overreaching.
The model does not freeze accounts or block settlement by itself. Enforcement actions require an authorized human and are executed through the deterministic engine under key control.
When a rule is ambiguous or a match is weak, the model says so and routes to human review. It is engineered to escalate doubt rather than resolve it with a confident guess.
The model answers compliance questions within the rules it was trained on and declines outside them. It will not improvise a legal opinion for a regime it has not ingested.
The compliance capability is undergoing production hardening and external audit alongside the rest of the platform. Claims about it are stated as design intent until that review closes.
Talk to us about regulated and compliance in a sovereign deployment.