SOVEX
CBDC Data Centers Sovereign AI Tokenization Deep Tech Architecture About Team Request access
Sovereign AI / Specialized models / Regulated and compliance

Regulated and compliance.

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.

Trained on the actual rules, with the citation kept intact

The model's authority comes from primary sources, versioned and traceable to their origin.

01

Primary-source training

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.

02

Sanctions and watchlists

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.

03

Jurisdictional layering

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.

04

Version-pinned rules

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.

Compliance decisions arrive with the rule attached, not as a verdict

The model's job is to show its work so a human officer can accept or overturn it.

01

Cited determinations

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.

02

AML pattern reasoning

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.

03

Explainable false positives

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.

04

Human-in-the-loop by design

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.

Every compliance action leaves a record an auditor can walk

The design assumes the output will be examined by a regulator, and structures itself accordingly.

01

Immutable decision trail

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.

02

Reproducible outputs

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.

03

Data residency

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.

04

Signed model provenance

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.

What the model refuses to do is part of the design

Honest compliance means the model states its limits instead of overreaching.

01

No silent auto-enforcement

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.

02

Uncertainty is surfaced

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.

03

Scope discipline

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.

04

Under external audit

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.

Build it sovereign.

Talk to us about regulated and compliance in a sovereign deployment.