Sovereign-grade adversarial-symmetric AI for sanctions, AML and OSINT.
Aletheia (Greek ἀλήθεια: unconcealment) is a domain-tuned, on-prem-deployable, audit-traceable LLM stack for high-stakes financial-crime decision support — a six-month feasibility study by London-based fintech and RegTech founder Vladislav Solodkiy, vetted by Tech Nation, addressing the United Kingdom's compliance-AI capability asymmetry head-on.
The asymmetry
Criminals have better LLMs than regulators.
The WormGPT, FraudGPT and DIG AI families industrialise sanctions evasion, beneficial-ownership obfuscation and dark-web vendor profiling. Tier-2 UK challenger banks have closed more than 10,000 correspondent-banking relationships since 2020. The FCA imposed £29 million in AML fines on UK challenger banks in 2024 alone. A single enhanced-due-diligence case still consumes 10–40 analyst hours. Cloud LLMs cannot be deployed in regulated UK banks. Closed-source RegTech platforms cannot be formally audited. Academic theorem-provers do not scale.
The architecture
(i) Continual fine-tune on investigative-journalism corpora
OFAC, OFSI, EU consolidated, FATF typologies, OpenSanctions, ICIJ, Companies House, BAILII — curated with provenance metadata, entity-linked, citation-preserving as a first-class output.
(ii) Evidence-gated inference controller
A first-order-logic-inspired policy layer. Citation gate, contradiction detection, missing-evidence detection, role-state-aware confidence threshold, escalation to human review when thresholds fail.
(iii) Compliance Refusal Index
The first published benchmark quantifying how mainstream LLMs perform on legitimate sanctions, beneficial-ownership and typology questions. ≈500 investigator-grade queries. Field-defining UK contribution.
(iv) On-premise reference deployment
Air-gapped. Operator-authorised. UK soil, UK-procurable commodity hardware. Audit-traceable. Sovereign-controlled — not jailbroken, not refusing.
Phase 1 deliverables · 1 August 2026 → 31 January 2027
| WP1 | WP2 | WP3 | WP4 | WP5 |
|---|---|---|---|---|
| Curated corpus (CC-BY-4.0) | Fine-tuned weights (Llama-3.1, Mistral, Qwen) | Inference controller (MIT) | Compliance Refusal Index (CC-BY-4.0) | White paper + Phase 2 plan |