Aletheia: verifiable AI for the UK's sovereign sanctions regime.
Problem
UK financial-crime defence is constrained by a measurable capability asymmetry between criminal actors and lawful investigators. Criminal actors operate jailbroken or self-hosted large language models — the WormGPT, FraudGPT and DIG AI families documented in Resecurity 2024 reporting and arXiv preprint 2508.12622 — to industrialise sanctions evasion, beneficial-ownership obfuscation and dark-web vendor profiling. Lawful investigators have no equivalent capability.
Cloud-hosted general-purpose LLMs cannot be deployed in regulated UK banks because data-residency, audit-trail and deterministic-inference rules forbid it. Closed-source RegTech platforms produce risk scores that cannot be formally audited. Academic theorem-provers do not scale to the noisy, contradictory natural-language reality of compliance evidence. The HMG Economic Crime Plan 2023–2026, the FCA's £29 million challenger-bank AML enforcement actions of 2024, and the roughly five-fold growth in OFSI Russia-sanctions case-load since 2022 frame this as a national-security priority.
Why current solutions fail
Three failure modes, three layers of inadequacy.
First, retrieval-augmented generation provides citation grounding but does not enforce it as a hard precondition for output — models can and do produce conclusions ungrounded in retrieved evidence. Second, constitutional AI constrains style, not chain-of-evidence — alignment refusals do not protect investigators from hallucinated UBO chains. Third, neuro-symbolic reasoning has demonstrated the principle in academic settings but has not been published at compliance-domain scale on UK-procurable on-premise hardware. No combination of these approaches has been benchmarked against operational sanctions and beneficial-ownership queries.
Approach
Aletheia's technical contribution is a novel system architecture, not a new base model. Two architectural novelties combine. First, continual-learning fine-tuning of an open-weights base on a curated investigative-journalism and regulatory corpus that preserves chain-of-evidence citations as first-class structured outputs. Second, an evidence-gated inference controller — a first-order-logic-inspired policy layer that validates citations, detects contradictions, surfaces missing-evidence gaps and escalates to human review when confidence thresholds fail. Both novelties live in the model and control layers, not the deployment substrate.
Outcomes
Phase 1 (1 August 2026 – 31 January 2027) produces five outputs, all under permissive open licences: a curated public-source corpus with provenance metadata; an MIT-licensed reference inference controller; the first published Compliance Refusal Index benchmark across mainstream and self-hosted LLMs; an on-premise reference deployment design; and a technical white paper with Phase 2 demonstrator roadmap.
Phase 2 deploys the demonstrator at two to three UK financial institutions with FCA, OFSI and NCA engagement on procurement-grade evaluation. Phase 3 commercialises through three named channels: direct on-premise licensing to UK tier-2 banks; white-label licensing to UK RegTech vendors required to meet the EU AI Act high-risk-system audit; and UK government procurement via Crown Commercial Service Digital Outcomes 6.
UK alignment
The work is performed entirely in the UK by UK-resident PAYE staff, with all exploitation routed through UK tier-2 banks, UK regulators and UK RegTech vendors. It advances the HMG Economic Crime Plan 2023–2026 (Cabinet Office, HM Treasury), the AI Opportunities Action Plan 2025 (DSIT) and the £500 million Sovereign AI Unit mandate, the National Security and Investment Act 2021 Area 1 (Artificial Intelligence) capability priority, the Economic Crime and Corporate Transparency Act 2023, and the FCA Business Plan 2024–2025 on AML enforcement.