90-second elevator pitch · 3-slide deck

Read it in ninety seconds.

The asymmetry. Criminals run jailbroken self-hosted LLMs (WormGPT, FraudGPT, DIG AI). UK regulators and tier-2 banks cannot. Cloud LLMs are blocked by data-residency. RegTech black boxes cannot be audited. The result: 10,000+ UK correspondent-banking relationships closed since 2020, £29m in 2024 FCA AML fines, 10–40 analyst hours per EDD case.

The architecture. Aletheia (Greek ἀλήθεια: unconcealment) is a domain-tuned, on-prem-deployable LLM stack. Continual fine-tune on a curated investigative-journalism corpus that preserves citations as first-class outputs. First-order-logic-inspired evidence-gated inference controller. Air-gapped, audit-traceable, sovereign-controlled.

The novelty. The first published Compliance Refusal Index — a benchmark of ≈500 investigator-grade queries that quantifies how mainstream LLMs perform on legitimate sanctions and beneficial-ownership questions. Field-defining UK contribution. CC-BY-4.0.

The ask. £140,000 grant (70%) on £200,000 total. Six months. Single applicant, UK-only delivery, TRL 3. Phase 2 demonstrator pathway at £1m, with three to four UK financial institutions and Sovereign AI Unit alignment.

One-page infographic

ALETHEIA ἀλήθεια — Sovereign-Grade AI for National Security & Financial Defence AIR-GAPPED · UK SOIL · AUDIT-TRACEABLE 01 · THE CAPABILITY GAP UK lawful investigators are losing the AI race. PER-CASE BURDEN 10–40 hrs Per enhanced-due-diligence case. >95% false-positive rate on rules. REGULATORY COST £29 M FCA AML fines on UK challenger banks in 2024 alone. DE-RISKING CRISIS 10,000+ UK correspondent-banking relationships closed since 2020. 02 · THE CRIMINAL AI ASYMMETRY CRIMINAL SIDE Self-hosted, alignment-stripped LLMs Industrialise sanctions evasion, beneficial-ownership obfuscation, dark-web vendor profiling. WormGPT FraudGPT DIG AI Source: Resecurity 2024; arXiv 2508.12622 LAWFUL INVESTIGATOR SIDE No equivalent sovereign capability Cloud LLMs (GPT-4o, Claude, Gemini) → blocked: data residency, audit trail. Closed RegTech (ComplyAdvantage, Quantexa) → blocked: no formal audit trail. Academic theorem-provers (Z3, Lean, Coq) → blocked: don't scale to natural language. "Aletheia exists to close this asymmetry — admissibly, audit-grade, on UK soil." 03 · THE ALETHEIA ARCHITECTURE Four-layer stack. AI is the contribution; hardware is the substrate. WP1 · CORPUS LAYER Curated provenance corpus · 5–10 M tokens · CC-BY-4.0 OFAC · OFSI · EU consolidated · FATF typologies · OpenSanctions · ICIJ · Companies House · BAILII — entity-linked, citation-preserving. WP2 · MODEL LAYER Continual fine-tune · Llama-3.1 · Mistral · Qwen-2.5 · LoRA / QLoRA Open-weights base trained on the curated corpus on UK-procurable cloud GPU; reproducibility manifest published with each checkpoint. WP3 · INFERENCE CONTROLLER ★ NOVEL CORE Evidence-gated · first-order-logic-inspired policy layer Citation gate · contradiction detection · missing-evidence detection · role-state confidence threshold · human escalation. MIT licence. WP4 · BENCHMARK LAYER Compliance Refusal Index · ≈500 investigator-grade queries · CC-BY-4.0 First published benchmark. GPT-4o · Claude 3.5 · Gemini 1.5 · Llama-3-70B vs Aletheia. Field-defining UK contribution. 04 · PHASE 1 SUCCESS THRESHOLDS Five capability dimensions. Baselines re-collected within Phase 1. DIMENSION BASELINE (LLAMA-3-70B) ALETHEIA TARGET Sanctions / UBO inference accuracy ~75% ≥85% UBO chain traversal recall (3-hop) ~55% ≥75% Citation precision (F1) ~25% ≥85% Unsupported-claim rate (lower better) ~40% ≤20% Refusal rate on operator-justified queries ~55% ≤5% PHASE 1 HEADLINE HYPOTHESIS ≥50% reduction in unsupported-claim rate vs unconstrained Llama-3-70B baseline, at no worse than 80% recall, on UK-procurable on-prem hardware. 05 · 6-MONTH ROADMAP · 1 AUG 2026 → 31 JAN 2027 M1 AUG M2 SEP M3 OCT M4 NOV M5 DEC M6 JAN WP1 Corpus · 5–10M tokens WP2 Fine-tune · LoRA/QLoRA on UK GPU WP3 Inference Controller · MIT licence WP4 Compliance Refusal Index · CC-BY-4.0 WP5 White paper + Phase 2 plan ▼ M2: Corpus v1.0 ▼ M4: Aletheia weights ▼ M5: CRI v1.0 ▼ M6: White paper · Phase 2 IFS 06 · PROJECT VITAL SIGNS TOTAL ELIGIBLE COSTS £200,000 INNOVATE UK GRANT (70%) £140,000 APPLICANT MATCH (30%) £60,000 DURATION · TRL 6 months · TRL 3

Aletheia at a glance — capability gap, architecture, validation thresholds, roadmap and vital signs. Renders crisp at any zoom (vector). All data points cited in the application body.

3-slide deck (text version)

SLIDE 1 · The capability asymmetry

Criminals have better LLMs than regulators. WormGPT, FraudGPT, DIG AI families industrialise sanctions evasion. UK has no equivalent capability for lawful investigators. The Office of Financial Sanctions Implementation Russia case-load has grown ≈5× since 2022. The FCA imposed £29m in 2024 AML fines on UK challenger banks. Tier-2 UK challenger banks have closed 10,000+ correspondent relationships since 2020. A single EDD case still consumes 10–40 analyst hours.

SLIDE 2 · The architecture (AI is the core, hardware the substrate)

(i) Continual fine-tune on a curated investigative-journalism corpus (OFAC, OFSI, EU consolidated, FATF, OpenSanctions, ICIJ, Companies House, BAILII) — citation-preserving, provenance-tracked, CC-BY-4.0. (ii) Evidence-gated inference controller — a first-order-logic-inspired policy layer with citation, contradiction, missing-evidence and confidence gates. (iii) On-premise reference deployment — air-gapped, UK soil, UK-procurable commodity hardware, audit-traceable. (iv) Compliance Refusal Index — first published benchmark, ≈500 queries, 5 baselines vs Aletheia.

SLIDE 3 · UK exploitation and Phase 2 readiness

Channel 1 — direct on-premise licensing to UK tier-2 challenger banks. Pricing £100k–£500k perpetual + 20% maintenance. Channel 2 — white-label licensing to UK RegTech vendors (ComplyAdvantage, Quantexa, Themis, Featurespace) required to pass the Refusal Index and EU AI Act high-risk-system audit. Channel 3 — UK government procurement via Crown Commercial Service Digital Outcomes 6 to FCA, OFSI, NCA. Conservative ARR £500k–£2m by Y2; £5m+ by Y4. Phase 2 demonstrator at £1m, twelve months, two to three UK financial institutions, Sovereign AI Unit alignment.

One paragraph for the journalist

"Aletheia is the first project to formally close the capability asymmetry between criminal LLMs and regulator LLMs in the United Kingdom — by combining a continual-learning fine-tune on investigative-journalism corpora with a first-order-logic-inspired evidence-gated inference controller, deployed on UK soil, audit-traceable by design, and benchmarked against the world's mainstream LLMs through the first published Compliance Refusal Index. It is not 'AI for compliance.' It is sovereign AI architecture, in service of a national-security priority HMG has explicitly named."

Deck

Aletheia · Sovereign-Grade Adversarial Compliance LLMs

Open on SlideShare.