01 · Task-Specialised Multi-LLM Division of Labour
How we use different LLMs for different tasks (the what and why)
Asking a single generalist chatbot or deep-research agent to research and write a multi-jurisdiction regulatory report creates correlated failure modes: the same model that misreads a repealed statute or treats secondary commentary as law also writes the confident summary and grades its own work. Our engine separates discovery, extraction, deterministic verification, and synthesis across distinct model families and non-AI software gates.
Primary-Source Retrieval
Search-grounded and long-context models locate official legislation portals, parliamentary gazettes, central-bank registers, supervisory rulebooks, and treaty series across 170 jurisdictions, capturing candidate provisions with exact section references, URLs, and verbatim quote spans.
Independent Cross-Model Audit
Heterogeneous frontier model families independently structure retrieved provisions against our fixed domain question schemas and audit whether the cited statutory passage actually supports the extracted posture—eliminating single-model hallucination and self-agreement bias.
Deterministic Code Verification
Pure non-AI software enforces schema conformance, HTTP source reachability, verbatim quote-span grounding against retrieved source text, ISO jurisdiction-code isolation, and Tri-Date consistency. Any claim failing a gate is blocked from publication as fact.
Bounded Reader Projection
Reasoning-optimised models generate role-specific views (MLRO, Legal Counsel, Board, CTO, and Co-Pilot answers) strictly over the already-verified record atoms—never from open-web model recall—so prose can never overwrite or invent underlying evidence.
02 · Evidentiary Integrity Enforced in Code
Deterministic code gates, Tri-Date clocks & three-state absence
Evidence integrity is enforced deterministically at build and publish time rather than asserted in prose. Every jurisdiction record across our monitors is governed by three structural invariants.
Tri-Date Separation
A single "Updated" timestamp hides whether the law changed or a crawler merely re-ran. Every record tracks three separate dates:
last_checked— when our pipeline last verified the primary source URL and instrument status;last_published— when the current monitor edition was compiled and deployed;last_changed— when the underlying substantive legal or regulatory rule actually changed.
Three-State Absence Taxonomy
General-purpose LLMs smooth over missing evidence with plausible-sounding interpolation. Our schema requires explicit gap states:
confirmed_absent— primary sources searched; no specific statutory rule exists on point (general law applies);not_yet_assessed— scheduled in our coverage queue and not yet evaluated;stale— previously verified source unreachable or awaiting re-verification after an amendment window.
Jurisdiction Quarantine & Quote Grounding
Automated gates verify that every cited authority belongs to the target jurisdiction's legal order (quarantining cross-jurisdiction bleed, such as metropolitan UK rules misfiled onto a Crown Dependency or Overseas Territory) and that extracted statutory quotes match retrieved source text.
03 · Human Accountability Across Development, Operations & Review
Three-layer human-in-the-loop (HITL) governance
Human oversight is built into the architecture of the engine, continuous weekly operations, and jurisdictional practitioner review—and every record states on its face which level of review it has received.
Human Legal & Systems Engineering
Qualified lawyers and systems engineers design the domain question trees, statutory hierarchies, permitted posture classifications, deterministic rejection gates, and benchmark evaluation rubrics. Language models cannot alter a schema, relax a verification gate, or override a failed provenance check.
Continuous Monitoring & Corrections Ledger
Human operators monitor weekly automated telemetry sweeps (citation reachability, schema conformance, crawler/indexation health), canary benchmark runs, and source-drift alerts. When a source changes or an error is identified, operators log a dated public entry in the Corrections Ledger rather than silently overwriting history.
Named Practitioner Sign-Off vs. Machine-Reviewed
Where a jurisdiction or module carries the Verified badge, a named local legal or regulatory practitioner has reviewed the structured module record, can override wording before publication, and signs with individual and firm attribution (contributor firms are licensed, never paid cash/revenue share to influence findings, and capped at ≤12% of verified items per firm). Records without practitioner sign-off are explicitly badged Machine-reviewed / content: ai_generated.
04 · Empirical Evaluation & Single Source of Truth
Why structured verification beats DIY AI deep research
Running an ad-hoc prompt in a frontier chatbot or deep-research agent produces fluent prose, but fails the evidentiary requirements of regulated compliance, legal advisory, and board reporting. All benchmark figures across our fleet are served from one canonical Single Source of Truth (/benchmark/summary.json).
Live head-to-head and M3 interpretation agreement scores are loaded from the single canonical Benchmark SSOT at a-i.gi/benchmark/summary.json (Public Benchmark).
| Dimension | DIY Frontier AI / Deep Research Prompt | Asymmetric Intelligence Engine & Monitors |
|---|---|---|
| Model Architecture | Single model family generates and self-validates its own prose in one pass. | Task-specialised multi-LLM pipeline with independent cross-model extraction & critique plus non-AI code gates. |
| Primary-Source Grounding | Frequently cites blog posts, law-firm marketing summaries, or dead URLs as statutory authority. | Source-tiered register enforcing primary legislation/regulator instruments (Tier A/1) and verbatim quote-span checks. |
| Temporal Precision | Confuses bill introduction, enactment, and in-force dates; no distinction between check date and change date. | Strict Tri-Date separation (last_checked, last_published, last_changed) on every record. |
| Missing or Unclear Law | Interpolates plausible rules from neighbouring jurisdictions to avoid leaving a blank. | Mandatory three-state absence (confirmed_absent, not_yet_assessed, stale) and cross-jurisdiction quarantine. |
| Reproducibility & Diffs | Non-deterministic: asking the same question twice yields different structure, citations, and conclusions. | Fixed domain schema across 170 jurisdictions with versioned weekly cycle diffs and point-in-time history. |
| Human Accountability | None—user must manually re-verify every citation from scratch before relying on it. | Three-layer HITL: human-engineered gates, weekly operator canary monitoring, and named counsel sign-off on Verified records. |
05 · Evidentiary Grading
Source hierarchy, confidence scale & severity ratings
Primary First, Never Anonymous
- Tier 1 / Tier A (Primary Official): Enacted statutes, official gazettes, central-bank/supervisory rulebooks, court judgments, and binding regulatory registers.
- Tier 2 / Tier B (Official Secondary & Multilateral): FATF/FSB/BIS/IMF/OECD mutual evaluations, consultation papers, and official statistical releases.
- Tier 3 (Corroborated Specialist): Named specialist legal/regulatory reporting, used only with explicit tier labelling and corroborated against primary instruments where available.
Five-Point Confidence Scale
- Verified: Checked and signed off by a named practitioner or analyst against the primary instrument.
- High: Multiple corroborating Tier 1 / Tier 2 official instruments with verbatim quote grounding.
- Medium: Single reliable official source or corroborated Tier 3 reporting where primary text is pending publication.
- Low: Thin or transitional sourcing; explicitly hedged.
- Unknown (
?): Explicit gap state; never filled by conjecture.
Severity & Trajectory
Regulatory and macro signals are rated for operational impact (Critical, High, Medium, Low) and directional trajectory (Escalating, Stable, De-escalating) with explicit effective-date and horizon bands so compliance teams can prioritise action.
06 · Procurement, Data Protection & Auditability
Enterprise governance, privacy & right to challenge
Zero Model Training on Customer Inputs
Subscriber Co-Pilot questions, watchlists, saved views, and generated MLRO/Legal/Board reports are logically isolated per account and organisation. Customer inputs and outputs are never used to train foundation models or shared across customer tenants.
Public-Law Corpus & GDPR
Our research corpus ingests only public statutory, regulatory, and multilateral instruments—no personal-data scraping. Account, billing, and session data are processed by Asymmetric Intelligence Limited (Gibraltar) in accordance with Gibraltar, UK, and EU GDPR standards (Privacy Notice · Terms · Licensing).
Challenge & Corrections Workflow
Every record links directly to its underlying primary sources and exposes a Report an error route (/legal/#contact or [email protected]). Confirmed corrections are logged with date, jurisdiction, prior text, and corrected citation in our public Corrections Ledger.
07 · Per-Product Methodology Directory
Dedicated methodology pages across all monitors
Each commercial intelligence product and the public-interest commons extends the shared engine methodology with domain-specific module schemas, statutory hierarchies, and coverage matrices. Select any monitor below to inspect its dedicated methodology page.
Advennt Methodology
Per-game licensing perimeters, B2C/B2B supplier regimes, advertising & sponsorship restrictions, AML/RG duties, and the 4-artefact publication contract across 170 jurisdictions.
World Payments Monitor Methodology
Fourteen payments-integrity domains (W1a–W13): PSP/EMI licensing, safeguarding, settlement & scheme access, APP fraud, FX controls, and corridor risk across 159+ records (14,769 traced citations).
Sentinel Methodology
AML/CFT statutory architecture, FATF mutual evaluations, sanctions implementation, beneficial ownership registers, and travel-rule supervision across 159 records (14,007 traced citations).
Cryptoassets Monitor Methodology
CASP/VASP licensing, MiCA and third-country equivalence, stablecoin reserve & redemption rules, qualified custody, and DeFi/staking perimeters across 162 records (12,188 traced citations).
Data Protection Monitor Methodology
Supervisory authorities, cross-border transfer mechanisms, breach notification clocks, DPO appointment thresholds, and automated decision-making rules across 169 records (11,013 traced citations).
Global AI Monitor Methodology
Horizontal AI statutes, frontier/GPAI model obligations, high-risk conformity assessments, sectoral AI rules, and compute/export controls across 159 records (6,419 traced citations).
Asymmetric Investor Methodology
Sovereign debt, credit, market-structure indicators, geopolitical regime-shift triggers, and explicit asset-class transmission horizons.
Commons Methodology (Seven Monitors)
Open CC BY 4.0 methodology governing the seven public-interest monitors: Democratic Integrity, Global Macro, FIMI & Cognitive Warfare, European Strategic Autonomy, AI Governance, Environmental Risks, and Conflict Escalation.
Public Benchmark & TRACE Hub
Warts-and-all head-to-head evaluation against Perplexity Sonar Deep Research and Google Gemini Deep Research (ARRF-TOPIC v1), M3 multi-model agreement, and the machine-readable SSOT JSON.