# Methodology, Multi-LLM Architecture & Human Governance — Asymmetric Intelligence (`a-i.gi/methodology`)

Canonical URL: https://a-i.gi/methodology/
Benchmark SSOT: https://a-i.gi/benchmark/summary.json
Public Benchmark Hub: https://a-i.gi/benchmark/

Asymmetric Intelligence powers eight commercial intelligence products and seven public-interest monitors from one source-disciplined research engine.

## 01. Task-Specialised Multi-LLM Division of Labour (The What and Why)

1. **Stage 1 — Primary-Source Discovery & 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.
2. **Stage 2 — Independent Multi-Model Extraction & Adversarial Cross-Check:** 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.
3. **Stage 3 — Deterministic Non-AI Code Verification (Zero-LLM Gate):** Pure non-AI software enforces schema conformance, HTTP source reachability, verbatim quote-span grounding, ISO jurisdiction-code isolation, and Tri-Date consistency (`last_checked`, `last_published`, `last_changed`). Any claim failing a gate is routed into an explicit gap state (`confirmed_absent`, `not_yet_assessed`, `stale`).
4. **Stage 4 — 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.

*IP Boundary:* We publish our verification stages, evidentiary hierarchy, deterministic gate rules, three-state gap taxonomy, corrections ledger, and benchmark scores in full. Exact model-to-task routing tables, ensemble weights, prompt decomposition chains, reconciliation heuristics, and calibration harnesses remain proprietary Asymmetric Intelligence IP.

## 02. Three-Layer Human-in-the-Loop (HITL) Governance

- **Layer 1 — Human Legal & Systems Engineering (Development & Gates):** Qualified lawyers and systems engineers design the domain question trees, statutory hierarchies, deterministic rejection gates, and evaluation rubrics.
- **Layer 2 — Weekly Operational Monitoring, Canary Sweeps & Corrections Ledger:** Human operators review weekly automated telemetry sweeps, canary benchmark runs, and source-drift alerts, and adjudicate dated entries in the public Corrections Ledger.
- **Layer 3 — Named Practitioner Sign-Off (`Verified` vs. `Machine-Reviewed`):** Records carrying the `Verified` badge are reviewed and signed by named local legal or regulatory practitioners (contributor firms are licensed, never paid cash/revenue share, and capped at <=12% of verified items per firm). Machine-reviewed records are explicitly labelled `content: ai_generated`.

## 03. Per-Product & Commons Methodology Directory

- **Advennt (Gambling Regulatory Intelligence):** https://advennt.io/about/methodology/
- **World Payments Monitor (`payments.gi`):** https://payments.gi/methodology/
- **Sentinel — Financial Integrity Monitor (`sentinel.gi`):** https://sentinel.gi/methodology/
- **Cryptoassets Monitor (`cryptoassets.gi`):** https://cryptoassets.gi/methodology/
- **Data Protection Monitor (`dataprotection.gi`):** https://dataprotection.gi/methodology/
- **Global AI Monitor (`artificial-intelligence.gi`):** https://artificial-intelligence.gi/methodology/
- **Asymmetric Investor (`asymmetric-investor.gi`):** https://asymmetric-investor.gi/methodology/
- **Asymmetric Intelligence Public Commons (`asym-intel.info`):** https://asym-intel.info/about/ (https://commons.a-i.gi/about.html)
- **Public Benchmark & TRACE Provenance Hub:** https://a-i.gi/benchmark/