Researching
Consensus Architecture in Agentic Banking
Why single-agent AI fails regulatory audits, and how to build self-policing systems using shadow prompts.
GovernanceAgentic AICompliance
✍️ Active Research: Compiling patterns for mitigating stochastic risks in heavily regulated environments.
The Regulatory Reality
In a creative workflow, an AI hallucination is a funny quirk. In a retail bank, an AI hallucinating an approved mortgage application is a multi-million-dollar regulatory breach.
Because Large Language Models (LLMs) are inherently non-deterministic, we cannot guarantee their safety through prompt engineering alone. Financial regulators (like APRA and the SEC) demand strict explainability that a standalone LLM simply cannot provide.
The Architectural Solution
We must surround the AI with a rigid, deterministic fortress. This upcoming paper explores:
- Consensus Architecture: Why an "Execution Agent" must submit its proposed actions to an independent "Evaluator Agent" before acting.
- Shadow Prompts: Equipping the Evaluator Agent with strict regulatory rules to police the Executor.
- The Deterministic Gate: Routing all AI-approved payloads through hard-coded Zod schemas and traditional business logic before hitting the ledger.
- Immutable Audit Trails: Generating cryptographically secure logs to satisfy the auditor's ultimate question: "Why did the AI do this?"