Sentinel Quant: A Human-in-the-Loop Multi-Model Portfolio Management System with Formal AI Governance Architecture

Recent advances in Large Language Models (LLMs) have demonstrated capabilities in financial reasoning, yet existing frameworks either operate as passive analytical tools or as fully autonomous agents without structured oversight. We introduce Sentinel Quant, a live-deployed portfolio management system that combines LLM-based market analysis with deterministic risk controls and a novel dual-model governance architecture. The system employs Claude Opus 4.6 as a supervisory reviewer that generates structured recommendations across six portfolio review domains, while Claude Sonnet 4.6 serves as the operational executor that implements only explicitly human-approved actions under comprehensive constraint validation. This separation ensures the advisory model has no execution capability and the execution model has no autonomous decision-making authority. The system manages three independent strategy sleeves (short-term momentum, mid-term trend, long-term ETF accumulation), supports dual-broker execution across Alpaca (US equities) and Trading212 (UK UCITS ETFs), and enforces eight independent risk checks that override all AI outputs. Deployed in continuous production since March 2026, we report preliminary operational metrics and discuss failure modes, safety mechanisms, and design principles derived from live operation. Our work differs fundamentally from recent multi-agent HFT systems such as QuantAgent [1] by addressing portfolio governance and execution safety rather than signal generation speed. Source code is publicly available.

This record contains the Sentinel Quant paper in PDF format. The paper presents a human-in-the-loop portfolio management system combining LLM-based market analysis, deterministic risk controls, structured dual-model governance, regime-aware portfolio operation, and dual-broker execution design. It documents the system architecture, governance workflow, preliminary live deployment observations, and operational failure modes identified during real-world use.