QUORUM — Autonomous AI Trading Desk

QUORUM is an autonomous AI trading desk built for Alpaca paper trading. Its core design separates reasoning from authority: language-model agents analyze structured market evidence and propose a target state, while deterministic code controls risk, position sizing, execution, protective orders, reconciliation, and the final decision path. The system uses an Analyst and Strategist reasoning layer, a deterministic Scout for market evidence, a Risk Officer with explicit veto rules, and a Decision Engine that converts the approved proposal into an executable intent. The model controls only the desired target state; it does not control price, position size, stops, risk limits, or broker execution. QUORUM also maintains an append-only decision ledger, durable client order IDs, continuous broker reconciliation, protective-order handling, circuit-breaker and kill-switch controls, and idempotent fill capture. Reasoning is cached only for the same symbol, timeframe, and bar, while failures are never cached. During development, QUORUM encountered a semantic routing bug where a risk veto was incorrectly interpreted as an instruction to exit an existing position. The incident produced a $3,052.16 realized loss in the paper account. The team reconstructed the event to within $0.13 of the modeled loss and added 16 regression tests targeting the failure mode. The incident is preserved as auditable evidence rather than hidden. The project also includes an options feasibility layer and read-only Alpaca options discovery/preflight. Full options execution is intentionally not claimed as implemented in this submission. QUORUM is paper-trading only. Its principle is simple: the model reasons, deterministic code decides, and reality is reconciled.