
Most trading-agent demos put the LLM in charge of finding alpha. We inverted that. The signal is deliberately a toy from public literature (Faber's 10-month SMA on SPY/QQQ/IWM/AAPL/NVDA); all the engineering went into discipline: deterministic risk gates that cannot be talked out of (worst-case-loss cap, −2% daily halt fed by the worse of our own P&L and the account's, structure-keyed idempotent dedup, a direction-flip guard reconciled against the broker's actual book every round), an LLM red-teamer on Alpaca's MCP server with veto-only power, pre-registered exit rules frozen in config before the contest, and a seeded random twin trading a shadow book as a built-in placebo arm. Every decision — including every mistake we caught — is one line in an append-only ledger, and the adversarial reviews that broke the agent before the contest could are committed in docs/ADVERSARIAL-REVIEW.md. Orders go out as atomic multi-leg spreads through the official Alpaca CLI; the MCP side is read-only by client-side allowlist. Account PA32VHBO5AOB ran it unattended from Windows scheduled tasks for the whole window.
4 Sep 2026