🤓 Latest Submissions
Argus
Argus is an autonomous trading agent that sells defined-risk SPY put credit spreads for weekly income. It exists to answer a question most trading bots never ask themselves: not "does this look good," but "would this survive an honest attempt to prove it wrong?" Each session, Argus proposes one spread, sized to a strict percent of equity. It passes through fifteen deterministic risk guards covering account state, liquidity, concentration, and a VIX-term-structure crash filter, all fail-closed on missing data. An LLM then reviews it but can only approve, shrink, or veto; a network-free function enforces that ceiling, so safety never depends on the model. Orders are sent as a single multi-leg limit order via direct Alpaca calls. A 15-minute monitor checks four exit conditions in priority order (orphaned leg, day-before-expiry, profit target, hard drawdown), and a recovery module reconciles the account against the broker's own state every cycle. It gives out a public read-only live dashboard of the real paper account (position, equity curve, decision log, no exposed controls), a local admin console for the operator, and an audit log where every decision, including every rejection, is recorded and reproducible offline. What makes it creative is the falsification layer. Before publishing a headline number, Argus checks whether that number is luck or not. A Deflated Sharpe Ratio weighs the result against all 31 ideas this project tried on the same data, since trying enough things eventually produces a good-looking number by chance. A Manipulation-Proof Performance Measure checks it a second way, chosen because ordinary Sharpe ratios can be artificially inflated by strategies shaped like this one: steady small gains, rare large losses. That same check once caught the project's own benchmark bug, one that had made the strategy look worse than it really was. The verdict is: a 0.56 Sharpe ratio, which cannot be statistically distinguished from the best of 31 lucky guesses.
4 Sep 2026
