Six rules behind Argus, a quant trading system run by GPT-6 Astra
The author says Astra generates a new quant strategy every hour and only three in a thousand survive, and lists six rules covering data, validation and risk control.
GPT-6 Astra writes a new quant trading strategy every hour, 24/7. 997 out of 1,000 get KILLED before they see a dollar. This is the manual for the 3 that survive. The system's called Argus, and every rule in it exists because an earlier version broke it first. 1. Operations owns the data and nothing else. Every number it reads about the past is POINT-IN-TIME - the print you'd have actually seen that day, never a later revision. Most bots cheat here without knowing it. 2. Researchers run as a swarm overnight, each assigned one class of mispricing - mean reversion, cross-sectional factors, the vol surface, event drift. REAL code, not takes. 3. The Validator is the whole game. It NEVER sees the reasoning, only the sealed strategy run against data it hasn't touched. Sharpe above 1.5. Drawdown under 15%. Hit rate above 55%. t-stat above 2.0. Then the two checks almost nobody runs - Deflated Sharpe and Probability of Backtest Overfitting. Most strategies die right here. 4. The PM sizes anything left at HALF-Kelly, under a volatility target, with a hard cap on correlation between positions. Full Kelly is how you go broke while being right. 5. Risk has ZERO negotiation. A 4% daily loss flattens the entire book. No override. 6. Execution works orders slowly instead of paying the spread. Paper only, until it's EARNED real money. One rule sits under all six: no agent grades its own work, and no kill switch waits on a bot's opinion of itself. Building the model was never the hard part. Building six different ways to say no to it was. Full paper's below.
This post quotes another post: “Running a nine-agent hedge fund with GPT-6 Astra”