BlackStar

Ask a language model what a stock will do and it will tell you — confidently, in full sentences, with a number. That number is often invented, and nothing about the way it is delivered tells you which numbers are real. That is the actual problem with an AI trading agent: not that it is wrong sometimes, but that it is equally fluent when it is wrong. BlackStar sells defined-risk vertical credit spreads on the Alpaca API. Its architecture exists to make hallucination structurally impossible. A language model is given exactly one job: naming what plausibly drives a stock — copper for an industrials ETF, rates for a bond fund. A naming task, which models are good at. Every one of those claims is then measured against real price history before it is allowed to vote, using bias-corrected estimators: Hurst exponents for trend persistence, Ornstein-Uhlenbeck half-life for mean reversion, Shannon entropy, and effective transfer entropy for genuine information flow. The corrections matter — raw transfer entropy scores 0.012 bits on pure random noise, so uncorrected, every driver looks informative. Corrected, noise scores zero. Fewer than a fifth of proposed drivers survive; the rest abstain. What survives is pooled by inverse-variance weighting, so the most precise agent is heard loudest rather than the most confident, then passed through three hard risk gates. Over 82 consecutive live passes it opened zero positions, because the premium on offer never covered the risk — and every refusal is logged with the number that caused it: credit/risk 18% below the 29% needed at 81% POP. Backtested out of sample it returns 4.11% over 344 trading days at 0.90% max drawdown, deflated Sharpe 0.74 across 26 counted trials. It does not beat buy-and-hold, and the dashboard says so itself. 459 tests.