CrossSignal: Auditable Cross-Market Options AI

Most autonomous trading agents optimize for producing a signal. CrossSignal is designed to produce a defensible decision. The agent synchronizes six market lenses: equity implied volatility, Treasury yields, credit, realized volatility, rate expectations, and options positioning. A deterministic engine identifies the strongest cross-market disagreement. Claude then creates a structured thesis, while a separate adversarial review identifies counterarguments, missing evidence, an alternative explanation, and a confidence penalty. CrossSignal converts approved signals into defined-risk SPY, HYG, and TLT option spreads. Before an order can reach Alpaca, deterministic controls verify confidence, data integrity, maximum loss, buying power, diversification, Greeks coverage, portfolio stress, liquidity, bid-ask quality, and drawdown. Claude cannot contact the broker or override a failed gate. Every cycle creates a SHA-256 Decision Contract containing the evidence, prediction, invalidation condition, risk result, and proposed portfolio before the outcome is known. The agent later scores the sealed forecast against inverse and cash counterfactuals. The project uses Alpaca's official MCP server for market data, account state, option snapshots, Greeks, paper orders, positions, and reconciliation. It also runs unattended Evidence Watch jobs in GitHub Actions with broker mutations disabled. The fresh $100,000 competition paper account currently has no orders or positions, so CrossSignal makes no P&L claim. Its latest cycles abstained because adversarial confidence fell below the fixed threshold and after-hours option quotes failed liquidity and bid-ask requirements. The agent did not weaken its controls merely to manufacture a competition trade.