Alpha Council

Alpha Council is an autonomous, options-native paper-trading desk with a governance chain modeled on a real trading floor. Each scan, a deterministic quant funnel narrows ~150 discovered symbols — including names injected the same morning from material news and SEC filings — to a handful of finalists on EVENT and MOMENTUM tracks. Each finalist gets a council: Bull, Bear, and Catalyst analysts, a scenario generator, a Python payoff engine pricing five real debit spreads under every scenario, and a GPT portfolio manager that proposes or abstains with underlying-price invalidation rules. A Claude Red Team attacks the proposal (PASS/MODIFY/VETO), one revision may shrink or switch structure but never add risk, and a deterministic Risk Constitution has the final, un-overridable word before any order reaches Alpaca. What makes it different is the measurement loop. Every decision is stored three ways — what GPT proposed, what Claude changed, what executed — so the P&L contribution of each governance layer decomposes into selection and sizing effects and is a queryable number, not a claim. Every refused candidate gets a shadow record, so the risk gates are audited too: what did our conservatism save, and what did it cost? After each close, an evolution agent writes lessons with evidence for and against and may propose a bounded Challenger strategy that must beat the Champion in shadow before a human can promote it; the code raises without that approval. Two live sessions, real numbers: 55 councils, 24 proposals, 19 Red Team reviews (18 MODIFY, 1 VETO), 11 Constitution approvals, four council-approved trades, and both evolution cycles declining to overfit a small sample — all for $7.74 of the $100 GenAI budget. Built on Alpaca's Trading API, IEX and Indicative market data, and the Alpaca MCP server as the control plane, with OpenAI and Anthropic models. Paper trading only, enforced in code.