Esscher

Esscher asks a simple question after an earnings release: once the market’s first reaction has passed, is there still a defensible trade—or should the agent stay in cash? Its loop is observe → decide → constrain → act → verify. Esscher freezes only information available at the decision cutoff, rejects late or incomplete evidence, and gives a bounded AI reasoner a structured view of the event. The agent can abstain. If it acts, it compares cash, shares, a long option, and a defined-risk debit spread as four expressions of the same directional view. The AI never has unrestricted order authority. A deterministic governor checks evidence freshness, account state, exposure, lifecycle rules, and duplicate-order protection before any Alpaca MCP paper action is eligible. Every stage emits a hash-linked receipt, so the path from source evidence to decision, permit, order intent, and reconciliation is inspectable and replayable. Current boundary: the public repository demonstrates deterministic point-in-time evaluation, synthetic fixtures, and fake-broker lifecycle tests. It does not yet claim a real Alpaca paper fill, live P&L, or proven alpha.