
MarginCalled does not predict the market. It looks for prices that cannot all be correct at once. Call prices must satisfy total positivity of order 2: for two expiries and two strikes, one product of prices must exceed another once strikes are forward-adjusted. Four contracts form a "rectangle". When the inequality fails on crossable quotes, the surface is internally inconsistent — not mispriced against a model, but provably inconsistent. The agent trades that and exits when it corrects. Every candidate passes an early-exercise theory gate before it can trade. TP2 is a theorem about European calls, so an American one only qualifies if early exercise is worthless — no dividend in the window, or a certificate from the source study. Crucially, an empty dividend list is not evidence of no dividend: anything past the provable horizon is marked unresolved and left alone. Survivors pass 16 deterministic risk gates, all evaluated with no early exit, including a re-test of the violation on fresh quotes immediately before the order goes out. Orders go over Alpaca's MCP server as multi-leg limits priced better than the indicative quote, so a bad quote produces a non-fill rather than a bad fill. Results: 192M rectangles scanned, 17,097 violations found, 43,566 episodes tracked to resolution. 98% reverted, median 11 minutes. Live, 21 round-trips — every one closed because its violation reverted, none on a timer. The account still lost $37, and that is the finding. The mispricing is worth ~2¢ and the bid-ask spread costs ~2¢, so execution decides the sign. An end-of-day study measuring mid-prices cannot discover this, because mid-prices are not tradeable. Every refusal is logged — 32,505 decisions with full evidence — and a language layer reads that log and explains them. It has no tools and cannot place an order
4 Sep 2026