EdgeStack: Evidence-Gated Trading Agent

**The problem.** LLM trading agents lose money confidently, because their rules come from vibes. Ask a model for a strategy and it gives you one — plausible, articulate, untested. We make the agent answer the harder question: what evidence does a trading rule need before it deserves to exist? **The answer.** EdgeStack was built with healthy skepticism: months of compute spent trying to disprove candidate edges, and the agent only trades what survived — 33 years of data, surrogate-null tests, drop-one-era checks, and two disjoint validation windows on an independently built engine where every parameter tuning failed validation and was rejected. The graveyard (Elliott waves, Fibonacci levels, five macro overlays, our own first options design) is published with the number that killed each idea. **What survived.** Long SPY overnight-only (Sharpe 0.89 vs 0.05 intraday), gated by the 12-month trend and a credit canary (adopted only after passing both validation windows, 0.80→0.98 / 0.65→1.02 Sharpe); a capitulation basket (+1.42%/event, t=4.27, 136 events/33y); and defined-risk options spreads on the same signal behind 14 deterministic gates. The unifying, thrice-measured finding: markets revert emotional moves and honor informational ones. **The machine.** Orders and account reads route through Alpaca's MCP Server v2 (REST fallback, every route journaled). The production engine reproduces the research record to three decimals; an append-only decision journal explains every trade — and every refusal — and feeds the public live dashboard. **Honest limits, on purpose.** These are risk-adjusted edges; a 5-session P&L window is mostly noise for every entrant. What EdgeStack demonstrates is the discipline that survives any market: rules that earned their existence, an engine that provably implements them, and an agent that can explain why it took a trade and why it did nothing.