Why an ensemble: re-asking one model gives near-identical answers, but different model families disagree in genuinely useful ways — cross-model agreement carries measurable predictive value, and their spread doubles as a danger gauge. Why price-blind: an analyst who can see the market price anchors to it; the price enters only later, in deterministic code — as the Reconciler's alarm and the 50/50 blend that turns extreme disagreement with the crowd into a small position rather than a bet-the-house moment. The Trading Desk is a human — nothing trades automatically. The proving ground: before any configuration touches real dollars it replays 12 months of history in a walk-forward simulation with an honest bankroll — finite cash, positions lock until resolution, and every model re-evaluation must be justified by a trigger and paid for. What makes it agentic: the analysts and the reconciler are LLMs in a multi-turn tool-use loop — each writes its own search queries, chooses which articles to read in full, and iterates until it submits a schema-validated verdict; nothing hands them a pre-built context blob. The same market-data layer is also exposed as a standard MCP server (Model Context Protocol; mates_mcp — five self-describing, read-only tools: resolution rules, prices, order book, search, market context) that any MCP client can connect to; it deliberately wraps no trading or account calls, so no agent can move money through it. Inside the simulator the identical tool loop runs in-process for speed and leak-tight determinism. Python + Flask, containers on GCP.