Connect Your Own Claude to My Forecaster -- Live MCP Demo

Any Claude -- Claude Code, Claude Desktop, or an SDK agent -- can connect to this system's MCP server and get receipt-backed answers. The server does not hand the model a paragraph to paraphrase. It hands back a structured envelope: a status the model must honor verbatim (ok / no_data / not_supported / refused / ambiguous), a verdict, sample sizes, p-values, the exact file the number came from, and the snapshot date. There is no room for the model to round up, borrow a stale number, or invent a dollar edge. When the data is absent, the honest answer is no_data, and the model is instructed to say so rather than fill the gap. Below are three real exchanges. Each one shows a different reason this matters.

Fail-closed answer engines are exactly what AI-engineering teams are trying to build right now: an LLM that can only relay validated facts, cannot hallucinate a number, and admits when it doesn't know. Retrieval-augmented chat usually means "the model paraphrases some documents and hopes." This is the harder version -- every tool returns a typed envelope, the model is contractually bound to honor the status, the numbers carry their own provenance and sample size, and the honesty rails (edge_claimed: false, the caveat ladder, the market-sharper disclosure) are enforced by the server, not the prompt. The demo isn't that the model gives good answers. It's that the model cannot give a dishonest one -- and when the market beats the model, the model is the first to tell you. That fail-closed contract, not any single forecast, is the transferable engineering.