AlphaGate — Trading Agents

Most LLM trading agents put the model in the decision seat: prompt in, order out. In options that fails for a specific reason — the loss function is asymmetric, and a language model has no calibrated sense of how much it can lose. A hallucinated ticker costs you a bad fill; a hallucinated naked short strangle costs you the account. AlphaGate inverts the arrangement. The LLM proposes structure — which spread, which expiry, which delta. Deterministic code disposes: every proposal must pass a Risk Gate that owns defined-risk enforcement, position and greeks budgets, and the kill switch. The Gate is pure, tested, and contains no model. Execution accepts only a gated order, and only the Gate can construct one — no bypass, no force flag, no debug switch. Naked short options are unrepresentable at the type level. Two agents share one paper account. The options sleeve trades a put credit spread rule (0.16 delta short, 0.08 delta wing, ~14 DTE) that an LLM-driven research lab in the same repository searched out and pre-registered against a sealed two-year window it had never read. The equity sleeve executes one strategy out of 414 hypotheses, put through walk-forward and its own sealed window. The two projects import nothing from each other; the seam is a JSON book carrying the strategy's fingerprint, and the account refuses any book whose fingerprint is not pinned. Every cycle — including every refusal — is written to an append-only journal with the inputs seen, the proposal made and the Gate's verdict, and rendered on a live dashboard. We claim only what the evidence supports: the option rule survived a pre-registered attempt to refute it, it was not confirmed, and the code stores those as two separate flags so no dashboard or README can flatten the distinction.