Autobelay - Long Premium, Short Leash

Judged account, Aug 31 - Sep 3: $100,000 to $105,096, +5.1% over 15 round trips, nothing held past expiry. The Kimi-K3 challenger ran the same code and the same gates on a different model and finished -7.0%, halted by the 2% cutoff. That gap is the result we defend. Autobelay is an autonomous options-trading agent - the brake is deterministic code. Thesis: buy defined-risk, short-dated options premium on the five most liquid names when an open-source model sees a concrete reason; code sizes every trade, stops it, and closes it before expiry. The model never touches an order. Every 10 minutes: (1) deterministic exits run first: stop-loss, take-profit, expiry-day close. (2) A snapshot is built through Alpaca's MCP server: account, positions, and per underlying a 12-contract menu (ATM and ~0.40-delta per side, three expiries) with IV and Greeks. (3) The model may call read-only research tools, then must answer with JSON proposals. Featherless models, one per account: Qwen3.8-Flash-Next judged, Kimi-K3/K2.6 challenging. (4) One risk gate that never negotiates: whitelist, notional cap, position count, expiration window, entry cutoff, daily-loss halt. Only our own code calls Alpaca's order tools. The accounting is the point. Every decision, order, rejection and exit is logged with the model and config behind it. The model's prose is audited against its own inputs: quoted figures against the prior it was shown, exit reasons against the account. The priors it is handed are Brier-scored nightly, so the inputs are graded too. The end-of-day critique comes from a model that did not trade that day: no account grades its own homework. The trading is autonomous; the risk envelope is human: runtime knobs expire at the close, hard caps need a pull request, and nothing at runtime can widen what the bot may lose. It self-hosts; halting an account or swapping its model is a phone control. MIT licensed and original to the event; pre-kickoff infrastructure disclosed in the README.