OneSpread: Evidence-Gated AI Options Agent

OneSpread is a solo-built options agent focused on an auditable decision-to-execution boundary. Its public demo is a synthetic replay of approvals, refusals, and order lifecycle states. AI logic: GLM-5, served through Featherless, reviews supplied market and news evidence. It must return structured JSON selecting a precomputed candidate or WAIT and cite supplied evidence IDs. It has no broker tools. A simple 5/20-bar IEX momentum filter supplies directional context; it is an uncalibrated prototype signal, not a demonstrated trading edge. Risk gates: deterministic Python constructs SPY bull-call or bear-put debit spreads with 7–14 days to expiry, $5 strike width, one contract per leg, and a $300 premium cap before costs. Only one active position and one daily entry attempt are allowed. After inference, the engine refreshes quotes, account state, and context. Stale or future-dated data, changed context, excessive price deterioration, old decisions, and calendar blackouts block entry. Alpaca infrastructure: the official Alpaca CLI reads market, news, and account data. alpaca-py validates multi-leg requests; a paper-only HTTP adapter submits them. SQLite records intent before submission, uncertain acknowledgements trigger reconciliation rather than blind retries, and pending orders are canceled after a timeout. Exit handling requires broker confirmation of a flat position. Validation: 22 offline tests, Ruff, and ty pass. GLM-5 passed 10/10 small synthetic integration cases, with 39.16-second median latency. These checks measure integration behavior, not alpha or general model quality. The public walkthrough demonstrates the architecture and synthetic failure cases; an autonomous model-selected broker round trip remains unverified. No profitability or production-readiness claim. Built with Python, uv, SQLite, Alpaca CLI, alpaca-py, Featherless, and GLM-5. MIT-licensed source, slides, and narrated demo are available in the public repository.