Three decision architecture experiment

Alpaca Options Agent trades defined-risk options spreads — bull put and bear call verticals, iron condors, and a directional debit-spread overlay — on liquid US equities, routed by market regime. A deterministic risk gate runs before any candidate reaches the model: daily-loss circuit breaker, concentration and correlation caps, a credit-to-width floor, a post-stop cooldown. The LLM can decline a gate-approved trade; it can never open one the gate rejected. Rather than build one agent, we ran three in parallel on the exact same imported risk code: this gate-then-decide submission, a fully autonomous zeroclaw agent with no candidate menu at all, and an independent teammate build — compared live on a public dashboard. Every option order goes through Alpaca's official MCP server. Where this account's data tier has no Greeks, IV history, or VIX feed, we compute them in-process and label every one a proxy. The ledger is unedited: profitability is the open question, not the claim.