ATLAS: Autonomous Trading, Limited-risk Alpha

ATLAS sells 7 to 14 day vertical credit spreads on twenty liquid names, on a dedicated Alpaca paper account, one cycle every fifteen minutes: measure, filter, argue, order, journal. The universe is fixed, not scanned. A scanner is where an earlier desk of mine spent ten sessions logging "no candidates" while the screener was never dry. Alpaca returns greeks and implied volatility as null on this plan, so ATLAS solves implied vol by bisection from the quote mid and takes delta from that, behind a quote-quality gate that runs first. Every rail exists because a real desk of mine failed that way. A spread goes in as one multi-leg order and comes out as one, and after every terminal state, including a submit that raised an exception after the broker had already accepted the order, positions are reconciled and any orphan leg is closed at once. That rail exists because a long leg once filled while its short leg was rejected, leaving a naked option nobody intended. Two collateral ledgers, because counting spread legs against a cash-secured cap once froze a desk shut for ten sessions. No naked short puts, ever. Two models then argue every candidate the gates allowed, and a referee written in code decides. They can veto a trade. They cannot cause one, size one, or price one, and the parser keeps three fields from a reply, so a model demanding five hundred contracts of something else has nowhere to land. It finished down 2.28 percent, and the system can say why. Every closed loser was a call spread and every winner a put spread, because the side rule sells calls on names below their twenty-day average, which left the desk short calls on the weakest names when the tape snapped back. That is a strategy wrong about direction for three days, not a system breaking. The dashboard prints the result with the best trade removed, a metric built to undercut its own headline. Five days is not a track record, and the page says so first.