Abitda

Most AI trading submissions build fragile bots that fire unhedged options orders and blow up the moment volatility spikes. ABITDA was built on a different premise: the options market doesn’t need another toy bot; it needs an institutional evaluation harness—a "SWE-bench for options finance"—that proves an autonomous agent can survive a crisis before it is ever granted broker execution privileges. ABITDA provides a standardized evaluation protocol (`OptionsAgentProtocol`) and a Python package (`pip install abitda`) that stress-tests candidate LLMs against deterministic bar-by-bar replays of 5 historical crises: the Aug 5 2024 Yen Carry Crash (+181% VIX surge), 2018 Volmageddon, the SVB bank run, 1987 Black Monday, and a calm-bull baseline. Each simulation outputs an institutional Fiduciary Scorecard graded A+ to F. On the Aug 2024 crash, our Floor Committee preserved 99.44% of capital (max drawdown -1.23%, zero Greek breaches), while an un-harnessed momentum bot suffered a -42.8% margin blowout. Before any agent action reaches the market, three code-enforced mathematical firewalls sit between model decisions and broker execution: 1. Greeks Gatekeeper: Hard analytical Black-Scholes limits capping aggregate book Delta at ±0.25 and Vega at $150. 2. Regime-Flip Emergency Exit: Liquidates short volatility exposure immediately upon macro shock detection (>12% VIX spike). 3. Statistical Edge Guardian: A rolling binomial Z-score monitor that autonomously self-suspends broker access if win-rate decays. An integrated FastMCP server exposes six quantitative tools to external agents (Claude, Cursor, Gemini CLI), while a live React/FastAPI cockpit operates at https://abitda.up.railway.app. Real execution is independently verified on Alpaca Paper Account PA382FDPI5IO with Options Level 3 clearance. The options market doesn't need more bots. It needs an institutional harness deciding which bots are safe to trade. That is ABITDA.