MSAR_HMM_Gamma_Trend_Strategy

A two-sleeve options agent trading SPY calls and puts on Alpaca paper. It decides once per session and is allowed to refuse at every step on a closed day, it exits without touching the account. Live and autonomous since 31 August. Sleeve 1 is a regime gate: dealer gamma from Alpaca open interest, a Markov vol filter, a 200-day trend veto. It is long or flat. Sleeve 2 takes the gamma direction and can be short. Max loss is structurally the premium. The differentiator is the audit. Ten thousand block-bootstrapped counterfactuals, a 400-shuffle permutation null, seven public checks. Out-of-sample Sharpe +0.93, deflated Sharpe 0.961, p<0.0025. Reality landed at the 48th percentile of its own distribution — typical, not lucky. Two of the seven checks FAIL, and they are on the dashboard rather than hidden. At 1x against real bid/ask: Sharpe 0.97 versus 0.95 for buy-and-hold, but a -10.0% max drawdown versus -25.4%. The drawdown, not the return, is the product. Most agents are built to find an edge. This one was built to destroy its own, and to publish what survived.