Berkshire Alpha

Berkshire Alpha is an autonomous options-trading agent whose models argue, and whose risk layer cannot be argued with. A deterministic quant layer picks the regime first — IV/RV, variance risk premium, 25-delta skew, VWAP deviation and volume-weighted momentum across a fixed 50-name universe ordered by measured 3–7 DTE options-chain liquidity, not market cap. No language model votes on it. Only then does the LLM layer run, narrowing as it goes: quant and news analysts, a Bull and a Bear drawn from deliberately different model families debating under a Disagree-or-Commit protocol where agreement is only valid if it cites newly introduced evidence, a trader that proposes a concrete structure, and three risk personas. Every output is Pydantic-validated with a one-retry-then-drop policy, and every strike is re-checked against the live chain in code. Then all of it is handed to a deterministic Python gate that no model can reach. Sizing is quarter-Kelly against a hard 2%-per-trade cap and can only ever reduce, never license. The number of trades a language model can authorise in this system is zero, and three import-graph tests fail the build if anyone ever wires one up. This is not a promise: on 1 September the gate refused an NVDA call spread that all three risk personas had unanimously approved, because Kelly f* came out at −0.00448 against a threshold of zero. That is decision 36, still queryable in the public read-only API. Then it did the hardest thing a trading system can do: nothing. An execution-layer defect put on a position the sizing model never sanctioned, portfolio delta reached 222% of its limit, and the gate froze new entries. Over the next two sessions the agent screened 350 candidates, blocked 46 on REDUCE_ONLY and originated exactly zero trades — spending $0.00 on inference, because the gate short-circuits the pipeline before the first token. Our own Reflector agent examined both sessions independently and voted HOLD.