Mandate

mandate trades options on an Alpaca paper account with no human in the loop. The whole strategy is a written charter, loaded at the start of every scheduled run. The agent reads the market, declares a volatility regime, and either acts inside its rules or stands down with its reasoning on the record. A standing inaction protocol treats "did nothing" as a decision needing approval, so silence is never ambiguous. The safety architecture is the project. The agent never holds an API key. Every call goes through a broker gateway serving only tools a human admitted through a two-key ceremony. The one order-placing operation is a defined-risk shim: it refuses any structure whose maximum loss it cannot compute, any position over a per-position dollar cap, and any order pushing the book past its aggregate cap. The caps are signed into a locked envelope; changing them, the shim, or the charter without the re-key ceremony quarantines every trading grant. Between runs, a deterministic supervisor with no model in it enforces exits stamped into a ledger at entry, under its own principal with no order-placing operation. A circuit breaker demotes the agent's placing authority on a threshold loss; a watchdog kills scheduled runs after a demotion; both principals write hash-chained audit tapes. The controls have already worked in public: on the development account, the breaker fired during a drawdown, the supervisor closed stopped positions while the agent stood demoted, and every refusal landed on the tape. The repo states the record's honest boundary: self-consistent tape, tamper-evident lock, never "tamper-proof." Built on Alpaca's MCP server through safe-agents, an open-source library for exactly this problem. Everything is MIT-licensed and reproducible on a paper account, with a live dashboard rebuilt from the tape after every run.