Argus

1. 1
   Picker proposes

   Fixed rules, calibrated from the ten-year backtest, propose one trade a day. No model, no discretion.

2. 2
   Guard 15 hard limits

   15 hard safety limits check the proposal. Any single failure blocks the trade, and none can be overridden downstream.

3. 3
   Reviewer veto / shrink only ? Verified live end-to-end: a real MCP account fetch plus a real Gemini call returned a genuine APPROVE decision.

   An AI (Gemini) may veto or shrink a Guard-approved proposal, never raise its size, never originate one. Enforced in code, not by the prompt.

4. 4
   Order to broker

   Sent to the broker as a single two-legged order once both Guard and Reviewer have cleared it.

5. 5
   Monitor every 15 min

   Every 15 minutes: a profit target, a forced close the day before expiry (assignment-risk protection), or a hard drawdown stop.

6. ↺
   Recovery runs alongside Order & Monitor

   An unrecognized broker position or a partial fill halts trading and closes the excess risk immediately.

The AI is deliberately the least-trusted component here: it sits after every hard limit has already passed, and the only actions it's structurally capable of are subtracting risk or refusing. Enforced by apply_reviewer_decision, not by asking the prompt nicely.

Full design rationale: OPTIONS_SYSTEM_PLAN.md. Full experiment log, including two retracted findings: EXPERIMENT.md.