Language models propose. Deterministic, unit-tested code disposes. Every decision it makes, and every one it refuses, is logged and replayable. It earns autonomy by proving itself first. Paper-only for now, and it claims no returns.
A real decision artifact, including the ones it refuses.
Returns are easy to claim and hard to prove. The architecture is the part you can actually verify, so we built these five things into it from the start.
Language models propose. Deterministic, unit-tested code disposes. Every capital decision passes hard gates, and no model reaches the broker directly.
Claude, OpenAI codex, Gemini and GPT cross-examine each call before it stands. Adversarial by design, a hedge against any single model's overconfidence.
Every decision, and every refusal, is hashed and trace-linked. You can rebuild it from the exact point-in-time data the agent saw.
Margin, leveraged ETFs and defined-risk options each sit behind deterministic caps. Auto-deleverage and broker-side resting stops are built to hold through a full outage.
It runs on point-in-time data with no look-ahead and grades itself on drawdown. The point is to prove the machine works before claiming any edge, and to say which is which.
The goal is an agent you trust to trade on its own. It gets there on evidence: each level unlocks only after the audit trail proves the one before it. Risk-reduction runs automatically from day one, and the deterministic rails hold at every level.
Anyone can show a winner. The rarer thing is a system whose limits live in code, not in a pitch. These are the rails it runs on.
One boundary, place_order, separates judgment from authority. No model output reaches the broker without passing deterministic, tested gates.
No lone genius model. Independent models cross-examine each call, and the disagreement is logged before the decision stands.
Every claim here is documented and council-vetted. Go read it.