01
Propose
The proposer role selects a trade from a filtered set of eligible options contracts.
Alpaca AI Trading Agents Hackathon
Devil's
Advocate
An AI trading agent that argues with itself before it trades. One AI role proposes. An adversarial role challenges. Code decides the final action.
Built with Alpaca · Python · LLM agents · GitHub Actions
Decision path
Market + portfolio statecontext
↓
Candidate filtercode
↓
ProposerAI
↓
AdversaryAI
↓
One revisionAI
↓
Deterministic risk gatecode
↓
Execute or No Tradefinal
How it works
Reasoning is flexible. Authority is not.
The AI can explore ideas and challenge itself. Deterministic code decides whether anything is actually allowed to happen.
02
Challenge
A second model attacks the reasoning, scores severity, and can trigger one revision.
03
Control
Hard rules check position size, collateral, exposure, daily limits, and fresh broker state before execution.
Proof it works
The interesting part is what the system refuses to do.
The project is built to preserve useful model reasoning without giving the model unlimited authority.
Deterministic intervention
4 contracts proposed
→ 2 contracts approved
→ 2 contracts approved
The models did not fix the sizing issue themselves. The deterministic gate did. It generated a safer substitute and re-checked the trade before execution.
Stand down
0 eligible
No proposal reached the model. The system chose no trade.
Gate tests
28 / 28
Latest deterministic enforcement suite passing.
Explore the evidence
The homepage is only the front door.
Each deeper page answers a different judge question without turning the homepage into a wall of information.
Evidence
Judge Report
Calibration, decision stories, live activity, limitations, and the compact audit ledger.
→
Trace
Decision Traces
Follow individual decisions from proposal through challenge, revision, gate, and final outcome.
→
Build
Source Code
Inspect the agent, deterministic gate, tests, automation workflow, and report builder.
→
Core idea
Models make judgments. Code controls authority.
That boundary is the product.
“A bad model decision should not automatically become a bad action.”