Every trade goes on trial before it goes on the tape.
Most AI trading systems ask one model for its best idea and take it. The Council does the opposite: one AI argues for the trade, a second is paid to destroy it, and a deterministic judge - pure rules, no model - delivers the verdict. Only ideas that survive cross-examination reach the broker.
01 · The bench
Six seats, one verdict
01
Mechanical
The Screener
Twice each morning a rules-only scanner sweeps US equities for stocks that are actually in play: at least 3x normal volume, a gap of 2% or more, real liquidity, and no earnings landmine that day. No AI involved - just hard gates.
02
AI - builds the case
The Analyst
A Claude model receives each candidate's full brief - price levels, volume profile, news - and writes the bull case as a structured trade plan: entry, stop, target, and the thesis in plain language.
03
AI - attacks the case
The Adversary
A second model with one job: kill the trade. It hunts for overhead supply, fake liquidity, crowded narratives, and weak levels. Some objections - like poor liquidity - are hard vetoes no score can override.
04
Deterministic
The Judge
No AI here. A fixed rulebook scores the surviving case: reward must be at least 1.5x the risk, the stop must be sensibly placed against volatility, and at most three ideas per session get through.
05
Deterministic
The Risk Gate
Position size is computed from a fixed 0.75% risk per trade. Daily loss beyond 2% or weekly loss beyond 4% halts the desk. Sector concentration and total exposure are capped. A kill switch overrides everything.
06
Interactive Brokers
Execution
Approved trades go in as bracket orders - entry, stop, and target placed together on the broker, never managed by hand. Unfilled orders are cancelled by 11:30. Everything is flat by the close. No overnight risk, ever.
The two AI seats run on Anthropic's Claude models. The judge and the risk gate are deliberately not AI: they are pure functions with truth-table tests, so no amount of persuasive prose can move a stop loss or widen a position.
02 · In session
A trading day, Eastern Time
First sweep of gappers and unusual volume before the open.
A final look at the tape minutes before the bell, catching late movers.
If the market never came to our price, the idea expires. No chasing.
Anything still open is closed. The Council does not hold overnight.
A nightly memo grades every decision - including the trades we refused.
03 · The rulebook
Hard limits, no exceptions
The Council trades intraday US share CFDs on a EUR 10,000 book. Every limit below lives in configuration, is enforced by code, and applies before a single order can leave the system. The AI seats cannot see, touch, or argue with any of them.
And because refusing a trade is also a decision, the Council logs a counterfactual for every idea it rejects - so we know not just how our trades did, but how our restraint did.
| Risk per trade | 0.75% of the book |
| Minimum reward-to-risk | 1.5 : 1 |
| Max trades per day | 3 |
| Max concurrent positions | 3 |
| Max per sector | 2 |
| Daily loss halt | -2% |
| Weekly loss halt | -4% |
| Overnight positions | Never |
04 · The road to real money
Shadow, then paper, then proof
Now · from Jul 21
Shadow mode
The full pipeline runs every session and logs every decision, but no orders are placed. We are checking that the data is clean and the arguments are sane before money - even paper money - moves.
Next
Paper orders
Real bracket orders on a paper account. Fill quality, slippage, and the evaluator's nightly calibration memos build the track record - published here as it happens.
Only if earned
Live capital
Real money is gated behind pre-registered performance criteria, exactly like every other book on this desk. If the Council never earns it, it never trades it.
The Council joins a desk that already keeps score in public.
Eight active books, a public archive of retired strategies, and a real-money account - every trade on the record.
Visit the lab