I built an AI trading system and then forbade it from picking stocks

Every AI trading bot I have ever seen online is a black box that claims an edge and asks for two things: your trust and your money. It shows you a green line going up and a testimonial. It never shows you the account statement, and it never tells you what happens when the market turns.

I did not want to be trusted. I wanted to be measured. So I built the opposite of that bot, and the first rule I gave it was that it is not allowed to pick a single stock.

What Aegis is, in one line

Aegis is a private research harness that paper trades a stock strategy across parallel accounts, grades every result against just buying the index, and puts a hard wall between "the computer has an opinion" and "money actually moves." No real money touches it until the paper phase earns the promotion, and that promotion is a decision I make by hand.

The whole project runs on one sentence I kept in front of me: the AI advises, the deterministic rules decide.

The one rule the whole thing lives by

The strategy itself has no AI in it. It is a plain, deterministic set of rules: a volatility expansion signal to enter, a fixed position size, a stop below and a target above. You could write it on an index card. Given the same market data, it makes the same choice every time, and I can read every line of it.

The language model sits on top of that as a veto and nothing else. It reads the recent headlines and the setup for a candidate the rules already want to buy, and it is allowed to say one thing: skip this one, and here is why. It cannot invent a buy the rules did not propose. It cannot size a position. It cannot decide when to sell.

That constraint is the entire point, and it comes from a simple asymmetry. A model that can only remove trades can, at its very worst, make me miss a winner. A model that can add trades, or size them, can blow up the account on a hallucinated headline. One of those failure modes costs me an opportunity. The other costs me the account. So the model gets the harmless job, and the boring rules keep the dangerous ones.

Why gate it at all, if it is only paper

Because the habit is the product. The pipeline is validation, then paper, then live, and each stage has to be earned by evidence from the stage before it. A backtest is not permission to paper trade. A good paper month is not permission to risk real money. Nothing advances because it feels ready. It advances because a number I agreed on in advance came in.

Paper trading sounds like the safe, low stakes part, and it is, but it is also where most of the honesty lives. Real fills, real spreads, real overnight gaps, a real broker that is the source of truth about what I actually hold. The strategy on the index card meets the market it will actually trade in, and the gap between the backtest and reality shows up here, cheaply, before it can show up expensively.

The honest part

Here is the thing the guide for this blog asks me to include, the part the black box bots never do.

Months in, the veto barely fires. The account that uses the language model and the account that does not are within a fraction of a percent of each other. And the strategy as a whole is roughly flat while the index it is trying to beat has climbed several points. Measured against just buying the index and going to the beach, my clever system is behind.

That is not the bot failing. That is the bot doing the exact job I built it for. It told me, with no real money at risk, that the AI layer adds almost nothing in its current role and that this particular strategy has no demonstrated edge yet. The most useful output of a measurement system is the measurement you did not want.

What it does not tell you

Two honest limits. First, it has only run through a rising market, which is the easy case for almost any long strategy. The real exam is a drawdown, when the stops are supposed to earn their keep, and I have not sat one yet. I am running it to the end of the year specifically to see that. Second, "the AI adds nothing" is a finding about this narrow, veto only role. It is not a claim about AI in markets generally. I gave the model the safest possible job and measured that job. That is all.

The rest of this series is about how the measuring works: the eight accounts I run in parallel to isolate one variable at a time, the learning loop that is deliberately designed to refuse most of what it is offered, and the benchmark that grades the whole thing against doing nothing. If you only take one idea from this post, take the wall: let the model advise all it likes, but never let it be the thing that decides.