There is a version of AI powered trading that is mostly marketing: point a language model at the market, ask it what to buy, act on the answer. It is a bad idea for a boring reason. Language models are good at reading and characterising text and they are not calibrated instruments for sizing financial risk.
The split
So we split the job in AlphaEngine along the line the technology actually falls on. The model reads news and grades the catalyst: is this a real, material event, or noise dressed up as an event? That is a reading comprehension task, and it is genuinely good at it.
Everything downstream of that grade is deterministic code. Position sizing, entry, stop placement, portfolio limits. The risk gate sits in front of all of it and cannot be overridden by the model. The model is allowed to lower risk. It is not allowed to raise it. If the model produces nonsense, the worst case is that the system trades less, not that it trades recklessly.
What the testing actually showed
The useful result was negative, and worth stating plainly: on price and volume signals alone, across a decade of data and after realistic costs, the strategies were flat to slightly losing. Not catastrophic, just not an edge. A great deal of published trading strategy stops being profitable the moment you charge it for spread and slippage.
Adding the catalyst grade changed the ordering in the direction the thesis predicted: higher graded events performed better than lower graded ones, consistently. That is a real signal. It is also, on its own, a thin one, which is why the system is still in testing rather than trading.
Why we are saying this before it works
Because the alternative is the industry norm, which is to describe a system as finished the moment it demos well. AlphaEngine has a fixed ladder to climb before it trades real money at any size: statistical validation on out of sample trades, then an extended paper trading period, then live at a fraction of intended size. If it fails a rung, it does not advance. That is the whole discipline, and it is the part worth writing down.