The AI reviews a day, a trade, a week or a period and answers questions about your statistics. The engine computes the numbers first, and a second, fast decision model picks the trades to review and checks the claims against the data.
It gives you a sentence with no arithmetic behind it: the numbers were generated along with the text. You can neither confirm nor refute it, and it sounds more convincing than the truth.
code computes, one model decides, another explains
How it works here
Metrics, limits, tilt and streaks come from an engine under tests. The model receives finished numbers and is kept out of the maths.
It answers narrow questions with a probability: does a trade break your rules, does a claim hold on the data, which agent should get the question.
The review model writes the reviews, the chat answers and the agents' work: the platform model or your own by key. It sees the computed numbers, your dossier and your rules, and every call goes into the trace.
Reviews of a day, a trade, a week, a period, a playbook or a prop account. The model sees the engine's metrics, your dossier and your rules, so it judges a trade by your style, not a textbook.
Five hundred trades won't go into a model whole. The engine sets flags: no stop, a loss beyond −1.5R, the best by R. The decision model scores each trade against the rules in your dossier, and 10–40 of the right ones go to the review.
The finished review is split into claims about the data, the ones with a number or a comparison. The decision model checks each against the facts and marks it. Advice like 'reduce your risk' isn't checked, because it isn't a fact.
In the journal before entry you enter the instrument, direction and risk. Code checks the remaining daily limit, trade count and tilt, and the decision model rates the entry against your rules. A hard limit in code outranks the model: risk above the remaining limit means stop.
On paid plans the AQREVI model answers by default, and it's included. You can connect your own key from Anthropic, OpenAI, Google, xAI, DeepSeek, OpenRouter and others, and then no operations are spent. On Free the model runs on your own key.
The call log is append-only: the run, the step, what went out, what came back, what it cost. The trace screen is built from that log, so there's no second version of events.
Where it differs
Trading AI usually looks at a chart and calls a direction. Ours doesn't read charts and never says long or short. It explains numbers already computed from your trades, and a second model checks whether each claim holds on the data.
Check it yourself
Open the trace of any answer: the request, the tool calls, the returned text and the cost in operations. Under a review sit the decision model's marks: supported, hypothesis or recheck.
The order
On a paid plan the AQREVI model is already connected. On Free, add your own key in settings.
About a metric, a day, a streak or a trade. The model gets the numbers, never access to computing them.
The decision model's marks sit under the review, and every number links to its metric. If you disagree, open the maths.
What it does not do
Questions
A model for short decisions. It answers a narrow question with a probability: yes or no, which option. It's cheap and fast, so it checks what would be costly for the review model: every trade and every claim. The mechanism doesn't depend on which model writes the review.
The journal lives with us, in your account's own database. The model provider only gets the answer's context: with the AQREVI model, the provider we use; with your own key, the one whose key you entered.
No. It reads your statistics. Other people have already sold you price forecasts.
It can't get a number wrong, because it doesn't produce numbers. It can misread them, which is why the decision model flags claims that don't hold and every number links to its metric.
In operations, not tokens. A chat question costs one, a day review three, a period review five, an agent answer ten. Free has 30 operations a month on your own key, Core 300, Pro 2,000, Apex 7,500.
The numbers under it are computed separately and open, and a second model has marked the claims. A confident sentence without arithmetic leaves you nothing but trust.