It refused to say the strategy worked.

What it wrote instead: not negative on this sample.

If you decide where client money goes, that sentence is worth more than the result.

The strategy trades pairs of crypto assets that usually move together. The test replayed twenty days of market data to see how its orders would really have been filled.

Then it explained why the result does not mean the strategy works.

A handful of days carried the result. Ten of the twenty days, added together, were negative. Twenty days, it said, was not enough to call the strategy ready to trade.

Next it listed every bias it could not remove from the test, and which way each one would push the result.

Then it named what the test had not modelled at all: how the trades would affect each other inside one portfolio, and how far our own orders would move the price beyond what the order book showed.

The test had also left out funding. On perpetual futures, the crypto contracts the strategy traded, one side pays the other at set times of day. Many trades were still open when a payment came due.

The next step it asked for was not to put the strategy live. It was a longer test, over 90 to 120 days instead of twenty.

Good numbers are easy to come by. Before client money goes near one, you want the list of ways it could be wrong.

Ask that of whoever brings you a result, a person or a machine: what it did on this sample, which days carried it and what the test left out.

Every result from our AI copilot comes with that list: what the test could not account for, and which way each gap would push the result.

If you want research reported to you that way, write to robert_yenokyan@bonton.ai.

What the test did not model at all: funding, how the trades would interact inside one portfolio, market impact beyond the visible order book, a placebo test and a correction for how many variants had been tried. The next step was a run over 90 to 120 days, because a handful of days carried the result.

The client was disappointed in our AI.

It turned out the client's result was the wrong one.

The client had a trading strategy with good results. It was written up and coded. They wanted to see whether our agents could rebuild it and get the same result. And how long that would take.

Our agents rebuilt it. Their result was nowhere near the client's.

So we dug into the client's implementation and found a couple of small things. The kind that look like no problem at all.

Their strategy looked at the closing price of each candle, the bars on a price chart, to decide when to trade. Then it got in and out at that same closing price.

But you only know a candle's close once the candle has finished. After that, the decision takes a moment to compute and the order takes a moment to reach the exchange. That delay is called latency. Their test had none, so it filled every trade at a price you could not count on getting.

On one trade the difference is tiny. Over millions of trades it adds up.

A proper test also charges realistic fees. And it counts the cost of your own order pushing the price against you as it fills. That cost is called market impact.

With the delay, realistic fees and market impact counted across all those trades, the strategy was not worth trading.

A tool that copies your result copies your mistakes along with it.

Our agents are built to call our validated backtesting libraries instead of writing that code themselves, wherever the library has what they need. Those libraries account for the delay, realistic fees and market impact. That makes the agents less prone to wrong results than an AI writing everything from scratch.

The agents are our AI copilot, and it is in private beta. They design, run and check trading research on our infrastructure. You see what they built and what came out in a web portal, or call them from your own code through an API.

Founding users help shape the product and their requests get priority. We talk to everyone who applies.

If you have a strategy with good results and want to know whether they survive a rebuild, email robert_yenokyan@bonton.ai with what you would use the copilot for.

You can check your own backtest for this without any tools. Take a few trades and compare the price each one was filled at with the price its signal was computed from. If they are the same number, the test is filling trades at a price you could not count on getting.