The questions people actually ask.

No. Every desk here is a simulation funded with $100,000 of pretend cash. There is no broker connection, no card, no way to deposit and nothing to withdraw.

The market data marking your positions is real. The money never is. That is the whole point — you get to find out how a strategy behaves before any of your own money is involved.

A backtest replays your rules over history and reports what would have happened. It is the cheapest way to learn whether an idea has any legs.

Most backtests flatter themselves in three ordinary ways, none of them malicious:

• They fill at the middle price. Real buying happens at the ask and real selling at the bid, and the gap between them is a cost you actually pay. We buy at the ask and sell at the bid — never the mid.
• They peek. If a rule can see even a few hours of the future, the result is fiction. A signal here reads only the day it fires and days before it, and the trade enters afterwards.
• They only show the window that worked. Rules tuned on a stretch of history will always look good on that stretch. We split the history and report what the rules did on the part they had never seen, next to the part they were built on.

None of that makes a result a prediction. It makes it an honest measurement of the past.

They are our engine’s read on a result — an opinion, printed beside your own judgement. They are not a gate. You choose what runs on your desk, always.

• LIVE — it held up on history the rules had never seen, on enough trades to mean something.
• SHADOW — the middle read. Promising, but it did not fully hold up on the unseen part, usually because there were too few trades to tell. Paper-trade it and keep watching.
• SHELF — our engine did not find an edge here that it could stand behind. That is our read on this history, not a verdict on you or on the idea.

You can deploy any strategy in any mode regardless of the read, and a strategy you never graded at all trades exactly like a graded one.

Real, split-adjusted daily bars and real daily option marks on liquid US names. Nothing on a result is a made-up price series.

One part is modelled, and we say so everywhere it shows up. Where we have a recorded two-sided quote for the option at the moment of the fill, we use the real bid and ask. Where no such quote was recorded, we model the spread at 2.5% each way around a real closing mark — you buy 2.5% above it and sell 2.5% below it — and every number that fill touches is labelled modelled rather than measured.

So a result tells you two things: what it did, and how much of what it did rests on a measured quote versus a modelled one. We would rather show you that split than average it away.

It means the engine found nothing it could honestly fill, so it declined to invent a result. Usually one of these:

• Your date range is outside the options window. Real option fills only exist from 2024-02-12 to 2026-05-20. A backtest range outside that finds no contracts to trade, so there is nothing to grade.
• Your rule never fired. A condition that no bar in the window ever met produces zero events. That is a real answer about your rule, not a fault.
• The names you picked have no option chains. Plenty of listed stocks and roughly half of ETFs simply have no usable chain in the store.

The fix is almost always to move the range inside the window first, then loosen the rule one condition at a time.

It means your AI gets its own credentials and its own paper desk. Each desk mints a key pair that belongs to that desk only, so your assistant can read the catalog, build a strategy, run a backtest, deploy a bot and report back without you copying JSON between two windows.

It sees exactly the numbers you see. There is no better answer behind the API, and the AI never sets a verdict, a fill or an EV — the engine alone does that.

Nothing, for as long as the founding era lasts. No card, no trial clock, nothing to cancel.

When we do charge, money will buy you room — more symbols, more runs, more desks — and never a different truth. The grading is identically honest on every plan.

No. That is not our job. Our engine gives you its read, and it sits beside your judgement rather than replacing it.

The same goes for your settings. Keep your stop where you want it — we grade the strategy you actually wrote, and if a different stop would have done better on this history we will show you by how much. We will not swap it for you and we will not refuse your idea.

No, and this is the invariant we guard hardest. A signal on a given day reads only that day and earlier — never a single bar past it.

Then the trade has to happen after the decision: a stock position enters at the next day’s open, and an option path starts strictly after the signal date. Which means a rule that would have been brilliant if only it could act on the same close it read gets no such favour here.

Not here. There is no real money in this product to lose.

Two honest warnings anyway. Nothing here is investment advice — we teach and we measure, we do not recommend. And a good grade is a measurement of the past, not a promise about the future. Markets change, and a strategy that held up for two years can stop working the week you fund it for real.

Because our score deliberately leaves it out. Testing a take-profit across a lot of history is one of the easiest ways to accidentally pick the number that happened to be luckiest, and a score built on that number is flattering rather than useful. So the graded comparison is “no stop” against two fixed stops, and nothing else.

That is a statement about our scoring, not about your trading. You can run a profit target on your desk; it just does not get to move the read.

Three of them carry most of the meaning: expected value per trade (what an average trade kept after the spread), win rate, and how many trades the whole thing rests on. A 40% average on nine trades is a story; a 3% average on four hundred is evidence.

Everything shown as a percentage is a percentage of the money put into that trade, so 7.0% means seven percent. If you read the raw JSON from the API, note the one seam: the summary figures are already percent (7.0 = 7%), while a per-trade pnl_pct is a decimal fraction (0.109 = 10.9%).

Probably not. A bot only acts when its conditions are actually met, and a selective rule can sit quiet for weeks. Doing nothing on a day it should do nothing is correct behaviour.

A bot that was never graded is also fine — ungraded is a normal resting state, not a pending or failed one, and it trades exactly like a graded bot. If you want a read on it, you ask for one; we never grade uninvited.

If you think it should have fired and didn’t, tell us with the strategy name and the day you expected — that is a report we want.