The part nobody wants to read
Most people who lose money trading do not lose it on one bad idea. They lose it by putting too much into one position, then trying to win it back the same afternoon.
A worked example. Say you have $10,000 and you cap yourself at losing 1% on any trade. That is $100. If your stop is 5% below your entry, the largest position you can take is $2,000, because a 5% loss on $2,000 is your $100. Change the stop to 10% and the position halves to $1,000.
That is position sizing, and it is the whole of risk management in one paragraph. The size follows from the stop, not from how confident you feel.
Guardrails are the automated version. You decide your limits once, while calm, and the terminal enforces them later, when you are not. Find them in Settings → Risk Management.
What you can cap
Max daily loss (%). The most you are willing to lose inside a window, as a percentage of your portfolio. This is the one that ends bad days before they become bad weeks.
Max position size (%). The largest single position you will hold. On a perpetual this measures your total exposure, not your margin: one Bitcoin contract is the same exposure at 1x as at 25x, and leverage only changes how much the exchange holds against it.
Max trades. How many trades you may place inside the window. This is the one that saves you from revenge trading, which is the specific failure of trying to win losses back immediately.
Leave any of them unset for no limit.
What happens on a breach
Each limit has its own action, so you can be strict about losses and relaxed about trade count:
| Action | Effect |
|---|---|
| Block all orders | Nothing goes out until you unlock |
| Block buy orders | You can still reduce or exit, but not add |
| Warn only | The order goes through, with a warning |
| Off | The limit is tracked but never acts |
When a limit blocks, a banner appears across the terminal and the Risk panel switches from All clear to Limit hit. Unlocking is a deliberate click.
Blocking buys rather than everything is the underrated option. A blown daily loss limit usually means you should stop opening new risk, not that you should be trapped in what you already hold.
The reset window
Your tracked profit and loss, trade count and any order lock reset on a schedule you pick: every 4 hours, every 12 hours, daily, or weekly. The Risk panel shows the current window’s numbers, when it started, and how long until it resets.
Order confirmation
The last guardrail is the one in your hand. By default you commit an order by pressing and holding the submit button, and a live order holds longer than a practice one. It costs half a second and it catches the order you did not mean to send.
If that half second is in your way, the same settings page switches it to a single click. The note under the button always says which one is in force. Nothing else changes: the caps above are enforced identically either way.
What the AI is allowed to do
The same page controls what the assistant may do without asking.
Auto-approve paper trades. Simulated funds only. The assistant places practice orders with no confirmation card.
Auto-approve live trades. Off by default, and granted one exchange at a time. The only way to grant it is to tick Don’t ask again on a live order card, so the first live auto-approval is always a deliberate act by you. Every exchange you have granted is listed here with a revoke button.
Auto-approval skips the confirmation card. It does not skip the guardrails. Every order the assistant places is checked exactly like one you typed.
Bots have their own guards too
A bot carries limits of its own, set when you create it: a daily loss cap, maximum trades per day, maximum position size, a cooldown after a loss, and a maximum losing streak. Those sit on top of the account-wide limits here. A bot that trips one of its own guards halts itself and tells you which one stopped it.
Why the AI cannot touch any of this
Keeping guardrails out of the model’s reach is deliberate, and it is the single most important design decision in the product.
A language model can be wrong. It can be argued into things. It can be manipulated by text it reads while researching, because it cannot always tell a web page’s instructions from yours. None of that is acceptable in the thing standing between you and a margin call.
So the limits are enforced by the code that sends orders, below the AI entirely. The assistant can analyse, warn, and propose. It cannot raise a cap, disable a block, or place a live order you did not approve.