What it can and cannot change

An assistant with your trading records in front of it is only worth having if the boundary around it is exact. This is that boundary, stated in full: Intelligence reads your data freely, and every change it makes stops at a confirmation card that names the change before anything is written. A few things sit outside the boundary altogether, and they are listed at the end.

How a change happens

Ask it to log a trade, and it does not log a trade. It puts up a card in the conversation: what it is about to do, the figures it resolved, and a Confirm button. Nothing is written until you press it. The facts on the card are written by the product from the arguments the change would actually run with, not by the assistant, so what you approve is what happens. Most rows on the card can be edited before you confirm; a fixed row says why it is fixed. Cancel closes the card and changes nothing.

You cannot hand it a standing permission from a confirmation. Every confirmation is a card in the conversation, and none of the four kinds of card offers Always allow: a card sits under one sentence about one proposal, and a control that answered for every future one would be read as answering for this one. If a permission was granted for a tool before that changed, it still runs, the card on the message says the change ran under it rather than offering you a button, and Settings › Intelligence is where you take it back. Anything that runs that way leaves exactly the same record as a change you pressed.

Every change that runs leaves a receipt card on the message that proposed it, and a row in the Activity log in Settings › Intelligence, so the answer to “what did it change” is never a matter of memory.

What it can change, once you confirm

• Trades: log a trade you describe, edit one, set its tags, move it to another account, or delete it. A deleted trade cannot be restored by you or by the assistant; the reversible alternative is a transfer.
• Journal and emotions: save a journal entry, request an AI draft for a day, delete an entry, and save or delete an emotion log against a trade.
• Trading accounts: create a manual account, rename or archive one, group accounts, record or remove a cash adjustment, and ask a linked account to sync now. It cannot connect a broker: that needs your sign-in, in the Link Account flow.
• Portfolio: add or edit a platform, record, edit, delete or restore a transaction, and set a monthly budget or a budget category.
• Goals, caps and alerts: set or retire a goal, set your trades-per-day discipline baseline, and arm or disarm an economic-calendar alert.
• Settings and memory: change profile fields and preferences, and remember or forget a fact about how you trade. Memory has its own article.

The journal draft has two protections worth knowing. It will not touch a day you have already marked as logged, a completed entry is refused before anything is written, so a stray request cannot overwrite work you finished. And asking twice for the same day does not queue a second draft; the repeat maps back onto the first request.

It also will not replace an entry that already exists for that day, edited or not. If there is something in that lane, it stays.

What it cannot do, no exceptions

• It cannot connect a broker or hold a credential, and it cannot turn a manual account into a synced one.
• It has no connection to a broker and no route to your money, it cannot place, modify or close an order, and it cannot withdraw.
• It cannot restore a deleted trade. Only Support can, and only within 30 days.
• It cannot delete your Alltra account or change how you sign in.
• It cannot make any change without the card, unless a standing permission you granted before those were withdrawn still covers that one tool.

Why that matters more than it sounds

Intelligence reads text you did not write with an assistant in mind, trade notes, journal entries, tags, symbol names. Any of that could contain a sentence written to look like an instruction. So everything it reads is handed to it inside a labelled data block, and it is told plainly that the contents are query results and never orders. A note that says “ignore your previous instructions” is a string in a field, and it is treated as one.

That is the reasoning behind the confirmation card as well. The instruction side of an assistant can never be fully trusted when the data side is open to the world, so the answer is not to trust it harder, it is to make sure nothing important happens until you have seen it and pressed.

Two things it will not say

• Anything forward-looking: it will not call a trade, recommend an instrument, or set a stop or a target for you. Everything is framed as what has already happened in your own record.
• A number it did not read: if the data does not answer the question, it is instructed to say so rather than to produce a figure that sounds right.

When a request falls outside all of that it declines, briefly, and points you back at what it can look at, a day, a setup, a stretch of the calendar. A refusal is not an error and nothing is wrong at your end.

One thing that is worth checking behind it: every figure it gives you exists somewhere in the Tracker, so a number you want to rely on can always be confirmed on the page that owns it.