Working with the assistant
What Radiant's assistant can read, what it can change, and what it always asks about first.
Radiant has an assistant that can read your map, answer questions about it, and propose changes to it. It is built on the same model catalogue you choose from in the chat panel, and it works through the same project permissions you already have.
Verifying your email
Chatting with the assistant and running tasks such as Generate title both need a verified email address. Until you verify, the composer is replaced by a prompt to verify it, and Generate title stays disabled.
Where you talk to it
Chat is the panel on the left of an open map. It knows which project you are in, and it can read and change that map.
It sees what you have selected
Whatever is selected on the map when you send a message is attached to that message automatically. A badge above the composer shows what is going along — a node's title, an edge's label, or a count when several things are selected.
This means "explain this node" works without you describing which node. It also means the selection is frozen at the moment you press send: selecting something else afterwards does not change a message already sent.
What it can do
Reading is always available: the map's nodes and connections, the compiled forecasting model and its diagnostics, and Metaculus search. It can also run a fresh simulation to answer a question, though it cannot open a run you saved earlier.
Changing the map is gated by the permission mode, and by default every change is proposed to you before it happens. See Permission modes and Approving map changes.
Ask it to tidy the map and it pushes overlapping nodes apart, leaving everything else where it is. Ask it to rearrange the map, or the nodes you have selected, and it lays them out again in the direction their connections and formulas flow; the connections between those nodes go back to Auto. Ask it to arrange the map around a node and it keeps that node where it is, with the rest in rings by how many connections away they are. Ask it to put nodes in lanes, say one per theme, actor, or stage, and it lays each lane out as a row, with the flow running left to right across them; lanes are only available through the assistant. Radiant works out the positions every time. Before it moves anything, it stores the map as a Before layout version; restore that from Versions to undo the layout.
Nodes the assistant adds without saying where go beside the node they connect to, or in a row under the map when they connect to nothing, so they never land on top of what is already there.
It can be wrong, and it can be misled
The assistant reads content from your map and from the web. Both can contain text that tries to instruct it. The approval step is the containment: read what a change actually does before you accept it, particularly on a map that imported content from somewhere else.
Chats and tasks are different things
A chat is a conversation. It keeps its history, you can rename it, and it stays until you delete it.
A task is one structured job with a defined output — naming a project, finding related Metaculus questions, suggesting a formula, pulling historical data for a base rate. Most tasks start from the menu on the thing they apply to, not by asking in chat. It runs in the background and tells you when it is done.
Most tasks finish by asking you to review the result before anything is applied. A running or review-ready task shows on the node it belongs to, and in the Tasks panel.
When a related-questions search turns up Metaculus questions Radiant suggested before, the review lists them under Suggested before, below the new ones. You can still select them. If nothing new turns up, the review says so.
Starting an AI forecast
Ask the assistant to get an AI forecast for a node whose question is already written, and it starts that task itself instead of estimating the probability itself.
Choose Quick or Full to start the forecast. The dialog shows estimated costs before you choose; the run is paid from your monthly AI credits and can take several minutes. While it runs, Radiant drafts the question's Key Factors and places them around the node, at no extra cost. Where a node already on the map names the same driver, it is linked to the question instead of being drafted again, and that link stays. When the run finishes, its forecast is recorded on the node, whether or not Radiant is open. A Full forecast replaces the drafted Key Factors nobody has changed with its own, and classifying their links uses more AI credits; a Quick forecast keeps the drafted ones. A drafted Key Factor someone has edited or connected to other nodes stays either way. If the run ends without recording a forecast, the drafted Key Factors nobody has changed are removed.
Use the remove action beside the AI prediction on the node to remove that run's prediction and its Key Factors. A Key Factor someone has edited or connected to other nodes is kept, and you are asked whether to remove only the prediction. Removing a forecast does not refund its credits, and its report stays in the Tasks panel. Open a recorded forecast's report from its prediction in the node's Forecasting section, or use View AI forecast report on the node for the latest run.
A node runs one AI forecast at a time, whoever started it. If its question changes before the run finishes, the forecast is not recorded; start a new one for the updated question.
Drafting a forecasting question
Ask the assistant whether something will happen, or how likely it is, such as "Will the Fed cut rates before July 2027?" or "What is the probability of a rate cut before July 2027?", and it puts the question on a new node and starts its question draft, where the drafting assistant writes a first proposal. It gives its own estimate in chat only when you ask for its quick estimate. Ask it to operationalize a node already on the map, or to draft that node's question, and it starts the draft on that node instead. Asking it to build a map or break a topic down gets that instead; when it cannot tell which you want, it asks.
Your request is enough to start the draft, in Manual and Auto alike: there is no separate approval. Its first proposal is paid from your AI credits, as Draft with AI on a node is. When the assistant cannot tell whether you want a question drafted or a map built, it asks first. In Read only it cannot start a draft. The chat then names the node and offers Continue draft, which opens the draft, and a new node is brought into view on the map; the node offers Continue draft too until the question is approved. You refine the question in the draft's own conversation, or ask the chat for a change, such as another deadline, and it changes the draft for you without a separate approval. A node the chat created for the question takes the new wording too, unless you have renamed it. Changing an approved question opens its draft again, and the node keeps the approved question until you approve it again. Once the question is approved the chat's button reads Approved, and if the node or its draft is removed it reads Draft removed. A node that already has a draft, or the same question asked again in the same chat, reuses that draft without asking again or paying for another first proposal.
In the draft, Edit opens the question for your own changes. Once the draft has a conversation with the drafting assistant, Save changes also asks it to check your edit and bring the other fields into line, which is paid from your AI credits. A draft with no conversation, such as one you started with Write it myself, just saves. If the check fails, your edit stays saved and Try again asks for the same check. Approve makes the draft the node's forecasting question; it waits for your edits to be saved and for any reply being written. You can then choose a Quick or Full forecast, or select Not now to leave the question ready without starting one.
Choosing a model
The model picker in the chat composer sets two things: which model answers, and how much reasoning it does before it does. Pick a model from the list, then use the Effort slider at the bottom of the picker, or press Tab while it is open, to choose Instant, Low, Medium or High. Some models always reason, so their slider has no Instant step.
Each model starts at its own reasoning level when you pick it. A new chat opens with the last model and level you chose in this browser, or GPT-6 Luna at High if you have not chosen one yet.
Both can only be changed while the chat is idle, and an existing chat keeps the settings it was using. When a model is retired, its chats move to the model that replaced it, or to the default model if nothing did. A reasoning level the new model does not offer becomes that model's own level.
AI credits
Agent work uses a monthly allowance of AI credits, shown as credits used and remaining behind the dial in the top bar. The allowance refreshes on your signup day each month, and unused credits do not roll over. When the allowance is exhausted, new AI messages and tasks are refused until the next refresh.
Stopping it
Stop ends the current turn. A map change already being applied may still finish, but nothing new starts, and the next message begins a fresh turn rather than resuming what you interrupted.