Edit parameters and advice

Steps

1. 1

   Step 1: Select a module

   Open the Workflow page, choose your project under Workflow project, and select a module card on the canvas.

   You should see: the Selection panel on the right shows that module's inputs and settings.

2. 2

   Step 2: Give each agent its advice

   Propose model, Implement, Validate and Train & tune each start their Selection panel with Your advice to the agents. Each agent reads only its own.

   • Good advice says what you know and the agent does not, for example “The signal has 60 Hz mains interference” or “Inference must stay under 10 ms.”
   • If the panel warns that Propose model reads your advice only when its “human_advice” input is connected, connect its optional Expert advice input, or the advice is never read. The “Single loop with your own advice” template, which Quick start opens, has it connected; the plain “Single research loop” template does not.
   • Advice is sent to the model provider and kept in version history. Do not paste secrets.

   You should see: each module you advised reads that text first on the next run.

3. 3

   Step 3: Decide what the results count for

   Select Train & tune and make the two choices under Run policy (required). They have no default; a run cannot start without them.

   • Health checks: Record only notes a failed check and keeps going; Block on failure invalidates that round.
   • Result standing: Diagnostic is for exploring; Scientific lets results be cited as evidence, and requires Block on failure.

   You should see: both Run policy choices on Train & tune are set, with no problem shown.

4. 4

   Step 4: Set the training budgets

   Still on Train & tune, expand Training settings. For each row choose Default or Set value, then enter a number. The reset button at the end of a row returns it to the default.

   For example, set Trial budget (minutes) to 5 and Formal budget (minutes) to 60. A formal round expected to take longer than 60 minutes is then refused before it starts, and the report says why.

   VRAM budgets (GB) are not in this table: turn on Settings › Show advanced tools, then select Edit advanced settings and rules in source. Every setting's meaning and default is on the Train & tune reference page.

   You should see: the heading counts your edits (for example “Training settings · 2 of … changed”), and edited rows say Changed, not saved.

5. 5

   Step 5: Set rounds and stop conditions

   In the Research loop panel, set Final iteration to the most rounds you are willing to run, for example 5. Then you can add stop conditions:

   • Add metric threshold: enter your task's Metric id exactly and a Threshold. The loop stops once the best valid Formal score reaches it.
   • Add no-improvement stop: enter Completed bodies, for example 2. The loop stops after 2 rounds in a row without a better score. It must be below the final iteration.

   Whether higher or lower is better comes from your task's metric; you do not set it here. Details are in Loops and routes.

   You should see: the Research loop panel states the round limit and which stop conditions are checked, in order.

6. 6

   Step 6: Review and save a new version

   Made a mistake? Use Undo (Ctrl+Z), Redo (Ctrl+Shift+Z) or Revert to saved. When it looks right, select Save, read the exact Python changes, enter a Version name and select Save version.

   To go back later, open Version history from the ⋯ menu, select View to see what changed, then Restore this version as a draft and save it like any other edit.

   You should see: Version history lists a new named version. The next run uses it; earlier runs are unchanged.