Working with AI
Give the assistant bounded strategy tasks and review its revisions and evidence.
The assistant can read strategy context, replace supported source fields, run historical tests, and inspect results. A request to improve a strategy can therefore create revisions and runs. Review those artifacts before publishing.
Give one testable task
Include the behavior you observed, the revision or run involved, and what should stay fixed. For example:
Inspect the selected baseline's first losing exit.
Explain which SJS condition caused it before changing the source.
Then try one entry filter. Keep position size, instrument, dates,
interval, capital, and fill delay unchanged.
Run a backtest and summarize the source change and resulting trade records.Use the actual run identifier when referring to an older result. For chart-specific investigations, save annotations first.
Send and inspect
Send saves pending draft changes before queuing the task. The assistant can then create newer revisions as it works. Inspect History and Changes to understand the resulting source, and select its run to examine Chart, Trades, and Logs.
Read explanations alongside the saved evidence. A higher summary P&L does not explain why behavior changed; inspect changed entries, exits, and assumptions.
Keep strategy ownership clear
Ask for SJS changes for the main algorithm. Introduce L3 only when you need its advanced capabilities. Policies are available through L3's frame scope; asking the assistant to use them directly from SJS will not make that API available.
The Lab's model tools do not publish apps or control live trading. Publication remains an explicit action in History.
Manage an active task
Use the task's cancel control when you want it to stop. Cancellation does not erase revisions or results already saved. Review the last completed checkpoint before continuing.
If you edit while the assistant creates a newer revision, the Lab preserves your local edits and shows a conflict notice. Follow the revision workflow to reconcile the versions.
Start the next task with a concrete finding from the previous run. For example, ask why an order remained pending before requesting a different entry rule.