Ground the answer
An assistant should identify the portfolio snapshot, risk analysis, signal, or market input behind an answer. Unsupported confidence is less useful than a clear statement that data is missing or stale.
Separate explanation from execution
AI can summarize, compare, simulate, and prepare a proposal. It should not create a hidden path around user confirmation, risk approval, or execution controls.
Preserve memory with boundaries
Goals, risk tolerance, avoided sectors, and prior decisions can improve future conversations. This context should be stored by the application, remain inspectable, and never replace current portfolio and market evidence.

