Checks the portfolio every morning, reads what the news means for her positions, asks the advisor for an explained view, reviews an optimisation proposal. Wants numbers she can trust and explanations she can check.
Needs every claim traceable: which news item, which model, which prompt version, which service computed the figure. Reads the audit trail, never the chat.
Keeps the deterministic core green, watches cost and latency budgets, opens circuit breakers before the demo does, gates every prompt change with the eval harness.
holdings, weights, return, Sharpe, volatility, VaR
live prices — or the snapshot, clearly marked
news → validated Insight: sentiment, tickers, event type
explained, cited answer; numbers from services
allocation, frontier, caveats; decide herself
Steps 1, 2 and 5 work unchanged (deterministic). Step 3 shows the last completed ingestion with a "paused" banner; step 4 says "advisor unavailable — here are your figures". Nothing is silently missing, nothing is invented.
request id → agent trace → each LLM call (prompt version, model, tokens, latency) → guard verdicts → cited Insight ids → the service calls that produced every number. Exportable as an audit record.