Research answers name their real sources and stick to data we hold

The research agent had several ways of producing an answer that looked better sourced than it was. We closed them.

Every question goes through the same path. Some kinds of question (comparisons, screens and general factual questions) went straight from your browser to the AI model without being metered, and could fall back to built-in tables of figures. They now go through our backend like every other question. That means they cost tokens, at the price the cost preview already showed. If we can't answer, the tokens are refunded and the reply says so. Deep research now takes the 120 tokens it lists and refunds them in full if no answer comes back, and playbooks are charged to your own balance, with a refund for any sections that fail.

Source labels name the real provider. Answers labelled prices and fundamentals as coming from providers that hadn't supplied them. Labels now name the source that actually served the data: Yahoo for prices and dividends, FRED for economic series, and company fundamentals that we curated by hand for financial years 2022 to 2024 are labelled as curated, not live.

No filler data. Some answers attached unrelated data so they looked complete: 30 rows of SPY prices next to a refusal, an oil chart for a question about copper, or a price table when you asked about fundamentals. Those are gone. If we don't have the data, the answer says so and explains why, and we no longer ask the AI model to fill an empty result from its general knowledge. Where an answer does rely on the model's general knowledge, it's labelled as training data.

One live coverage count. Research and paper trading quoted different universe sizes in different places (about 80, 89, 144 or 590 symbols). They now show one live count of the symbols we hold price data for, read from the same prices the backtest engine uses. Today that's 175 symbols, and the figure updates itself when coverage changes. Ask about a symbol outside it and you get a clear out-of-coverage answer.

Smaller fixes. The consumer price index "year-on-year" change was measured against an 11-month-old base, and the 10-year yield move reported in basis points was really a percentage change; both are corrected. The loading screen no longer plays canned steps or holds the answer back until its animation finishes.