Every research figure now links to its primary source

From 28 September, every figure in a Research Agent answer carries a numbered marker that links to where it came from, with the date: the SEC filing for company figures, the FRED series page for economic data, the price history page for prices and dividends, and the page itself for web sources. Where an AI model writes prose, it writes only from figures we supply, and every number it writes is checked.

Update, 13 October 2026: one kind of answer did not follow this until 13 October. The Research Agent page sent many questions that named a company to a bull, base and bear view, which an AI model wrote with no data and no sources. That view has been removed, and those questions now get answers from our data.

Company figures from SEC filings, with the right years. On 27 September we moved company fundamentals from a hand-curated table to SEC EDGAR 10-K filings, after an audit found 45 of that table's 102 rows mislabelled or wrong. Asked how Nvidia's revenue changed over the last two years, the old answer said revenue grew 383.8% in "FY2024". That compared fiscal 2025 with fiscal 2023, under the wrong year. The answer now takes fiscal-year labels from the filings: $60.9B in fiscal 2024 (the year ended 28 January 2024), $130.5B in fiscal 2025, up 114.2%, and $215.9B in fiscal 2026, up 65.5%, each linked to its 10-K. Year-on-year changes are only worked out between consecutive fiscal years, every metric you ask about is answered or declined by name, and operating margin is included when you ask for it. If the live SEC lookup is unavailable, the answer comes from a dated offline copy of the same SEC data and says so.

AI prose from our data only. Comparisons, sector views, screens and investment theses used to be written from the AI model's general knowledge. Now the model gets a fact sheet (our stored prices, 10-K figures, sector map, dividends and headlines) and writes only from it. Every number it writes is checked against the fact sheet. A sentence with a figure that isn't there is removed, and the answer says a sentence was removed. The figures themselves are always listed under the prose.

"Why" questions answered from cited web pages. Questions such as "Why did Nvidia's margins rise?" now use a live web search. Each claim links to the page it came from, with the publisher and date, and figures still come only from the data. If the search is switched off or fails, the answer says the "why" part couldn't be answered; it never falls back on the model's memory.

Economic data with dates. Economic series are refreshed daily from FRED. Answers give the date of the latest observation and the series' real frequency (the 10-year Treasury yield series we use, for example, is a monthly average), and 12-month changes are worked out from the index levels.

UK and European shares. Ask about a London or European listing by its symbol, such as VOD.L or SAP.DE, and you get price, return, drawdown and dividend answers in its own currency. Fundamentals and comparisons for companies outside the US are declined with the reason: we have no licensed source for their filings.

Declined rather than guessed. Earnings and analyst ratings are declined, at no charge, because we have no primary source for them, and they are no longer advertised. Rankings by a metric we don't hold, such as P/E, are declined instead of being sorted on missing values, and "lowest" now sorts lowest first.