The Research Agent no longer gives bull, base and bear cases

Until 13 October, a question to the Research Agent such as "is NVDA a good buy?" came back as three cases, bull, base and bear, each with a percentage and a range of returns. None of that came from our data. An AI model wrote it with no figures in front of it, so the percentages and the ranges had no source. We have removed that view.

Why. It broke three of our own rules. Every figure should link to where it came from. Nothing should be worded as a forecast or as odds. And nothing should read as advice to buy or sell.

It reached further than opinion questions. The Research Agent page sent that view any question that named a company without one of a short list of words such as "P/E" or "revenue". So "How risky is NVDA?", "NVDA last 30 days" and "What are Tesla's main risk factors?" also came back as bull, base and bear cases when asked on the page, although the research service answers each of them from data. Of the 70 example questions we published for the Research Agent, the page sent 27 to that view. All 27 now go to the answers built for them, at the prices we publish for those answers.

What you get now. Ask for a view of a share and you get a report from our data. The figures are pulled first: how the share has done and how risky it has been, how it is valued against its own last five years, growth and margins from its filings, and the risks its latest annual report lists. An AI model then writes the words from those figures only. Code checks every sentence it writes. A sentence with a figure that is not in the data, or with no source number, is removed. So is a sentence that says to buy or sell, or gives odds, a price target or a forecast. The report says how many sentences were removed. Every figure is listed under the report with its source and date.

What it will not say. Whether to buy or sell. The report opens by saying so. It gives no odds and no forecast.

Three questions to ask next. Under the report are three questions, each with its price on the button: how risky the share has been, whether it is expensive compared with its own history, and what its annual report lists as risks. One tap sends the question.

Shares in London and Europe. We have no licensed source for company filings outside the United States, so a report on a UK or European share covers its price figures only, and says so.

What it costs. A report is 40 tokens, the same as any research answer an AI model writes. If the model's words are not available, you get the figures alone for 10 tokens.

"Why" questions. A "why" question that names a company, such as "Why is NVDA down today?", gets the 40-token answer from web pages, each one cited. A "why" question no longer switches Deep Research on by itself. You switch it on when you want it, and its price is shown before you send.

Older answers. A bull, base and bear answer already in your research history still opens. It now carries a line saying it was an AI model's own estimate with no source.

Credit spreads. The Research Agent no longer shows the high-yield or investment-grade credit spread. Those two series belong to ICE, and we do not have ICE's permission to reproduce them. Ask for one and the reply says so. It points to the investment-grade spread on the Market Pulse page, which we work out ourselves from figures the US Treasury and the Federal Reserve publish.