The failure nobody catches
Everyone worries about AI inventing statistics. Fair enough. But an invented number is the easy problem. You can check a number. The dangerous failure is quieter. It's the sentence that reads like analysis and isn't grounded in anything.
It looks like this. A source reports that exports rose 12%. Another mentions a new trade agreement. Neither connects them. The AI writes: “Exports rose 12%, driven by the new trade agreement.” Every fact in that sentence is true. The word driven is invented. And it's the only word that matters, because it's the one a reader will act on.
Or the report says a company is “likely repositioning.” Nobody said that. It's a guess wearing a hedge. Or a plan that was set to begin in March quietly becomes something that began in March. Or one company's ranking gets stretched across the three companies listed beside it.
None of these trip a fact-checker. Every name, date and figure survives inspection. What's been added is confidence the evidence doesn't support — and confidence is what you're buying a research report for.
Two jobs, two systems
The usual way to build an AI research tool is to have it search the web and write the report in the same breath. That's the root of it. When a tool searches and writes at once, it can't separate what it just read from what it already thought it knew.
We split the job in two, and give the halves to two different systems.
The first system only collects. It isn't allowed to write prose, do arithmetic, or convert a currency, and anything it returns that isn't a fact is thrown away. Its only output that survives is a list of facts. Each one arrives with its source, its date, and a word-for-word quote from the page it came from.
The second system only writes. It never touches the internet. It can only work from the list the first system built. If something isn't on that list, there's nothing to build a sentence out of.
That split is the foundation. What sits on top of it is a set of rules about what the writing system is allowed to say.
What gets collected
A fact doesn't make the list just because it was found. It has to clear a set of rules first.
A quote, or it doesn't count. Every fact carries the exact sentence from the source that backs it up — the number, the units, the date, the scope. A link on its own isn't evidence.
Numbers stay as published, and nothing gets worked out. No converting or rescaling to make something fit — if a source reports in yen, you'll see yen. Growth rates, shares and totals aren't calculated while facts are gathered. Only what a source actually states gets recorded.
No invented cause. This one matters most. A fact is recorded as what happened — never with a “because” clause the source didn't write. If a source reports a change without explaining it, we record the change.
Percentages carry what they're a percentage of. A share with no stated base isn't a usable fact, so it doesn't become one.
Where it comes from matters. Company results come from the company's own filings. Prices come from the operator's own price list. Social posts, videos and forums can add context, but they never supply a number, a date or a ranking.
What the writer is allowed to say
This is the part that does the heavy lifting, and it's the part most people never think to ask about.
The rule underneath all of it: evidence outranks emphasis. A sourced claim is never sharpened, simplified, broadened or strengthened to make it land harder. What the source said is the ceiling.
No cause the sources didn't establish. “Because,” “due to,” “driven by,” “as a result of” — these only appear when a source actually makes that link. Where the evidence doesn't explain something, the report says the cause isn't established. That's more useful to you than a confident guess, and it's honest.
No speculation in our own voice. “Likely,” “probably,” “presumably,” “it's reasonable to assume,” “would appear to” — these are banned as the report's own analysis. The only exception is when a named source says it and we quote them directly. Softening a claim doesn't excuse it from needing evidence. A hedge isn't a source.
No superlatives without proof. “First,” “only,” “largest,” “leading,” “unprecedented” — every one of these needs a source that actually says it. And a ranking for one company doesn't license ranking the others sitting next to it in the same list.
Forecasts stay forecasts. Something “expected to complete” in March doesn't become something that completed in March. If the evidence describes a plan, the report describes a plan.
Scope stays attached. A division's revenue is labelled as a division's revenue, not the company's. A regional figure isn't quietly promoted to a national one.
One claim, one source. A sentence has to stand up against a single piece of evidence. Two facts from two different sources can't be welded together into a relationship that neither of them states. That's the most common way an AI manufactures an insight, and it's specifically blocked.
No filler. Sections are as long as the evidence supports and no longer. Three well-sourced findings beat six thin ones. If a section is short, that's because the evidence was.
No names the sources didn't give. If a source names a role but not a person, the report names the role.
When the evidence isn't there, we tell you
This is the part most tools don't do.
Before writing anything, we collect everything the report needs and count it. If a section hasn't turned up enough solid evidence, we don't write around the gap. We shorten the section, drop it, or stop and come back to you.
You'll hear from us before the report is written, not after. Sometimes that means we say the coverage isn't strong enough and ask whether you'd like to change the focus.
We'd rather tell you the data isn't there than hand you a confident page that's holding air. A gap in the evidence is a finding in its own right, and often a useful one.
Then we check the writing
Once the report is drafted it goes through around twenty-five checks before anyone sees it. Some look for numeric faults — a figure that doesn't match the table above it, a total that doesn't add up, two sources quietly disagreeing.
The rest read the prose against the evidence behind it. Does the cited source actually support this claim, or just sit near it? Does the sentence describe the same company, the same period, the same basis as the fact it rests on? Has a conclusion crept in that the evidence doesn't carry?
Every fault those checks confirm has to end up in one of four places. It gets fixed, it gets removed, it gets disclosed to you in the report, or it gets logged against the report. A confirmed fault never sits in the middle, noticed and ignored.
And once something has been dealt with, it's locked. The same fault can't quietly reappear in a later report without the system catching it.
What we don't claim
We'd rather be straight with you about the edges.
We can stop a claim being invented. We can't make a source correct. If a credible publisher prints the wrong figure and we quote them accurately, that figure appears in your report with an accurate citation. To reduce the risk we go to primary sources first — regulators, filings, official statistics — and where reputable sources disagree, we show you both rather than quietly picking one.
We don't publish an accuracy percentage. We could put a number on this page. Plenty of people do. But we haven't run the controlled study that would make it true, and putting an unproven figure on a page about not inventing figures would be a strange way to earn your trust. When we've measured it properly, we'll publish it here.
We show our working. Every report ends with its sources, what those sources disagreed on, and what we looked for and couldn't find. You don't have to take our word for any of it. That's the point.