How to check a number an AI tool gives you

A fluent answer is not a checked answer.

An AI chat tool builds its answer by predicting likely next words, one after another. That makes its answers read smoothly whether or not the facts in them are right. A figure can have the right units, the right number of decimals and a confident sentence around it, and still be wrong, out of date, or about something else.

So the tone of an answer tells you nothing about its accuracy. What tells you something is whether you can find the figure, in those words, in a document you can open. The three habits below turn an answer you cannot trust into one you can check in a few minutes.

Ask for the document and the page behind every figure.

For each number in an answer, ask which document it came from, who filed it, on what date, for which period, and where in the document it sits. Then open the document and find the passage. If the tool cannot name a document, treat the number as unsupported until you find it yourself.

Here is why the period and the kind of document matter. Hashrate is the computing power a bitcoin miner runs, counted in exahashes per second and written EH/s. Ask what hashrate MARA reached by the end of 2025 and a careless answer can say 75 EH/s. That figure is real, but it is the target the company set in a press release on Jul 1, 2025. The shareholder letter filed on Feb 26, 2026 reports 66.4 EH/s of what MARA calls energized hashrate at Dec 31, 2025: the hashrate its energized miners could produce in theory if all of them were running, including any that are temporarily offline, estimated from the manufacturers' specifications. Both numbers are the company's own; only one answers the question.

The same trap runs the other way. Bitdeer set a 50 EH/s target for the end of 2025 and reported 55.2 EH/s of self-mining hashrate for December. An answer that quotes the target as the result understates it. A count also needs its name: the same Bitdeer report lists 58.0 EH/s of proprietary hashrate and 71.0 EH/s under management, so an answer has to say which one it means.

A short record for each figure keeps all of this apart. Three of the four documents in it are exhibits to filings with the SEC: an 8-K is the form a US company files to report news, a 6-K is the counterpart for a company based abroad, and exhibit 99.1 is the attachment where companies usually put a press release or letter. The fourth, MARA's June 2025 production update, is a press release on the company's website, so it has a date but no filing date.

What to record for each figure before you use it. The figures are the companies' own, from the documents under Sources. Dated is the day a filing reached the SEC or the day a press release came out.

What to record for each figure before you use it. The figures are the companies' own, from the documents under Sources. Dated is the day a filing reached the SEC or the day a press release came out.

Figure	Document	Dated	Kind
MARA, 75 EH/s by the end of 2025	June 2025 production update, a press release	Jul 1, 2025	Target
MARA, 66.4 EH/s energized at Dec 31, 2025	Q4 2025 shareholder letter, 8-K exhibit 99.1	Feb 26, 2026	Result
Bitdeer, 50 EH/s by year-end 2025	November 2025 update, 6-K exhibit 99.1	Dec 15, 2025	Target
Bitdeer, 55.2 EH/s self-mining in December 2025	December 2025 update, 6-K exhibit 99.1	Jan 12, 2026	Result

Not stated is a good answer.

When the documents do not contain a figure, the right answer says so. A tool that fills the gap with a plausible number has made something up, even if the number happens to be close. When you ask a question, say plainly that the answer should come only from the documents you name and should read not stated where they are silent.

A real example comes from IREN. Its quarterly report filed Feb 5, 2026, for the quarter ended Dec 31, 2025, gives a net electricity cost per bitcoin mined of about $34,400. Its quarterly report filed May 8, 2026, and its annual report filed Aug 27, 2026, show the cost of revenue for bitcoin mining as a total and give no cost per bitcoin. So the honest answer to "what did IREN spend per bitcoin in the quarter ended Mar 31, 2026?" is that the report for that period does not say. A figure for it that comes with no document is an estimate, and it should be labelled as one.

Do the arithmetic outside the chat window.

A language model predicts text, and a sum is text it can predict wrongly. Take the filed inputs out of the answer, put them in a spreadsheet or a calculator, and compute the result yourself. It takes a minute and removes a whole class of error.

For MARA: 66.4 minus 75 is minus 8.6, and minus 8.6 divided by 75 is minus 11.5 percent. That is the shortfall the promised and delivered article prints. If an answer gives a different percentage from the same two inputs, the inputs or the arithmetic are wrong, and you now know which to check.

Where these habits come from.

The ideas in this article come from the free Clief Notes courses by Jake Van Clief, including Bas Rosario's lessons on context and on diagnosing a weak answer, which advise taking a job like adding up rows away from the model and giving it to a spreadsheet. As Jake Van Clief puts it in the lesson Computational Orchestration: "A simple regex in the right spot beats GPT-4 in the wrong spot." A regex is a short rule that finds a pattern in text, and GPT-4 is a large AI model; his point is that the right tool in the right place beats a bigger tool in the wrong place. The courses are free at skool.com/cliefnotes. The mining examples and the wording here are MinerTerminal's own.