How every number is checked

All four run on every headline figure. Each takes under a minute, and each has already caught a live error.

1. Scale: does the figure fit the company's own recent size?

The figure is compared against the prior quarter and the year-ago quarter before anything else. If it implies a company halved or doubled without the source saying so, it is wrong until proven otherwise.

What went wrong: Roblox was published at $801m of quarterly revenue when the company was already running near $1.5bn. That single comparison would have stopped the whole article.

2. Arithmetic: do the level, the comparison and the percentage multiply out?

If a source says "$111.2bn, up 16.6% from $94.0bn", the division gives +18.3%. At least one of the three numbers is wrong, so the source is unreliable for all of them. Percentages and levels must agree or neither is published.

What went wrong: This caught a bad Apple revenue figure, and it settled a KOSPI close: 6,595.45 reconciles against a 5,593.71 prior close at +17.91%; the widely syndicated 6,695.45 does not.

3. Guidance: is the reported figure inside what the company guided?

A result outside management's own range is possible but rare, and far more often it signals a bad source than a blowout. Anything outside the range is treated as suspect until the release itself has been read. This check now runs as a dataset too: the guidance axis of the earnings expectations database records the guided range and the printed figure for every tracked quarter.

What went wrong: The wrong Apple revenue figure sat above Apple's guided 14-17% growth range.

4. Sources: two independent, at least one primary or tier-1

Company press release, shareholder letter, SEC filing or investor-relations page first; CNBC, Reuters and Bloomberg second. Aggregators and SEO content farms never establish a figure on their own. When a primary source cannot be reached, the article says so and rests on two tier-1 outlets that agree.

An intraday snapshot, a regular-session close and an after-hours print are three different numbers, and conflating them has been the single most frequent error on this site. Every price move names its session. The close is preferred, and the article says which close.

What went wrong: The Nasdaq's "+1.6%" was a mid-morning read published as a full day; it closed +2.48%. Reddit was published at +12% from the pre-market and closed down 20.76%. Microsoft was published at 2-3%, then 8%, then 9.5%, all on the same day.

Before anything is called a beat, the figure is taken apart.

Operating or non-operating. Amazon's $5.75 EPS contained a large gain from revaluing an equity stake. Quoting it as a 216% beat would have been meaningless; operating income was the honest comparison.

One-off items. Of Apple's $2.02 EPS, 11 cents was tariff refunds, and 2 of the 50.1 gross-margin points were the same refunds. Ex-refunds it was a two-cent beat and a margin inside guidance, which is a different article.

Revenue or bookings, reported or adjusted. The article names which one, because for some companies the forward number and the recognised number diverge sharply.

Corporate actions. A "21,251% rally" in DFNS was a reverse-split artefact. Splits, spin-offs and offerings are checked before any move is quoted.

The four checks catch numbers that are wrong. They do not catch a set of numbers where each one is defensible and the set does not close. Three habits cover that.

Share count, price and total are multiplied out whenever they appear together. "555,555,555 Class A shares at $135 to raise $86 billion" is $75bn. The raise was right and the share count was incomplete: the greenshoe added 83,333,333 shares, so 638,888,888 were sold, which does give $86.2bn. Nobody had to be wrong for the sentence to mislead.

No numerator without its denominator. "911.5 million shares unlock" against a 555.6m float is +164%; against 13.1bn shares outstanding it is 7%. Both are true and they support opposite conclusions, so the reader gets the base every time. The lock-up calendar carries both bases on every tranche for this reason.

Dollar values of share tranches are dated. The same 911.5m shares were reported at $123bn, $116bn and about $100bn within two weeks because the price moved. The share count is the durable number; a dollar figure is quoted with the price or date it was struck at.

A source's hedge is resolved, not repeated. "A further 455.8m shares could release under certain conditions" left a large number hanging. The condition was a price trigger at $175.50 against a stock near $110. It could not trigger, and saying so was the whole value of the sentence.

1. Primary source first, then the tier-1 outlet, then arithmetic consistency with figures already verified.
2. If it cannot be resolved, the safest defensible framing is published and the contested precision is left out. Samsung's July 31 close was reported anywhere from 25.6% to 28%, so the article said "about 27%" and led on the verified $1 trillion market-cap fact.
3. An unverified magnitude never goes in a URL. A slug without a number survives a revision; one with -9-percent in it does not.

Figures supplied by the site's own editor are verified the same way. "MSFT up 15%" was 9.5%; "SNDK up 32%" was 26.29%; "AAPL down 9%" was 6.6% at the time and 8.1% at the low. The verified figure is what gets published.

If the direction or the magnitude was wrong, the URL was wrong too. The piece is republished at a corrected address, inbound links are fixed, and a correction section records what the old piece claimed so the error stays visible. Smaller factual fixes are corrected in place and dated.

The same rule applies to the site's own calls. The track record logs every published play, including the losers, and its methodology says how each is graded. Spotted an error? Send it in and it is checked against the sources above.

This protocol is the standard behind every figure in the site's open datasets, which are published under CC BY 4.0. To cite the standard itself:

Atul Ghandhi, "Figure Verification Protocol: How Every Number on This Site Is Checked", Regards of Wallstreet, updated 2026-08-22, https://www.regardsofwallstreet.com/methodology/verification