AXHUB DEEP READ · 05

What it takes to publish a single number
— the rules we use to filter figures

Reading time about 8 minType operating log — editorial principles disclosedFor all our readers, and teams who write reports

The cards, lectures, and articles on this site are full of figures. Which means there are just as many chances to get them wrong. So we made rules for publishing figures, and we've kept a record of what the rules actually caught — 1 correction, 1 discard, 2 softenings, 1 hold. We're disclosing those rules and that record. Because in an age when AI pours out writing, we believe this tedious procedure is what sets us apart.

1The principle — into the library first, only registered ones into content

The rule is one sentence. Every figure that appears in content must first be registered in the case library.

Registration has a form — who, what, the measured result, the source and date confirmed. If even one of the four boxes can't be filled, it isn't registered, and if it isn't registered, it can't go into a card or an article.

The effect of this structure is simple. "Where did I see that again?" disappears. Even if one figure appears in several pieces of content, its basis lives in one place, and if it's wrong, you only fix one place.

2Five labels — the same number carries different weight

Not every figure holds the same standing. We label them in five tiers.

Primary-confirmedA figure we read and confirmed directly in the primary source (the original report, an official announcement, a statute). Fit for a headline.
Secondary citationA figure quoted by another outlet, where we haven't yet seen the original source. We label it and use it only as directional evidence — no standalone headline.
Vendor claimA figure released by a company that sells the tool or platform. It may be true, but there's a stake involved, so we note "vendor claim" every time we cite it.
Self-reported/anecdotalA first-person social-media review or interview. Cited only for narrative and perspective, never used as if it were a verified figure.
TrackingAn interesting story whose original source we couldn't find. When used in the text, we state "tracking" and write it as a tendency rather than an assertion.

As a reader, read it this way — if you see a figure that appears to carry no label, that's our mistake.

3What the rules actually caught

A rule only lives if it has cases. These are all real records.

Correction — 7.8% was actually 7.6%Several blogs were quoting "value-added up 7.8% at firms that adopted AI." When we found and checked the original survey (KCCI's SGI), it was 7.6%. A 0.2-percentage-point difference, but we corrected it before registering, and left a record of the correction in the library. In a chain of blogs citing blogs, numbers slip a little at a time.
Discard — the source article didn't contain that numberWe found an attractive figure, "public-institution adoption from 34% to 77%," but when we opened the article cited as its source, the number wasn't there. We didn't register it, and used it in no content. A link being present doesn't mean the evidence is — you have to open it to know.
Softening — we deleted "eight in ten"An early draft of one of our cards had the sentence "eight in ten adoptions stall this way." Plausible, but with no source. It got caught in an internal audit and was changed to "this is the common shape among teams that stalled." Your own writing is the most dangerous — a plausible sentence reads like conviction even without a source.
Hold — a good story, but no originalA domestic case of an individual who built thirty apps was an article candidate, but we couldn't secure the original post, so we excluded it. Instead we swapped in two individual cases that had press verification. The better the story, the more you want to lower the bar for verification — that's the moment the rule does its work.
Dating it — 96% is a 2021 numberA bank chatbot's "96% answer-similarity rate" is a fine figure, but it's a report from five years ago. Published as is, it reads as current. We stamped the date into the text and attached "latest figure being tracked." A figure's shelf life is part of its source.

4The more famous the number — attach the counterargument

The figure "95% of enterprise AI pilots fail" is quoted everywhere. We cover it too — but alongside the counterargument.

There's criticism that the study defines success narrowly (recognizing only profit-and-loss impact within six months) and that its core figure is a directional estimate based on 52 interviews. So in our content, the 95% is used not as a "probability of failure" but as a "list of failure mechanisms."

Stated as a principle, it goes like this — the more famous a number, the more we find and read at least one piece criticizing it before we publish it.

5If you want to try it — five rules for team reports

This procedure isn't just for a content site. It carries over directly to team reports and executive presentations. In an age where AI writes the first draft, it's become more necessary, not less.

  1. Open the original for every figure. If it's a citation of a citation, two clicks to the original source. If there is none, drop the figure.
  2. Attach a label. Primary-confirmed / secondary / vendor / self-reported — a one-word footnote is enough.
  3. Write the date. "96% (as of 2021)" and "96%" are different sentences.
  4. For famous numbers, find the counterargument. If there's no counterargument, you just haven't looked hard enough yet.
  5. Keep a record of what you filtered. The more corrections and discards pile up, the more the remaining numbers earn trust.

You can have AI do it — "make a table of every figure in this document with a link to the source original and whether it's confirmed." Just, the last click is a person's.

6The gist

Generation got cheap; verification didn't. That's why verification becomes the differentiator.

That's also why we disclose even our correction records — a record of fixing it when we were wrong is more credible than a claim of never having been wrong.

The things made with these rules → Full Learn contents · Deep Read contents

The correction, discard, softening, and hold cases in the text are all real entries recorded in the AXHub case library and work log (2026-07-02 to 05). Detailed sources for the surveys and reports mentioned are in the corresponding library entries.