The number you measured once, now wrong in four documents

You measured something. It went into a summary. The summary went into a plan, the plan into a deck, and someone quoted the deck in an email. Then the underlying thing changed.

One measurement is now wrong in four places, all of which look authoritative, and nothing connects them. This has always happened. AI-assisted work makes it considerably worse, for a reason worth understanding, and there are three habits that contain it.

Why is this worse now?#

Because documents multiply faster than corrections travel.

Producing a summary used to cost an afternoon, which limited how many existed. Now a figure can move from a spreadsheet into a briefing, a proposal, a set of talking points and a client update in a morning, each one a plausible standalone document.

Corrections did not speed up. Fixing the source still requires knowing which four documents copied it, and nobody wrote that down, because at the time it was one number in one summary.

The result is a specific, recognisable failure: several people confidently citing different versions of the same figure, all of which were correct when they were copied.

The three shapes it takes#

The carried figure. Somebody quotes a number from a previous document rather than from the source, because the previous document is closer to hand and looked reliable. Each hop adds distance from the measurement and none of the hops are recorded.

The relabelled figure. A number keeps its value and loses a word. A rate measured one way gets described as measuring something slightly different, and the new description is entirely plausible. This is the expensive one, because both readings sound correct and nothing downstream catches it.

The re-based figure. The population changed and the number did not. A percentage measured against one group gets presented against a larger one. The temptation is a find-and-replace on the label, which quietly asserts a measurement nobody made.

Three habits that contain it#

1. Cite the source, not the sentence. When a figure enters a document, record where it came from and the date it was read. Not "revenue is up 14%" but "revenue is up 14% (dashboard, read 12 August)".

This costs a few words and does the most work of anything here, because it makes staleness visible without anyone checking. A reader seeing a date from three months ago knows what to do.

2. Never copy a number you have not re-derived. If a figure is going into something durable, go back to where it was measured and read it again. Not the document that carried it. The place it was measured.

The rule sounds excessive until the first time a re-derivation produces a different answer, which happens more often than anyone expects.

3. Do not put fast-moving numbers in prose at all. For anything that changes weekly, reference where it lives rather than what it says. "Current spend is in the tracker" ages correctly. "Current spend is $4,300" is wrong within a week and looks authoritative forever.

A test worth running once#

Take the most-quoted number in your work. The one that appears in the deck, the update and the summary.

Go back to where it was originally measured and read it again.

Then find every document that carries it and check whether they agree, with each other and with the source.

Most people running this test for the first time find at least one disagreement. The point is not the disagreement. It is discovering how many documents carried it, because nobody knew that number either.

I am about to put a number into a document that other people will use. Check my handling of it before I do. Ask me:
1. What is the number, and what exactly does it measure? Make me state the units and the population it was measured against.
2. Where did I get it from, and is that the place it was measured or a document that quoted it?
3. When was it measured, and how often does the underlying thing change?
4. What else will carry this number once I publish it?
Then tell me whether this number should appear as a value at all, or whether I should reference where it lives instead. If I got it from a document rather than the source, tell me to go back and re-derive it.

Why AI makes the last habit hardest#

Assistants are very good at producing polished prose containing figures. Asked to write an update, a model will happily state a number it read in the material you gave it, in a confident sentence, with no indication of where it came from or when.

That output is not wrong. It is a faithful copy of what you provided. But it has stripped the provenance, and the resulting sentence is the most quotable version of the figure that will ever exist.

If you give a model numbers without their sources, it will hand you back numbers without their sources, in better prose.

Ahead pairs you with an expert operator who works on your own setup, in your own tools. If you have a figure that appears in more than three places and you are not certain they agree, that is a good first hour.