← CHANGELOG
Two ways to be wrong while every check you ran came back clean.
16 August 2026·5 min read·Onkydra · Faith Ogundimu
Most advice about checking your work assumes errors are arithmetic. In practice the ones that survive are the ones where the arithmetic is correct and an assumption underneath it is not, because everything you would naturally check comes back clean. Here are two from the last week, one financial and one statistical, that failed in the same shape.
The seat that was billed for seven days
Our profit-and-loss statement carried one line for a Google Workspace seat. It was filed as an estimate, taken off a checkout screenshot: 8.10 euro plus 23% VAT, so 9.96 euro. The rate was right and the VAT rate was right.
The vendor's account-activity export says the seat cost 2.24 euro. The 14-day free trial ran to 24 July, so the seat was billed for 25 to 31 July only. Seven days, not a month. The error ratio is 9.96 divided by 2.24, which is 4.44, and 31 over 7 is 4.43. The whole of the error was one unstated assumption about a date range.
8.10 euro plus 23% really does equal 9.96 euro. Every check anyone would run on that line passed, because the only wrong thing in it was a fact nobody had a source for. It moved the filed expense total from 22.31 to 13.91 US dollars, on a statement small enough that a single line item is a third of it.
The replacement is checkable in two independent ways, which is the standard the estimate never met. The recorded VAT is 0.42 on 1.82, or 23.08%, which is Ireland's standard rate. And seven days of a 31-day month at 8.10 euro is 1.83, against a recorded 1.82: a one-cent rounding difference that confirms the list rate and the seven-day period at the same time.
The effect size that was a mean difference
The second one cost a day rather than a month. We had measured that an immediate-early gene panel falls in single-nuclei libraries relative to single-cell libraries, and the sentence we published reported it as Cliff's delta 0.63.
0.628486 is that panel's mean difference in the same block of the same file. Its Cliff's delta is 0.959596. The effect was quoted about a third smaller than it is, under the name of a different statistic.
What made it durable is worth more than the correction. Three figures were pulled from one record and two of them were right: the p value of 1.9e-06 and the Cohen's d of 3.46 sitting next to it were both correct and both checkable. Anyone verifying the sentence would confirm two of three and stop, which is what verifying a sentence usually means.
What the two have in common
In both cases the wrong number was the only one in the sentence with no source behind it, and it was surrounded by numbers that had one. An inference that reconciles internally is not evidence. Consistency tells you that the numbers you have agree with each other, and says nothing at all about whether one of them came from anywhere.
The two fixes are correspondingly dull. Every figure on the financial statement is now read off a vendor record with nothing inferred, and the numbers on it are asserted against the rendered document by a script, so the three artefacts cannot drift apart quietly. The effect size is quoted by its path in the file it comes from, so the name and the value are read together or not at all.
This is the same failure as the transposed axis in our knockout table, which shipped for five days with the test suite green. There the number that had no independent source was the one the reader and its own test agreed about. See the post on that one for the version of this with a longer fuse.
The habit that would have caught both
Ask of every number in a sentence: which of these could I be wrong about without anything else in the sentence changing? That is a shorter list than the numbers themselves, and it is where to spend the checking. In the invoice it was the period. In the effect size it was which statistic the value belonged to. Neither is the sort of thing recomputing the sum will find.