Four numbers were published this week that are large enough to be repeated all year. Here they are in the form they will be repeated in:
- One in five school districts cancelled school because of the building.
- Federal agencies spent $9.5 billion paying people who were not working.
- Repealing the power plant carbon rules saves $160 billion.
- The original reasons for the shareholder proposal rule did not hold up.
Every one of those is a fair summary of a real document. And three of the four documents contain a sentence, usually in smaller type, that tells you how wrong the number might be. Finding that sentence is the entire skill.
One: the number that publishes its own range
GAO estimates that 19 percent of school districts cancelled school in 2024-25 over a facilities problem, affecting 2 million students and 3.5 million days of learning. Directly under the chart:
Note: The 95 percent confidence intervals for these estimates are 15 to 24 percent of districts; 1.5 to 2.4 million students affected; and 2.5 to 4.5 million days of learning lost.
Plus the survey's own vital signs: "overall weighted response of 62 percent," and "estimates presented have a margin of error of no more than plus or minus 7 percentage points."
This is the best case. You are told the point estimate, the interval, the response rate and the margin. You can decide for yourself whether to say "one in five" or "between 15 and 24 percent," and both are honest. Note what the interval does not do: it does not make the finding soft. The bottom of the range is still a million and a half children.
Two: the number that names the direction of its error
GAO's $9.5 billion in paid administrative leave comes with a defect report attached. Agencies "reported 144 percent more paid administrative leave used in pay periods with a public holiday in 2023 through early 2025," and GAO states the rule plainly: "Holidays should not be reported as paid administrative leave, and OPM has issued guidance to help address this data issue."
GAO states the consequence in the direction it runs: the limitations "could overstate the actual amount used."
That is a different and in some ways more useful kind of honesty than an interval. There is no tidy ± here. There is a named mechanism — holidays coded wrong — a measured size, and an arrow. You now know that $9.5 billion is a ceiling-ish figure rather than a midpoint, which changes how you would use it in an argument.
It also tells you something the number alone cannot: OPM "does not plan to retroactively fix such historical errors in data released to the public." The error is documented and staying.
Three: the number that changes by a third depending on one assumption
EPA's savings figure for the power plant repeal is not one number. From the rule itself: the present value of compliance cost savings over 2026-2047 is "$160 billion using a three percent discount rate and $95 billion using a seven percent discount rate."
Same repeal. Same years. Same dollars. One choice of discount rate moves the headline by $65 billion. Neither figure is a trick — publishing both is the correct practice, and the reason to publish both is precisely that the choice is an assumption rather than a measurement.
Then the note under the table, which is the sentence that matters most:
For the reasons the EPA describes in section 4 of the RIA, the Agency has not monetized several impact categories, including the potential health and welfare impacts of changes in emissions. Therefore, this table does not include those impact categories.
So the $160 billion is not net of anything on the health side. It is not that EPA weighed those effects and found them small; the document says they are not in the arithmetic. A savings number with one column deliberately empty is still a real number — but saves $160 billion and saves $160 billion in compliance cost, with the health effects not counted are different sentences, and only one of them is what the document says.
Four: the claim with no number under it at all
The SEC's fact sheet gives this as a policy reason for rescinding Rule 14a-8: "many of the justifications that were originally provided to support adoption of Rule 14a-8 either have not been substantiated in practice or are less compelling today."
Read that as an empirical claim, because that is its grammar. Something was expected; it did not happen. But not substantiated by what? No count of proposals, no measure of what they did or failed to do, no study named — in the press release or in the two-page fact sheet, which are the documents this desk read. The full proposing release may well quantify it; this desk has not read it and will not pretend otherwise.
That gap is not proof the claim is wrong. It is the reason the claim cannot yet be checked, which is a different thing and worth saying out loud. When a number is absent, the useful question is not "is this true" but "what would have to be counted for this to be testable, and did anybody count it?"
The one-minute habit
For any number you are about to pass along:
- Find the source's own sentence about error. A confidence interval, a margin, a stated direction of bias, a data-limitation paragraph, a range across assumptions. It is usually within a paragraph of the number, often in a footnote or a table note.
- Ask what is not in it. Not monetized. Excludes. For the agencies in our review. Among districts responding. These phrases are the boundary of the number, and they are always written down.
- If there is no such sentence, say so when you repeat the number. Not as an accusation — as a fact about what you know. "The Commission says the original justifications were not substantiated; the documents I read do not quantify it" is a stronger sentence than either repeating it flatly or dismissing it.
Three of this week's four documents did the work for you. That is a better ratio than most weeks, and it is only visible if you look for the sentence.