This site refuses to build when one of its own figures goes stale.

The usual way a figure dies is quietly. It was true when it was published, nothing happened to the page, and five years later it is a statistic about a market that has changed shape.

A "last updated" date does not fix this, because it records when somebody touched the page rather than when anybody checked the number. A date nobody is obliged to renew is decoration.

Why the class rather than a global rule

Because "review your content every six months" is advice that treats a price and an Act of Parliament as the same object.

A leaderboard changes without notice and without a version number; thirty days is generous. A statutory instrument rarely changes and is catastrophic when it does and you did not notice; a hundred and eighty days with a hard stop is right. A randomised trial published in 2019 has a figure that will not move, but the paper could be retracted or restated, so it gets read once a year. Arithmetic never expires, and saying so out loud is better than a date nobody will honour.

The expiry case is separate and stronger

A prediction with a horizon — "40% of enterprise applications by the end of 2026" — does not decay. It stops being a forecast on a known date and becomes a claim about the past, at which point it is either wrong or it is history, and either way it cannot keep being quoted as a forecast.

That date is written down beside the source when the source is a prediction, and hitting it is treated as harder than going stale.

What it actually costs

Less than expected, and not where expected.

The cost is not re-reading sources. It is choosing the class honestly when you know a stricter one will interrupt you. Labelling our own Vercel bill as "live" would force a re-read every thirty days; labelling it "dated" means once a year. We chose "dated" because the note reports three specific days rather than a running total, and that reasoning is written in the source entry so it can be disagreed with.

The second cost is real but rare: a build blocked on a stale figure in an unrelated note, on the day you wanted to fix a typo. The thirty-day grace exists for exactly that, and it is a grace rather than a licence — the report prints from day one.

What we would tell another firm

You almost certainly do not need this. It is heavy machinery for a site with 310 sources and a stated position that every figure names its origin.

What is worth stealing is smaller and free: beside each number you publish, record where it came from and when you last looked. Not in a spreadsheet — beside the number, where the next person editing the page will see it. That alone prevents the most common failure, which is not staleness but amnesia: nobody remembers where the figure came from, so nobody can check whether it still holds.

The related discipline is refusing to publish what fails checking, which cost us nine figures and is set out in the note on the ones we removed. The whole register is public at what we checked, with every source, its class and its next date.

What this does not tell you

A re-check date does not make a figure true. It makes it re-read, which is a lower bar and a different virtue — the 1,445% claim we struck was not old, it was wrong on the day it was published, and no cadence would have caught it.

Reading the source against the sentence catches that, and it is a separate exercise. So is looking for what has happened since: on 24 September 2026 that second exercise found notes whose sources were right when read and had since been overtaken, which a cadence would have caught only when it came round.

Nor does the system judge. It tells a human that something is due; a human decides whether the figure still holds. Automating the reminder is easy and automating the judgement would be the mistake — which is the same boundary we draw in every system we build, and it is the argument in the note on what an agent may decide and the audit layer.